Modelling the Stability and Determinants of Household Food Insecurity: A

Multivariate Longitudinal Ordinal Logistic Regression Approach

by

Jemal Ayalew Yimam

submitted in accordance with the requirements for the degree of

DOCTOR OF PHILOSOPHY

in the subject

STATISTICS

at the

UNIVERSITY OF SOUTH AFRICA

SUPERVISOR: Professor John O. Olaomi

2019
 DECLARATION

Name: Jemal Ayalew Yimam

Student Number: 50879235

Degree: PhD

Exact wording of the title of the thesis as appearing on the electronic copy submitted for
examination: 16

I declare the above thesis is my own work and that all the sources that I have used or
quoted have been indicated and acknowledged by means of complete references.

I further declare that I submitted the thesis to originality checking software. The result
summary is attached.

I further declare that I have not previously submitted this work, or part of it, for
examination at UNISA for another qualification or at any other higher education
institution.’

___ ___September. 2019________

Signature Date

i|Page
 Acknowledgements

First of all, I praise Allah, the almighty, for enabling me to complete this long journey. I have
been very fortunate to have been supported by many people. First and foremost, my deepest
appreciation goes to my Supervisor, Professor John O Olaomi. He is the most wonderful mentor
who has always been a true inspiration and a role model. As my supervisor, he has given me
constant support, guidance, encouragement, and good humour. I have learned a great deal from
his expertise and experiences in research on modelling food insecurity determinants. I am greatly
indebted to him for his patience, generosity and understanding as well as his insightful
suggestions.

I would like to thank both Wollo University and University of South Africa for their valuable
and unreserved arrangement of studying this education and granting both the ethical clearance
and the financial support to accomplish this long journey work.

My special thanks also go to South Wollo Zone Woreda and Kebele leaders for their positive and
genuine cooperation provided to me during data collection. Many thanks also for the Agricultural
Agent workers participated as supervisor and data collector for their contribution to successfully
accomplish the data collection.

I wish to thank my parents for their unconditional love and support throughout my life. Without
it, I could have never made it this far. I am also grateful to my mother for her prayers and
understanding.

Finally, I owe tremendous amount of gratitude to my wife, Seada Bezabih, and my children,
Sumaya, Ameterrehman, Huzeyfa and Abdurrehman. Seada has been the wind beneath my
wings, enabling me to fly so high to accomplish my goals. It was a long journey, but thanks to
her love and support; it was incredible.

iii | P a g e
© Yimam JA, UNISA 2019
 List of Abbreviations

2SLS : 2-Stages Least Square

AMH : Ali-Mikhail-Haq
C : Copula(s)
CFSI : Composite Food Security Index
EM : Expectation Maximization
ERHS : Ethiopian Rural Household Survey
FAO : Food and Agricultural Organization
FCS : Consumption Score
GEE : Generalized Estimating Equations
GLS : Generalized Least Squares
HDDS : Household Dietary Diversity Score
HFIAS : Household Food Insecurity Access Scale
HHS : Household Hunger Scale
IEE : Independent Estimation Approach
IFM : Inference Function for Margins
IV : Instrumental Variable
MAHFP : Months of Adequate Household Food Provisioning
mAIC : modified Akaike Information Criterion
ML : Maximum Likelihood
MLE : Maximum Likelihood Estimation
MPL : Maximum Pairwise Likelihood
OLS : Ordinary Least Square
PCAI : Principal Component Analysis Index
PCC : Pair-Copula Construction
PMF : Probability Mass Function
SEM : Structural Equation Modelling
SSP : Stepwise Semi-parametric Estimator
USDA : United States Department of Agriculture
WASH : Water Supply, Sanitation, and Hygiene
WFS : World Food Submit

iv | P a g e
© Yimam JA, UNISA 2019
 Summary

Multivariate longitudinal ordinal data are collected for studying the dependence between
multivariate ordinal outcomes, the changes over time and associated determinant factors. This
emanates from the interdependence of the three dimensions of household food security statuses,
the stability of these dimensions over time and the additional contribution of covariates on the
dependence structure.

It is generally known that the random effect models have a lack of population-averaged
interpretation for non-normally distributed outcomes in analysing ordinal data. In this thesis, we
propose an alternative model for analysing multivariate longitudinal ordinal data with application
to the household food insecurity by developing a pair copula construction (PCC) and cumulative
logit marginal distributions-based model using the full maximum likelihood estimation (MLE)
method. The simplified log-likelihood function of the D-vine pair copula multivariate discrete
random variables was obtained with its parameters estimated.

Data were collected from 646 households living in selected rural Woredas of South Wollo Zone
of the Amhara Regional State, Ethiopia from June 2014 to June 2015 three times at six months
interval. Multistage cluster sampling was employed to select representative Woredas and
households. The household food security status was determined using both the quartile score and
composite index. Three distinct pair copula models with cumulative logit version were employed
for multivariate, longitudinal and multivariate longitudinal ordinal data applicable for household
food security.

The first model was employed to assess the dependence between food security dimensions and
their corresponding determinant factors simultaneously. The copula parameter of this model
indicated that household food security dimensions have significant and positive pairwise
dependence. The marginal parameters showed that smaller land size, shortage of rainfall,
cultivating once a year, and the presence of disease were positively associated with chronic to
mild food insecurity in all dimensions. Moreover, cold agro-ecology and market price increase
were associated with household food insecurity at availability and accessibility dimensions.

The second model was used to assess the stability of household food security over time and the
determinant factors. The copula parameter revealed that individual household food security

v|Page
© Yimam JA, UNISA 2019
status is not stable over time. Moreover, the marginal parameter indicated that presence of crop
disease, market price increase and medium agro-ecology were the significant recurrent factors
for households to have chronic to mild food insecurity throughout the study period. One-time
cultivation per year was the temporal significant factor for household food insecurity.

The third model was developed for measuring the dependence between the three dimensions,
namely, their stability over time, the effects of the covariates both on the dependence structure,
and stability over time simultaneously. The copula parameter of the population-average
cumulative logit model revealed that food security dimensions were positively dependent to each
other and the individual household food security status is not stable over time.

The marginal parameter of this model provided that lower agro-ecology, shortage of rainfall,
presence of cultivation disease, increased market price, use of pesticides, cultivating smaller
types of cereal crops, and cultivating once per year were positively affects the household food in-
security in availability dimension. On the other hand, lower agro-ecology, increased market
price, herbing small amount of livestock, hot agro-ecology and small farmland size positively
affect the household food insecurity in the accessibility dimension. Furthermore, households
headed by wife, divorced/widowed marital status of the household head, shortage of rainfall, and
small farmland size positively affect the household food in-security in utilisation dimension.

This model provided a population-average interpretation with acceptable computational

challenges in multivariate longitudinal ordinal data analysis. The study suggests that food
security situation analysis is a multidimensional so that over-sighting the three dimensions over
time simultaneously provides detail household food security situation than the single dimension.
The pair copula population-average cumulative logit model addressed all the food security
dimensions simultaneously, and the model found computationally effective. Therefore, we
suggest this model to apply for other application areas for not extremely large number of
outcomes and covariates.

Keywords:

Food insecurity; chronically food in-secured; composite food index; multivariate ordinal
outcomes; longitudinal ordinal outcomes; multivariate longitudinal ordinal outcomes; marginal
model; cumulative logit, pair copula; full maximum likelihood

vi | P a g e
© Yimam JA, UNISA 2019
 Tables of Contents

DECLARATION ............................................................................................................................. i
Acknowledgements ........................................................................................................................ iii
List of Abbreviations ..................................................................................................................... iv
Summary ......................................................................................................................................... v
Tables of Contents ........................................................................................................................ vii
List of Tables ................................................................................................................................ xii
List of Figures .............................................................................................................................. xiv
Chapter One .................................................................................................................................... 1
Introduction .......................................................................................................................... 1

Chapter Two.................................................................................................................................... 8
Literature Review................................................................................................................. 8

2.1 Statistical Models in Food Security .............................................................................. 8

Multivariate Ordinal Models ................................................................................. 8

Longitudinal Ordinal Models .............................................................................. 10

Multivariate Longitudinal Ordinal Models ......................................................... 13

2.2 Definition and Concept of Food Security ................................................................... 16

2.3 Food Security Measurements and Determinant Factors ............................................. 18

Chapter Three................................................................................................................................ 24
Data .................................................................................................................................... 24

3.1 Study Area and Population ......................................................................................... 24

3.2 Data Collection Instrument ......................................................................................... 24

3.3 Sampling Design and Procedure................................................................................. 25

3.4 Sample Size ................................................................................................................ 26

3.5 Data Collection ........................................................................................................... 27

3.6 Data Collection Periods .............................................................................................. 28

3.7 Measuring Food Security............................................................................................ 29

vii | P a g e
© Yimam JA, UNISA 2019
 3.8 Characteristics of Some Study Variables ................................................................... 33

Chapter Four ................................................................................................................................. 35

Methodology ...................................................................................................................... 35

4.1 Introduction ................................................................................................................ 35

4.2 Basics of Copula Theory ............................................................................................ 36

Definition and Properties of Copula Theory ....................................................... 36

Copula Density .................................................................................................... 38

Families of Copulas ............................................................................................. 39

Elliptical Copulas ............................................................................................ 39

Archimedean Copulas...................................................................................... 41

Pair Copula Construction................................................................................. 43

Dependency Measures......................................................................................... 45

Measure of Concordance ................................................................................. 45

Tail Dependence .............................................................................................. 46

Dependencies Characteristics of Bivariate Copula Families ........................... 47

4.3 A Pair Copula Construction Approach for Multivariate Ordinal Data ....................... 48

PCCs in the Continuous Case.............................................................................. 49

PCCs in the Discrete Case ................................................................................... 55

4.4 Pair Copula Construction for Longitudinal Ordinal Data .......................................... 62

Pair Copula Construction for Longitudinal Continuous Data ............................. 63

Pair Copula Construction for Longitudinal Discrete Data .................................. 65

Selection of Pair Copula Families and Parameter Estimation of the D-Vine ..... 68

Selection of Pair Copula Families ................................................................... 68

Parameter Estimation ....................................................................................... 70

4.5 Pair Copula Construction for Multivariate Longitudinal Ordinal Data ...................... 70

viii | P a g e
© Yimam JA, UNISA 2019
 Pair Copula Construction for Multivariate Longitudinal Continuous Data ........ 72

Pair Copula Construction for Multivariate Longitudinal Discrete Data ............. 75

D vine in Multivariate Longitudinal Discrete Data ......................................... 76

Selection of Pair Copula Families ................................................................... 79

Parameter Estimation of the D-Vine ............................................................... 80

4.6 Statistical Software ..................................................................................................... 81

4.7 Variable Selection....................................................................................................... 81

Chapter Five .................................................................................................................................. 82

Analysis of the Household Food Data ............................................................................... 82

5.1 Internal Validity Assessment of the Data Collection Tools ....................................... 82

5.2 Characteristics of the Study Participants .................................................................... 82

5.3 Household Food Security Analysis ............................................................................ 86

5.4 A Pair Copula Construction Approach for Multivariate Ordinal Data with Application
to Household Food Insecurity Data ....................................................................................... 88

Kendall's tau Correlation Coefficient .................................................................. 88

PCC Selection ..................................................................................................... 89

Estimation of the Copula and Marginal Parameters............................................ 90

Effects of PCC on the Univariate Cumulative Logit Model ............................... 93

5.5 A Pair Copula Construction Approach for Longitudinal Ordinal Data with
Application to Household Food Insecurity Data ................................................................... 96

Kendall's tau Correlation Coefficient .................................................................. 96

PCC Selection ..................................................................................................... 97

Estimation of the Copula and Marginal Parameters............................................ 98

Effects of PCC Model on the Univariate Cumulative Logit ............................. 101

5.6 A Pair Copula Construction Approach for Multivariate Longitudinal Ordinal Data
with Application to Household Food Insecurity Data ......................................................... 103

ix | P a g e
© Yimam JA, UNISA 2019
 Kendall's tau Correlation Coefficient ................................................................ 103

PCC Selection ................................................................................................... 104

Estimation of the Copula and Marginal Parameters.......................................... 105

Effects of PCC Model on the Univariate Marginal Cumulative Logit Model .. 109

Chapter Six.................................................................................................................................. 113

Discussions ...................................................................................................................... 113

6.1 Discussions for the Findings of the Pair Copula Based Multivariate Ordinal Model
Application to Food Security Data ...................................................................................... 113

6.2 Discussions for the Findings of the Pair Copula Based Longitudinal Ordinal Model
Application to Food Security Data ...................................................................................... 117

6.3 Discussions for the Findings of the Pair Copula Based Multivariate Longitudinal
Ordinal Model Application to Food Security Data ............................................................. 119

6.4 Potential Policy Implementing Strategies................................................................. 123

Chapter Seven ............................................................................................................................. 126

Conclusions, Recommendations and Future Works ........................................................ 126

7.1 Conclusions .............................................................................................................. 126

7.2 Recommendations .................................................................................................... 127

7.3 Limitations and Weaknesses of the Study ................................................................ 128

7.4 Future Works ............................................................................................................ 129

References ................................................................................................................................... 132

Appendices .................................................................................................................................. 143
Appendix A: Questionnaires ................................................................................................... 143

Appendix B: Internal consistence analysis of the data collection tools. ................................. 154

Appendix C: The joint probability distribution based on the D-vine pair copula was displayed
as follows................................................................................................................................. 137

Appendix D: R codes for the log-likelihood D-vines ............................................................. 147

Appendix E: Plagiarism Report............................................................................................... 168

x|Page
© Yimam JA, UNISA 2019
 Appendix F: Ethical Clearance Approval ............................................................................... 170

Appendix F: Language editing certificate ............................................................................... 172

xi | P a g e
© Yimam JA, UNISA 2019
 List of Tables

Table 3. 1 Sample size allocations for the selected Woredas and then to selected Kebeles with-in

the respected Woredas .................................................................................................................. 27

Table 3. 2: Cut-off points for household access scale .................................................................. 31

Table 4. 1: Kendall’s tau, upper and lower tail dependence for bivariate copula families

(Dissmann, 2010). ......................................................................................................................... 48

Table 5. 1: Summary measures of non-time varying variables of the households ....................... 83

Table 5. 2: Summary measures of the time-varying variables of the households ........................ 85

Table 5. 3: Household food security statuses of the three food security dimensions in all data

collection rounds (f stands for frequency and % for percent) ....................................................... 87

Table 5. 4: The Composite Food Security Status of households in the three data collection rounds

....................................................................................................................................................... 88

Table 5. 5: Nonparametric correlation coefficients and their significant α values in brackets of

the household food security status of the dimensions................................................................... 88

Table 5. 6: The Kendall’s tau for the pseudo data computed from the cumulative logit marginal

distribution for the application data availability, accessibility and utilisation.............................. 89

Table 5. 7: The summary of copula families by applying Algorithm 1 to the data and estimating

modified Akaike Information Criterion (mAIC) to select the best fit copula families. ................ 90

Table 5. 8: The estimates of dependence parameters using the selected pair copula for the

application data of multivariate ordinal household food security status. ...................................... 91

Table 5. 9: Summary results of pair copula based cumulative logit model parameters estimates

for the household food security data. ............................................................................................ 92

xii | P a g e
© Yimam JA, UNISA 2019
Table 5. 10: Summary results of the univariate cumulative logit model parameters estimates for

the household food security data................................................................................................... 95

Table 5. 11: Nonparametric correlation coefficients of the household food security states of the

three data collection phases .......................................................................................................... 97

Table 5. 12: The Kendall’s tau for the pseudo data computed from the cumulative logit marginal

distribution for the application of three-phase longitudinal data. ................................................. 97

Table 5. 13; Summary of copula families by applying Algorithm II to the data and estimated

modified Akaike Information Criterion (mAIC) to select the best fit copula families. ................ 98

Table 5. 14: The estimates of dependence parameters using the selected pair copula for the

application data of longitudinal household food security status. .................................................. 99

Table 5. 15: Summary results of the marginal parameters of pair copula based longitudinal

cumulative logit model for the household food security data. .................................................... 100

Table 5. 16: Summary results of the marginal parameters of the univariate cumulative logit

model for the household food security data ................................................................................ 102

Table 5. 17: Nonparametric correlation coefficients of the household food security states of the

three combined dimensions......................................................................................................... 103

Table 5. 18: The Kendall’s tau for the pseudo data computed from the marginal model of the

cumulative logit marginal distribution for the multivariate longitudinal data. ........................... 103

Table 5. 19: Summary of copula families using Algorithm I to the data and estimated modified

Akaike Information Criterion (mAIC) to select the best fit copula families. ............................. 104

Table 5. 20: The estimates of dependence parameters using the selected pair copula for the

application data of multivariate longitudinal household food security data. .............................. 106

xiii | P a g e
© Yimam JA, UNISA 2019
Table 5. 21: Summary results of pair copula based marginal model via cumulative logit model

parameters estimates for the household food security data. ....................................................... 107

Table 5. 22: Summary results of the univariate marginal model via cumulative logit model

parameters estimates for the household food security data. ....................................................... 110

List of Figures

Figure 3. 1: Framework of food security of the household at three follow-up time points at six-

month interval ............................................................................................................................... 33

Figure 4. 1: A D-vine tree representation for m = 5. .................................................................... 53

Figure 4. 2: A C-vine tree representation for m = 5. .................................................................... 53

xiv | P a g e
© Yimam JA, UNISA 2019
 Chapter One

Introduction

This study attempted to fill the methodological gap of jointly modelling the stability and
determinants of the household food security for each dimension using multivariate longitudinal
ordinal outcome. Since, multivariate joint modelling has a pretty advantage over separate
modelling because it has relative efficiency to estimate the model and provide a powerful test of
significance than the univariate (Gueorguieva, 2001; Molenberghs and Verbeke, 2006).

Among previous studies for multivariate longitudinal ordinal outcomes, the majority of them
were concentrated around random effect models. Among these, random effect models in the
context of item response theory (Liu, 2008, Liu and Hedeker, 2006), random effect models in
the context of latent variable model (Cagnone et al., 2009), random effect models using subject-
specific model (random intercept models) (Choi, 2012, Verbeke et al., 2014) and random effects
models by introducing a continuous distributed random variable underneath the ordinal
outcomes, some form of latent variable models (Laffont et al., 2014) were employed so far.
However, the random effect models have lack of a population-averaged interpretation for non-
normally distributed outcomes and some computational challenges (Abegaz et al., 2015,
Nooraeea, 2015).

Despite numerous studies on multivariate longitudinal outcomes, relatively little research has
been conducted on the marginal models for ordinal outcomes on the context of population
average. Among these, Generalized Estimating Equation (GEE) was one of the most popular
models. GEE was proposed for binary outcome and continuous outcome (Rochon, 1996).
Furthermore, Gray and Brookmeyer (2000) proposed multivariate longitudinal models for
continuous and discrete/time-to-event response variables using GEE. Another marginal model
tailoring GEE was proposed for measuring multicity measured ordinal outcomes (Huang et al.,
2002). Even though GEE is popular and provided consistent estimators for the regression
parameters in the population-average interpretation; when the focus of the analysis includes
certain aspects of the association structure, the construction of the joint model using GEE
becomes more complex as it implies making assumptions about the within-outcome, the
between-outcome, and the cross outcome association, and inferences of interest can be very

1|Page
© Yimam JA, UNISA 2019
sensitive with respect to the assumptions made (Verbeke et al., 2014). Moreover, GEE cannot
incorporate the contribution of determinants of the outcome in determining the dependence
between outcomes since it measures it using a working correlation independent of the
determinants.

Another alternative model that helps for parameter estimation for non-normally distributed
continuous or categorical data was the quasi least squares (Chaganty and Naik, 2002). A
marginalised bivariate model using Kronecker product (KP) covariance structure was also
proposed to handle two longitudinal ordinal outcomes (Lee et al., 2013). Marginal models using
maximum likelihood estimation (MLE) on the context of multivariate t-copula were developed
for multivariate longitudinal regression model for ordinal responses (Abegaz et al., 2015). The
computation of the probability mass function for a discrete multivariate copula including
multivariate t-copula requires 2m for evaluation while the pair-copula construction (PCC)
method reduces the computation challenge of discrete multivariate copulas which requires only
the evaluation of 2m (m − 1) bivariate copula functions (Lennon, 2016, Nicklas, 2013;
Panagiotelis et al., 2012, Stöber et al., 2015).

In the context of PCC, a general framework for modelling multivariate repeated measurements
using PCC (Shi and Yang, 2016, Shi and Zhao, 2018) and copula-based GLMM models by
combining random-effects models and the D-vine copulas (Zhang et al., 2019) have been
proposed for investigating multivariate longitudinal data with mixed-types of responses. All
these copula-based models have yet not implemented in multivariate longitudinal ordinal
outcomes.

Hence, the current study further attempted to resolve both the population-average interpretations
in the random effect models and computational challenges of the multivariate copulas using pair
copula construction for multivariate longitudinal ordinal data. Furthermore, pair copula
construction requires the determination of the appropriate marginal distributions based on the
nature of the outcomes (Aas et al., 2009, Czado, 2010, Nelsen, 2007, Panagiotelis et al., 2012).
Since the responses of the current study variables are ordinal, the natural choices for ordinal data
are cumulative logit or probit models. Hence, we selected cumulative logit model since the scale
of the logistic is greater than the normal and this made the interpretation easier for logistic
version and popular in many fields (Choi, 2012). In this thesis, the pair copula construction and

2|Page
© Yimam JA, UNISA 2019
cumulative logit marginal distributions are incorporated for the development of multivariate
longitudinal ordinal model using the full maximum likelihood estimation (MLE) method in
particular for the current application data.

The research problem informing this thesis lies in that, reviewed literature demonstrates that
multivariate longitudinal ordinal outcomes both in the random effect and population-average
have some limitations in providing the appropriate interpretations or computational challenges.
The random effects model lacks the population-average interpretation and the population-
average models are limited in number and have computational challenges. As a result, analysing
multivariate longitudinal ordinal outcomes need improvement with the hope of incorporating the
population-average interpretation that can resolve computational challenges. Furthermore,
literature demonstrates that food security experts do not amply utilize the rigour of statistics.
Despite the availability of robust statistical tools that have the rigour to satisfy the quest for
assessing the stability and determinants of food insecurity, existing analysis of survey data
heavily depend on the rudimentary, exploratory or descriptive statistics that lack depth. As a
result, modelling stability and determinants of food insecurity analysis lack the efficiency of
scientifically established evidence.

This thesis is, therefore, aimed at achieving the following main objectives:
1. To develop population-average multivariate longitudinal ordinal models using Pair
Copula Construction, with application to household food insecurity; and
2. To jointly model the stability and determinants of household food insecurity using
multivariate longitudinal ordinal model approach.

Furthermore, this thesis is, therefore, aimed at achieving the following specific objectives in line
with the main objectives:

1. To jointly model the household food security determinants for the three food security
dimensions and the dependence between them using Pair Copula Construction based
Multivariate Ordinal data analysis;
2. To jointly determine household food security dependence between successive time
periods and respective determinants using Pair Copula Construction based Longitudinal
Ordinal data analysis;

3|Page
© Yimam JA, UNISA 2019
 3. To develop population-average based Pair Copula Construction models for Multivariate
Longitudinal Ordinal Data;
4. To apply the population-average based Pair Copula Construction for jointly modelling
the stability and determinants of household food insecurity for each dimension; and
5. To assess food security situation in selected Woredas of South Wollo Zone, Amhara
Region, Ethiopia.

In this light, the study sought an answer to three specific questions:

1. Does the Pair Copula Construction approach alleviate the lack of the population-average
interpretation of the random effect models in modelling multivariate longitudinal ordinal
outcomes?
2. Does the Pair Copula Construction approach alleviate the existing computational
challenges of population-average marginal multivariate copula models in modelling
multivariate longitudinal ordinal outcomes?
3. Does the population-average based pair copula construction of the multivariate
longitudinal ordinal model fits for modelling the stability and determinants of household
food insecurity?

To provide answers to these questions, this thesis aims at exploring the rigour of models that
help in resolving the population-average interpretation of the random effect models and the
computational challenges of the multivariate copula models in the population-average version in
modelling multivariate longitudinal ordinal outcomes. By doing so, the study hopes to strengthen
the theory of statistics in alleviating the aforementioned limitations and enlarge statistical
methods in capturing the complete information of the multivariate longitudinal ordinal data
analysis. Moreover, this thesis aims at exploring the rigour of models that may help in predicting
the determinants of household food insecurity for each dimension simultaneously, generating
indices for the dependence of the dimensions, and the stability of the three dimensions over time.
By so doing, the study hopes to strengthen food security monitoring, evaluation and reporting
systems toward more robust, statistics-based predictive analysis. The usual analysis approaches
often shy away from such approaches in a misconception that statistical methods are complicated
and user-unfriendly. The study specifically taps into recent work in constructing jointly
estimating the stability and determinants for each dimension.

4|Page
© Yimam JA, UNISA 2019
Household food security becomes the multifaceted problems with multidimensional serious
impacts since 1970 (Abafita and Kim, 2014). The widest acceptable definition of food security
states that the existence of food security occurs when “all people, at all times, have physical,
social and economic access to sufficient, safe and nutritious food that meets their dietary needs
and food preferences for an active and healthy life”(Pinstrup-Andersen, 2009). This definition
consists of four important interlinked dimensions, namely; physical availability of food,
economic and physical access to food, food utilisation and stability of the other three dimensions
over time (Abafita and Kim, 2014, FAO., 2014, Napoli et al., 2011). Moreover, the usual
category of household food security levels “food secured” or “food in-secured” were further
disaggregated into “severe food in-secured”, “moderately food in-secured”, “mildly food in-
secured” and “food secured” (Capaldo et al., 2010, Hunnes, 2013). Hence, the first three food
security dimensions take these categories to classify food security status of the households.

Modelling the determinants of household food security should be assessed using the first three
dimensions simultaneously over time to oversee the entire household food security situation and
the stability over time (Capaldo et al., 2010, FAO, 2008). Knowing that the levels of food
insecurity are ordinal in their nature (Magaña-Lemus et al., 2016) and hence modelling of food
insecurity is the question of ordinal data analysis and each dimension can be considered as a
response factor which have ordinal outcomes. Therefore, the methodological issue of food
insecurity is the generalization of developing multivariate longitudinal ordinal data analysis.
Furthermore, each of the dimensions has non-normal correlation to each other. Hence,
identifying the determinant of one or two of the dimensions will not reflect the entire food
security situation (FAO, 2008, FAO., 2014).

Numerous studies were conducted on household food security determinants separately for each
dimension or in a composite index (Abafita and Kim, 2014, Aspelund, 2002, Birhane et al.,
2014, Endale et al., 2014, Etana and Tolossa, 2017, Méthot and Bennett, 2018, Moroda et al.,
2018, Motbainor et al., 2016, Ngema et al., 2018). However, the findings have been quite mixed
and conflicting. This is owing to some of the food access proxy indicators which have served as
food availability proxy indicators and vice versa. Similarly, some of food access proxy indicators
have served for utilisation proxy indicators. This implies that understanding several concepts
associated with the definition of food security are necessary before examining the determinants

5|Page
© Yimam JA, UNISA 2019
of food security. Moreover, despite the fact that FAO recommended that the food security
dimensions has been addressed simultaneously, there is no work conducted on the subject that
can accommodate the determinants for each dimension jointly as well as the stability and
determinants for each dimension over time jointly in a single model.

Longitudinal food security household survey was conducted in selected rural Woredas of the
South Wollo Zone of Amhara regional state in Ethiopia to illustrate the PCC model in
multivariate longitudinal ordinal data. Ethiopia is one of the poorest countries in the world, and
about 90% of the populations live in the rural areas. The problem of food insecurity has
continued to persist in many rural households of the country. The seriousness of the problem
varies from one area to another depending on the state of the natural resources and the extent of
development of these resources (Asmamaw et al., 2015, Endalew et al., 2015). Rural food
insecurity is one of the defining features of rural poverty, particularly in the moisture-deficit
northeast highland plateaus and some pastoral areas of Ethiopian(Agidew and Singh, 2018). The
study area, the South Wollo Zone, is among these areas, which is mostly affected by food
insecurity (Agidew and Singh, 2018, Asmamaw et al., 2015).

Numerous studies have been conducted in Ethiopia regarding the subject of food security with
different results with particular recommendations, and various measures have been taken
(Abafita and Kim, 2014, Abdu et al., 2018, Abegaz, 2017, Agidew and Singh, 2018, Ahmed et
al., 2017, Asmamaw et al., 2015, Assefa, 2015, Bashir and Schilizzi, 2012, Birhane et al., 2014,
Castro, 2000, Endale et al., 2014, Endalew et al., 2015, Etana and Tolossa, 2017, Moroda et al.,
2018, Motbainor et al., 2016, Negatu, 2004, Nigussie and Alemayehu, 2013, Shone et al., 2017).
However, only two studies have been conducted for the last 15 years in South Wollo Zone in
particular in Tewuledery (Agidew and Singh, 2018) and Sayint (Asmamaw et al., 2015)
Woredas. These studies did not reflect the entire food security situation of the zone and they are
concerned only on the food access dimension.

On the other hand, studies conducted by (Castro, 2000, Negatu, 2004) covered the overall
situation of the zone. However, the studies were conducted about 15 years ago, in which the
level of food security and the economy as a whole are very different from this time. Many socio-
economic factors have been changed in the South Wollo Zone and the country too. Therefore, it
is important to update the stability and determinants of household food security or insecurity

6|Page
© Yimam JA, UNISA 2019
situation incorporating all potential factors in the recent models. In this regard, our model is
more robust than the previous which will lead to a more accurate result.

This thesis consists of seven chapters including this chapter as it is organized in the following
chapters. Chapter 2 provides the review of food security concepts, definition, measuring
methods, and determinant factors. The statistical models that can serve for multivariate,
longitudinal and multivariate longitudinal ordinal outcomes are also discussed. In Chapter 3, we
describe the data used for this thesis including source and type of data, sample size and sampling
procedures, data collection procedures and methods and food security measuring methods.

The methods employed for modelling the determinant factors of food security are discussed in
Chapter 4. In this chapter, we presented a pair copula construction based marginal cumulative
logit model for multivariate, longitudinal and multivariate longitudinal outcomes to estimate the
dependence between the ordinal outcomes and their respective determinants. We developed
algorithm to select appropriate bivariate copula families to represent the dependence measures in
the model.

In Chapter 5, we presented the analysis household data using the methods presented in Chapter
4. It consists of the findings of the pair copula construction-based multivariate ordinal
cumulative logit model for assessing the dependence between food availability, accessibility and
utilisation, and their respective determinants. It also presents the findings of the pair copula
construction-based longitudinal ordinal cumulative logit model for assessing the dependence
between food security status of households in the three rounds and the respective determinants.
Lastly, the findings of the new population-average pair copula construction approach for
multivariate longitudinal ordinal data to measure the dependence between food security
dimensions, the stability over time and the respective determinants simultaneously are presented.

In Chapter 6, we present the discussion for the findings presented in Chapter 5. In Chapter 7, we
present the contribution of this thesis for statistical methods and future works for further
developments. We ended the thesis with the questionnaire, simplification of the full maximum
likelihood and log-likelihood functions of the developed models as appendices.

7|Page
© Yimam JA, UNISA 2019
 Chapter Two

Literature Review

The issue of food security can be addressed using either of the three models including
multivariate ordinal models, longitudinal ordinal models and multivariate longitudinal models.
The statistical models employed so far for multivariate ordinal outcomes, longitudinal ordinal
outcomes and multivariate longitudinal ordinal outcomes were reviewed as follows.

Multivariate Ordinal Models

Among different previous studies for multivariate ordinal data, (Gange, 1994) developed
Generalized Estimating Equations (GEE) methods for correlated ordinal responses which
extends the model developed by (Liang and Zeger, 1986) for correlated binary data. GEE method
is still the applied model with different versions.

On the other hand for performing multivariate ordinal data analysis, Structural Equation
Modelling (SEM) with two stage methodology Maximum Pairwise Likelihood (MPL) -
Generalized Least Squares (GLS) method was developed (Liu, 2007). Large sample simulation
studies of this method showed that the parameter and standard error estimates, and the test
statistics are acceptable. However, standard error formulae underestimate empirical variability
for small sample size less than 200.

Moreover, a multivariate non-linear model for ordinal responses was also developed (Aspelund,
2002). In this model, a linear- by-linear log linear model with independent estimation approach
(IEA) along robust standard errors was used. IEA performed well if only the marginal parameter
estimates were of interest.

Multivariate ordered probit model with pairwise likelihood inference was employed for
multivariate ordinal responses in the continuous latent variable model (Kenne Pagui and Canale,
2016). The model was applied in PLordprod R package and found that the model reduced the
computational problems related to the calculation of a q dimensional integral for each single

8|Page
© Yimam JA, UNISA 2019
likelihood using full likelihood. However, the model is unable to compute both the mean of the
latent variables and the first threshold.

The composite likelihood methods with a latent variable specification were also applied using
both the probit and the logit link functions for multivariate ordinal regression model (Hirk et al.,
2018). The model was executed in mnormt R package; both link functions resulted into recovery
for high correlation parameters whereas for low correlation both of the link functions were not
recovered.

In line with the existing methods, copulas have been popular tools for modelling multivariate
outcomes since copulas have several attractive properties. The first attractive properties of
copulas are allowing us to construct separately the dependence structure from the joint
dependence structure and the marginal probabilities. The second properties are invariant, under
continuous and increasing transformations. The third one is, unlike correlation, they do not
require elliptically distributed for the marginals. Lastly, they can be used to measure tail
dependences of the joint distribution (Syring, 2013) . “A copula is a function which joins a
multivariate distribution function to its one-dimensional marginal distribution functions (Nelsen,
2007).” With a closed form, MLE is straightforward for copula functions. For m-dimensional
data, the probability mass function can be computed using 2m finite differences of the copula
function. As a result, the approach is computationally intensive, and becomes infeasible for high
dimensional problems (Panagiotelis et al., 2012).

Another extension of copulas that have attractive properties for discrete data have been
developed. This extension is called elliptical copulas, in particular Gaussian copulas that can
capture both positive and negative dependence under closed marginalization. As many elliptical
copulas, including the Gaussian copula, cannot be written down in closed form, MLE through
taking finite differences is not a feasible option (Panagiotelis et al., 2012). In a similar vein, to
estimate models based on Gaussian copulas, Bayesian methods have been used (Pitt et al., 2006).
In general, both Frequentist and Bayesian techniques discussed above are computationally
intensive, and may not be applied easily to higher dimensions (Panagiotelis et al., 2012).

Among copula methods applied for multivariate ordinal data, a multivariate ordered logit
regression with the notion of multivariate copula was modelled (Dardanoni and Forcina, 2008).

9|Page
© Yimam JA, UNISA 2019
This model describes how the joint distribution of a set of ordinal response variables depends on
exogenous regressors. The nature of the main properties of the marginal parameterization and the
global interaction copula was found to be nonparametric. The model was found to be an efficient
model for estimation purpose for small response variables.

In response to these challenges, a pair-copula construction (PCC) method has been developed for
multivariate copulas using only bivariate copulas. PCC was originally developed for continuous
random variables and then extended for discrete random variables (Panagiotelis et al., 2012,
Syring, 2013). The advantages of this approach are one PCC provide a highly flexible framework
for constructing copulas exhibiting a wide range of dependence characteristics. Second, the
computation of the probability mass function for a discrete PCC only requires the evaluation of
2m (m − 1) bivariate copula functions, whereas the multivariate copulas requires 2m for
evaluation. As a result, MLE is feasible even for higher (Panagiotelis et al., 2012).

Longitudinal Ordinal Models

GLM extended to the longitudinal setting in two types of generalizations includes subject-
specific and marginal models (Fitzmaurice et al., 2009; Koper and Manseau, 2009). The subject-
specific models are the class of generalised linear mixed models (GLMM) that consider the
association between ordinal outcomes within a subject by treating some of the model parameters
as random variables. Maximum likelihood (ML) estimation method is commonly used to
compute the fixed and random effect parameters. The consequence of random effect in
longitudinal ordinal outcome is that the association is always positive and interpretations of the
fixed parameter estimates for the population of subjects are not straightforward (Fitzmaurice et
al., 2009).

Among different previous studies for analysing clustered data with ordinal responses in the
GLLM classes, mixed-effect model was introduced for the first time by (Harville and Mee,
1984). The estimates of the random effects were approximated by Taylor series expression. This
was also advanced for parameter estimation purpose through numerical quadrature method with
one random effect by (Jansen, 1990).

The marginal model (population-average) interpretation is not obvious owing to the complication
of integrating out the random effects since generally assumed as normally distributed. To

10 | P a g e
© Yimam JA, UNISA 2019
overcome these issues, different random effect models were proposed. The random effect model
that captured the limitation in marginal model was developed which is the mean response
depends only on the fixed-effect and not the individual effect with complex correlation structure
(Tutz and Hennevogl, 1996). Furthermore, a maximum likelihood utilizing quasi-Newton
algorithms with Monte Carlo integration of the random effects was developed for the random
effects of the longitudinal ordinal data (Lee, 2008). The model can be executable in any software
which have independence proportional odds model (IPOM). However, the models were
particularly useful in longitudinal analyses with a moderate to large number of repeated
measurements per subject.

In contrast with GLMMs, marginal models consider the association between ordinal outcomes at
population level. In the class of marginal models, Fitzmaurice and Laird (1993) developed a
marginal model or population-averaged model with maximum likelihood (ML) approach for
repeated ordinal responses that captured only the mean response effects on the particular
specified predictors of interest but not on the individual effects (Fitzmaurice and Laird, 1993).
Since ML approach for fitting marginal model is awkward, GEE was developed as alternative
method for the first time in place to fit marginal model through cumulative logit (Lipsitz et al.,
1994). GEE provides consistent estimators for the regression parameters when the model has
been correctly specified even if it has the limitation of treating the association structure as
nuisance parameters. Moreover, GEE2 was also developed for modelling the association
structure using global odds ratio, while in marginal model, it is considered as a nuisance
(Heagerty and Zeger, 1996). However, the interpretation of the association structure is difficult
for ordinal outcomes because GEE2 does not lead to a multivariate distribution for the ordinal
outcomes and thus complicates the interpretation of association structures. On the contrary, to
improve of the existing models during the century, Perin (2009) developed a model called
“alternate formulation of alternating logistic regressions model” using orthogonalised residuals
to consider the association structure in marginal models for longitudinal ordinal data. Similarly,
an alternative logistic regression (ALR) was proposed to provide insight and some advantages in
the marginal model estimated via GEE and subject-specific models estimated via GLMMs
(Bhatnagar et al., 2015). The model was executed in SAS/STAT version 9.3 and the model
behaved similarly to marginal models estimated via GEE for mean effects. However, it was
difficult to ascribe clustering to the correct level, particularly for ALRs.

11 | P a g e
© Yimam JA, UNISA 2019
To improve the association structure different forms of GEE have been proposed. Of these
among the recent one, local odds ratios parameterization structure executed in multgee R
package (Touloumis et al., 2013) and weighted score methods executed in weighted Scores R
package (Nikoloulopoulos, 2017) are found. The latter allowed latent correlation structure for the
selection of the correlation structure not restricted to an exchangeable or unstructured one and
reduces the computational challenge for large dimensions.

Alternatively, Nooraee et al., (2016) developed an approximate marginal logistic distribution

model for the analysis of longitudinal ordinal data that can accommodate majority of the
limitations in the existing models. The model was executed in existing packages in R and
provided comparable interpretation with GEE. Moreover, the model can be applied without
having to use additional analysis such as multiple imputation over the other methods (GEE) if
incomplete outcome data fulfils the ignobility assumptions and sample size are not too small
(n>=100). However, the model expected to be sensitive to strong deviations from the
multivariate t-distribution for latent variables for estimation of the correlation coefficient.

In line with the existing methods, among recently developed models for modelling repeated or
longitudinal outcomes, the most popular tools are copula. A copula model with bivariate copula
function was one of the presented model-to-model repeated ordered categorical data
(Vandenhende and Lambert, 2000). In this model, the standard cumulative regression models
were used to model the marginal distributions and the copula function was used to model the
dependence between repeated responses. Even though, the copula models addressed both the
marginal parameters and association structures in repeated ordinal data as well, it might not be
suitable to quantify dependence over the bounds.

Another extension of copulas that have attractive properties for longitudinal ordinal data has
been developed. This extension is the multivariate ordered probit model on the basis of
multivariate copula representation for obtaining the maximum likelihood estimates of the
parameters of longitudinal ordinal model (Kurada, 2011). The model was executed in Mprobit
package in R software and the result showed that the model was computationally challenging to
implement it.

12 | P a g e
© Yimam JA, UNISA 2019
In response to these challenges, a pair-copula construction (PCC) method has been developed for
multivariate copulas using only bivariate copulas. PCC was originally developed for continuous
random variables and then extended for discrete random variables (Panagiotelis et al., 2012,
Syring, 2013). The advantages of this approach are one PCC provides a highly flexible
framework for constructing copulas exhibiting a wide range of dependence characteristics.
Moreover, the computation of the probability mass function for a discrete PCC only requires the
evaluation of 2m (m − 1) bivariate copula functions whereas the multivariate copulas requires 2m
for evaluation. As a result, MLE is feasible even for higher and it was executed in R software
(Panagiotelis et al., 2012).

Multivariate Longitudinal Ordinal Models

The statistical analysis of multivariate longitudinal ordinal data for assessing the changes across
time can be addressed either by reducing the dimension of the multivariate longitudinal data to
univariate longitudinal data using some kind of summary measures, or jointly addressing the
associations/dependencies across multivariate covariates and the changes across time points. A
statistical review for the first part was produced in different literatures; Verbeke et al., 2014).
Several approaches for jointly modelling multivariate longitudinal data have been proposed in
the statistical literature that includes three main classes. These are the subject- specific (random
effect) models, the marginal (population-average) models and full specification of the
multivariate distribution for the outcomes (Copula models).

Among previous studies for multivariate longitudinal ordinal outcomes, majorities of them were
concentrated around random effect models. One approach of this model in the context of item
response theory that can handles three-level multivariate ordinal outcomes in longitudinal
settings and can accommodates multiple random subject effects was developed using iterative
Fisher scoring solution for estimating all required parameters and their corresponding standard
errors (Liu, 2008; Liu and Hedeker, 2006). The model was implemented in the GAUSS language
(GAUSS 3.6). Another approach for this random effect models in the context of latent variable
model was also developed to account for the correlation between the time points using item-
specific random effects with a full information MLE method (Cagnone et al., 2009). A
FORTRAN program was written to implement the model. The models can be extended in many
different directions but more difficult to be implemented computationally.

13 | P a g e
© Yimam JA, UNISA 2019
Similarly, in the setting of random effect models, another model was also developed using
subject-specific model (random intercept models) for the longitudinal part and conditioning these
random effect models to account for the repeated independent cross-sectional outcomes. One
extension of this model was also developed that can generate multivariate correlated random
effects across the repeated cross-sectional outcomes from the subject-specific random effects
model which varies across each cross-sectional outcome (Choi, 2012, Verbeke et al., 2014).
Furthermore, another extension of random effects models was developed to relax the
independence assumption on the conditional distribution given random effects by introducing
some type of latent variable models (Laffont et al., 2014). The model was implemented in R
package with probit mixed effects model with a latent variable interpretation. The authors
pointed out that the model worked well for their application data even if the probit model offers
less flexibility than other (logistic) models, requiring only a limited number of parameters to be
estimated. Therefore, extensions of the model can be considered to accommodate more complex
situations.

Even though random effect models have provided many advantages especially to compute
correlations among outcomes through random effects, they have lack of a population-averaged
interpretation for normally distributed outcomes and some computational challenges (Abegaz et
al., 2015, Nooraeea, 2015).

In line with random effect models, very little marginal models were developed for multivariate
longitudinal categorical or ordinal outcomes with the hope to resolve the limitation of
population-average interpretation in random effects models. One of the alternative model that
helps for parameter estimation for non-normally distributed continuous or categorical data was
the quasi least squares (Chaganty and Naik, 2002). Other marginal model approaches that
tailoring GEE was proposed for measuring multicity measured ordinal outcomes (Huang et al.,
2002) since GEE was implemented for binary and time-to event outcomes through combining
two GEE models for the two outcomes, using an autoregressive-type working correlation matrix
for the intra- and inter-outcome dependence over time (Rochon, 1996). Furthermore, Gray and
Brookmeyer (2000) proposed multivariate longitudinal models for continuous and discrete/time-
to-event response variables using GEE approach the popular population-average interpretation
model. GEE approach still has limitation in modelling multivariate longitudinal outcomes since

14 | P a g e
© Yimam JA, UNISA 2019
it treats association as nuisance and measures it using working correlation. Another marginalised
bivariate model using Kronecker product (KP) covariance structure to capture the correlation
between processes at a given time and the correlation within a process over time (serial
correlation) for bivariate longitudinal ordinal data was employed (Lee et al., 2013). The model
was implemented in R package but limited for only two longitudinal ordinal outcomes.

On the other hand, alternative model in the marginal models using the MLE method on the
context of multivariate t-copula was developed for multivariate longitudinal regression model
for ordinal responses, through a computationally efficient Monte Carlo EM approach (Abegaz et
al., 2015). The computation of the probability mass function for a discrete multivariate copulas
including multivariate t-copula require 2m for evaluation (Panagiotelis et al., 2012).

In the third classes of modelling multivariate longitudinal outcomes, the full specification of the
multivariate distribution for the outcomes was implemented specially using multivariate copula
models. A general framework for modelling multivariate repeated measurements was also
proposed for mixed type of outcomes (Shi and Yang, 2016; Shi and Zhao, 2018). The
longitudinal observation of each response was separately modelled using pair copula
constructions with a D-vine structure. Then the multiple D-vines were then joined by a
multivariate copula. The model was executed in R package using zero inflated Poisson
regression and sequential approach was used for inference purpose.

In line with the existing PCC models, copula-based GLMM models have been proposed for
investigating multivariate longitudinal data with mixed-types of responses by combining
random-effects models and the D-vine copulas (Zhang et al., 2019). The D-vine copula measured
the correlation between multiple responses measured at a given time point. Furthermore, the non-
parametric maximum likelihood method was used instead of specifying the random effects
distribution. The model was executed in R package using c continuous and binary outcomes and
the result showed that the non-parametric models were more efficient and flexible than the usual
Gaussian models. However, the model converged slowly when the number of mass points K is
large. All the copula-based model reviewed so far have not yet been implemented in multivariate
longitudinal ordinal outcomes.

15 | P a g e
© Yimam JA, UNISA 2019
Hence, modelling multivariate longitudinal ordinal outcomes using pair copula construction will
reduce the computational challenges of the probability mass function evaluation. Therefore, this
model can resolve both the population-average interpretations in the random effect models and
computational challenges of the multivariate copulas.

All the literatures reviewed herewith for multivariate, longitudinal and multivariate longitudinal
ordinal models did not mean they are part of the analysis. However, we tried to show the
evolution of the multivariate longitudinal ordinal model developmental process until the recent
year. The thesis established itself on the implementation of the PCC model for multivariate
longitudinal ordinal outcomes for jointly assessing the stability over time, the dependence
between ordinal outcomes and the determinants for each ordinal outcome at the same time.

Since 1970 food insecurity has brought multifaceted problems with multidimensional serious
impacts and became the first debating issue on the development that concerns the whole of
mankind (Abafita and Kim, 2014). Since then, the issue of food security has discussed and
diversified immensely worldwide. The widest acceptable definition of food security was also
acquired during the World Food Submit (WFS) held in 1996, which states that the existence of
food security occurs when “all people, at all times, have physical, social and economic access to
sufficient, safe and nutritious food that meets their dietary needs and food preferences for an
active and healthy life”(Pinstrup-Andersen, 2009).

This definition consists of four important interlinked dimensions, namely, physical availability of
food, economic and physical access to food, food utilisation and stability of the other three
dimensions over time (Abafita and Kim, 2014, FAO, 2014, Napoli et al., 2011). Therefore, the
function of food unavailability, food inaccessibility, inadequate utilized food and instability of
food availability, accessibility and utilisation over time at household level, resulted in household
food insecurity (Etana and Tolossa, 2017).

Physical availability of food: The availability refers to the physical existence of food. It
addresses the “supply side” of food security and be it from own production or on the markets so
that the supply is adequate, of appropriate quality, varied and contributes to a healthy diet. On
national level, it is a combination of domestic food production, commercial food imports and

16 | P a g e
© Yimam JA, UNISA 2019
exports, food aid, and domestic food stocks. Moreover, on household level, it could be from own
production or bought from the local markets (Godfray et al., 2010). Different scholars employed
different types of methods to measure availability dimension of household food security status.
Various studies conducted by FAO used ‘‘dietary energy intake” as a measure food security in-
terms of food availability (Coates, 2013). Moreover, the “Months of Adequate Household Food
Provisioning (MAHFP)” has served as a measure of household food security status in-terms of
availability dimension (Carletto et al., 2013, Moroda et al., 2018). On the other hand, both the
total annual household production of corn and bean; and the total annual corn and bean
consumption per capita have served as a measure of household food security status of the
availability dimension (Coates, 2013, Méthot and Bennett, 2018).

Economic and physical access to food: The presence of food availability in the community
does not mean that the household accessed the food so that the accessibility is the demand side of
food security. The accessibility refers to the purchasing power of a household/individual,
infrastructure and existing food price at national or regional level (FAO., 2014, Pinstrup-
Andersen, 2009). The affordability of the food available at the region or community was
evaluated by economic access to food and the deliverability of the available food to all people
who need it was also captured by physical accessibility to food (Assefa, 2015). In general, access
is ensured when all households and all individuals within those households have sufficient
resources to obtain appropriate foods for a nutritious diet because an adequate supply of food
(food production and availability) at the national or international level does not in itself
guarantee household level food security (Assefa, 2015, Carletto et al., 2013, FAO., 2014,
Hunnes, 2013). Various indicators of food access were employed to measure the status of food
security including, the annual net household income (Méthot and Bennett, 2018); the Household
Food Insecurity Access Scale (HFIAS) (Carletto et al., 2013, Coates, 2013, Etana and Tolossa,
2017, Méthot and Bennett, 2018, Moroda et al., 2018); the Household Hunger Scale (HHS)
(Ballard et al., 2011, Méthot and Bennett, 2018); the Months of Adequate Home Food
Provisioning (MAHFP) (Bilinsky and Swindale, 2007, Méthot and Bennett, 2018) and the
Household Dietary Diversity Score (HDDS) (Méthot and Bennett, 2018, Swindale and Bilinsky,
2007).

17 | P a g e
© Yimam JA, UNISA 2019
Food Utilisation: Proper utilisation of food is very important because the availability and
accessibility of food do not reflect the appropriate food security situation (Assefa, 2015).
Utilisation is commonly understood as the way the body makes the most of various nutrients in
the food (FAO., 2014). It is directly linked to a safe and adequate diet, water availability and
quality, sanitation systems, and is influenced by water-borne, food-borne, vector-borne, and
other infectious diseases (Hunnes, 2013, Pinstrup-Andersen, 2009). In addition, sufficient energy
and nutrient intake by individuals is the result of good care and feeding practices, food
preparation, and diversity of the diet and intra-household distribution of food (FAO, 2008,
Hunnes, 2013, Pinstrup-Andersen, 2009). In general, it is socio-economic and biological aspects
of food. The composite score of the “Household Dietary Diversity Score (HDDS)” and the
facilities in-terms of access and use of water supply, sanitation, and hygiene (WASH) were
suggested as a measure of food security indicators in the dimension of utilisation (Carletto et al.,
2013, Moroda et al., 2018). On the other hand, Consumption Score (FCS) was employed as a
measure of food utilisation dimension (Méthot and Bennett, 2018).

Stability: Stability depends on local and regional food production (food availability) and on the
reliability and price of food imports (food access) (Cohen and Garrett, 2010, Hunnes, 2013). On
the other hand even if one’s food intake is adequate today, one is still considered to be food
insecure if she has inadequate access to food on a periodic basis, risking a deterioration of your
nutritional status (Hunnes, 2013). Therefore, stability depends on the availability, access and
utilisation dimensions of food security (FAO., 2014).

Different types of food security measurements were employed for different purposes. We
reviewed recently employed food security measurements and the corresponding statistical
models employed to determine associated factors of food security. The most used food security
measuring tool and statistical models to determine its associated factors are “Household Food
Insecurity Access scale (HFIAS)” and multivariable logistic regression model, respectively.
HFIAS is used to assess the household food security status in-terms of food access (accessibility
dimension) and effects to action since 2007 (Coates, 2013).

18 | P a g e
© Yimam JA, UNISA 2019
Several studies were conducted in Ethiopia that revealed different factors affecting household
food (in)-security status using HFIAS and multivariable logistic regression model. First we
reviewed the recent studies implemented using both of the two methods together and continues
for other methods. All the researches reviewed herewith in this thesis that were measured by
HFIAS method assessed the household food security access dimension. The study conducted in
Addis Ababa by (Birhane et al., 2014) showed that lower monthly income of the household,
household headed by uneducated household heads, daily labourers, and government employees
were more likely to have higher food insecurity whereas households living in government rental
houses were less likely to be food in-secured.

A community based cross-sectional study conducted in Farta District, Northwest Ethiopia

indicated that households headed by females, lack of education, large family size, few or absence
of livestock, absence of income from off-farm activities, lack of irrigation and lack of perennial
income were identified as associated factors for food insecurity (Endale et al., 2014). Similarly,
(Motbainor et al., 2016) conducted a community based comparative cross-sectional study in east
and west Gojjam zones of Amhara Region and the results revealed that five or above family size,
non-merchant women, household monthly income less than 560 ETB, illiterate mothers, rural
residential area, highland agro-ecology and lack of livestock were positively affects household
food insecurity. Moreover, (Shone et al., 2017) conducted a community-based cross-sectional
study in West Abaya District, Southern Ethiopia and the results indicated that households headed
by female, households headed by persons aged >65 years, households with larger family size and
owning smaller farm land were increased the risk of being food in-secured.

On the other hand, the study conducted in Addis Ababa and Arisi Zone of Oromia Region by
(Etana and Tolossa, 2017) showed that lower education status, poor economic status,
unemployment status and study sites provided a statistical significant effect for households to be
food in-secured. As studied by (Tantu et al., 2017) in Wolaita Sodo Town, the result indicted that
single household head, greater than two dependent members, households headed with daily
labourers, lower monthly income and low monthly food expenditure have positive and
significant relationship with food insecurity. Moreover, (Abegaz, 2017) analysed the pooled data
of the sixth and seventh round of the Ethiopian Rural Household Survey (ERHS) using binary
multivariable logistic regression model revealed that rain shock, lack of off-farm income, and

19 | P a g e
© Yimam JA, UNISA 2019
region of the households were negatively associated with food security. Moreover, as studied by
(Abdu et al., 2018) in Assayita district in Afar region through multivariate regression models
revealed that age, parity, and having >2 children below five years of age were statistically
associated variables with household food insecurity.

The “Household Dietary Diversity Score (HDDS)” was also used to determine food security
status as proxy indicator. (Moroda et al., 2018) conducted a study in Ethiopia using logistic
regression models. They found that low educational status, small farmland size, small total
annual income, far distance from health facilities, access to irrigable land, far distance to road
transport, far distance to input/output markets, frequent drought and the in-availability of
supporting organizations were positively associated with household food in-security situation in
the utilisation dimension. This paper addressed both the determinants of household food
accessibility and utilisation dimensions. Moreover, a study conducted in South Africa by (Ngema
et al., 2018) using binary logistic regression revealed that education, and receiving infrastructural
support (irrigation), positively influenced the food security status of households. However,
household income and access to credit showed a negative correlation. The work addressed the
general outlook of food security. It does not indicate which food security dimension was
addressed.

The “coping strategies index” has served as a means of proxy variable to measure food security
status. (Napier et al., 2018) conducted a study in Durban, South Africa using logistic regression
model. They found that larger household size, households spending between R700 and R900 on
food monthly and households purchasing food from street vendors or informal community shops
were indicators of food in-security.

The “calorie intake” tool was also employed in Pakistan as a proxy food security indicator as
conducted by (Ahmed et al., 2017) using binary logistic regression model. The results showed
that family size, monthly income, food prices, health expenses, the market accessibility factors
(road distance and transportation cost) and debt were identified as the main factors influencing
the food security status of rural households. This work provided an input in the household food
availability dimension.

20 | P a g e
© Yimam JA, UNISA 2019
The study conducted in the Sekyere-Afram Plains District of Ghana using both “USDA
Household Food Security Scale” and a binary logit model revealed that households headed by
unmarried people, large household size, small farm size, absence of off-farm income generating
activities and farmers without access to credit were identified that leads households to be food
in-secured (Mensah et al., 2013; Zeray, 2017). The work of Mensah et al. (2013) addressed the
deteminats of food security intermis of household food access. Similarly, (Kelly and Pemberton,
2016) conducted a study in eastern rural area of Grand Bahama Island with the same procedure
revealed that higher educational level of household head, high monthly income, and access to
community gardens were statistically significant predictors for food security.

Habyarimana (2015) conducted a study in Rural Households in Rwanda using “Food

Consumption Groups Score” and probit model. The study revealed that rural households headed
by females, large household size, limited household's farm animal, and small household asset
index, were significant variables for household food security. Moreover, the study also found
that limited household food acquisition level, large household food acquisition problem, small
amount of household spending level, small amount of monthly food expenditure, small percent
of land suitability per cell, large amount of soil erosion index per village, reduced coping
strategy index and being membership to agricultural cooperative were significant variable for
household food in-security (Habyarimana, 2015). This research addressed the determinants for
household food accessibility dimension.

The study was conducted to determine predictors of household food security in-terms of food
access in Mexico using the “Mexican Food Security Scale” and Ordered probit model (Magaña-
Lemus et al., 2016). They indicated that households include those with younger, less-educated
household heads, headed by single, widowed or divorced women, with disabled household
members, with native language speakers, with children, as well as rural and lower-income
households were more likely to be food insecure.

Bashir and Schilizzi (2012) conducted a meta-analysis that showed education level, household
head’s age, input availability, technology adoption, farm size, land quality, price of inputs, and
credit were associated with household food security of the availability dimension. On the
contrary, income, distribution of income within the household, household size, total earning
members, and family structure were associated with the access dimension of household food

21 | P a g e
© Yimam JA, UNISA 2019
security. Moreover, gender and expenditure on food and health are considered as determinants of
utilisation aspect.

The studies reviewed so far addressed a single dimension among the four dimensions of
household food security. The next review intended on the composite multidimensional index of
food security. The most known composite index of food security analysis was principal
component analysis index (PCAI). Abafita and Kim (2014) employed PCAI to compute the
composite food security index of food availability, accessibility and utilisation in Ethiopia. An
instrumental variable (IV) regression models using 2-Stages Least Square (2SLS) was applied to
select the significant predictors and the findings indicated that participation in off-farm activities,
education of the household head, household size, livestock possession, rainfall index, fertilizer
use and per capita consumption expenditure were statistically significant determinants with
positive impact on household food security. On the contrary, remittance and credit access had a
negative and statistically significant impact on household food security (Abafita and Kim, 2014).

Similarly, (Mbolanyi et al., 2017) followed the procedure of (Abafita and Kim, 2014) for the
study conducted in rangeland area of Uganda using Ordinary least square (OLS) and the result
indicated that age of the household head, male household head, On-Farm Income and household
head level of education (second degree or above) positively affected the food security of
households. On the other hand, as studied by (Wineman, 2016) in rural Zambia on three food
security components of the households (food quantity, food quality and food stability) using
multinomial logistic regression. The author found that both rainfall and temperature have a
significant impact on a household’s food security score.

The situation of food security is very difficult as it is the result of complex interaction between
numerous variables. For instance, some of the food access proxy indicators have served as food
availability proxy indicators and vice versa. Similarly, some of food access proxy indicators have
served for utilisation proxy indicators. This implies that understanding several concepts
associated with the definition of food security are necessary before examining the determinants
of food security.

The composite multidimensional index of food security conducted so far did not consider the
contribution of each dimension for the determination of household food security status.

22 | P a g e
© Yimam JA, UNISA 2019
Similarly, previous studies did not address the associated determinate factors for each dimension
at a time in a single model. The dependence between the dimensions and the respective
predictors for each dimension did not undertake in a single model simultaneously.

These gaps can be seen in three ways taking all the dimensions together. First, the dependence
between food availability, accessibility and utilisation and the predictors for each dimension can
be addressed in the statistical models of multivariate ordinal data analysis. Second, the stability
of the composite multidimensional index of food security of the three dimensions can be
addressed using the statistical models of longitudinal ordinal data analysis. Lastly, the stability of
the three food security dimensions over time and predictors for each dimension can be addressed
using multivariate longitudinal ordinal data analysis.

23 | P a g e
© Yimam JA, UNISA 2019
 Chapter Three

Data

This study was conducted in South Wollo Zone, one of the 11 zones in the Amhara Region State
of Ethiopia. South Wollo is located in the North East of Ethiopia with latitude and longitude of
11°07'59.99" N 39°37'59.99" E. Dessie is the capital of the zone which is 401 kilometres away
from the capital city of Ethiopia, Addis Ababa. South Wollo has a population of 2,518,862, of
whom 50.4% and 49.6% are women and men, respectively. The largest ethnic group of the zone
was Amhara which accounted for 99.33% of the total population. Moreover, 70.89% of the
population subscribe to Muslim religion, and 28.8% were practising Ethiopian Orthodox
Christianity (CSA, 2007).

South Wollo has 18 rural Woredas and two urban Woredas. Each Woreda has Kebeles which is a
smallest unit in the administration of the zone. The target group for the thesis is households who
are farmers living in the rural Woredas. This choice will minimise the error which will come
from the heterogeneity lifestyle of the households because the living style and sources of food
security dimensions are the same.

The quantitative data collection instrument was developed through extensive review of similar
literatures and recent studies (Ballard et al., 2011; Bashir and Schilizzi, 2012; Carletto et al.,
2013; Castro, 2000; Coates, 2013; Cohen and Garrett, 2010; Godfray et al., 2010; Hunnes, 2013;
Napoli et al., 2011; Negatu, 2004; Ryu and Bartfeld, 2012; Biesalski et al., 2017; de Bruin and
Gresse, 2018). Data were collected by using semi-structured questionnaire, which allowed study
participants to express more additional information and their opinions.

The questionnaire has five parts; the first part covers area identification, the second covers
demographic and socio-economic characteristics of the households, the third part covers farming
activities in relation to agriculture activities, environmental and climate change conditions. The
fourth part intended on the information related to food security status of the households in each

24 | P a g e
© Yimam JA, UNISA 2019
dimension includes Availability, accessibility, utilisation, and the stability of the three
dimensions over time. The availability dimension questions were customised from several
researchers (Carletto et al., 2013, Coates, 2013, Godfray et al., 2010). The questions for
accessibility used in this study was the Household Food Insecurity Access Scale (HFIAS)
(Swindale and Bilinsky, 2007). The utilisation tool was obtained from (Carletto et al., 2013,
Faber et al., 2009). The last part covers the copying strategy that the households applied to
overcome the hardship and crises of food security. The questionnaire is first prepared in English
and then translated into Amharic (local language of the respondent’s) attached in appendix I.

Administratively, Ethiopia is divided into 11 regions. Subsequently, regions are divided into
zones. Similarly, zones are further divided into Woredas, the smaller administrative unit. Each
Woredas is further subdivided into the lowest administrative unit called Kebele. For the current
study, South Wollo zone from Amhara region was selected. In this area, the food security
situation has yet not been updated for the last 15 years after (Castro, 2000, Negatu, 2004). South
Wollo zone has 18 and 2 rural and urban Woredas, respectively. The rural Woredas are assumed
to have uniform agro-ecological and homogenous in cultivation strategies. Hence, three-stage
sampling procedure is the ideal sampling methods. This implied that sample of primary units
(Woredas) were selected from the total rural Woredas of south Wollo Zone, then sample of
secondary units (Kebeles) were chosen from each of the selected primary units (Woredas) and
finally, sample of tertiary units (households) were chosen from each selected secondary unit
(Kebeles). Hence, three rural Woredas were determined as optimal sample size using the
ordinary cluster sampling formula as

2
(𝑍𝛼/2 +𝑍𝛽 ) 𝑀𝑉 2 (1.96+0.84)2 (18)(0.001)
𝑚= 2 = = 2.867 ≈ 3.
(𝑍𝛼/2 +𝑍𝛽 ) 𝑉 2 +(𝑀−1)𝑑2 (1.96+.84)2 (0.001)+(18−1)(0.05)2

2
Where (𝑍0.05/2 + 𝑍0.2 ) = (1.96 + 0.84)2 at 5% level of significant and 80% power, 𝑀 = 18 is
the number of rural Woredas in South Wollo zone, 𝑑 is the degree of precision and taken to be
0.05 and 𝑉 2 = 0.001 is the ratio of the variance of the error term and the variance of the food
security proportion 𝑃 = 0.60 of the study conducted in Guraghe zone, Southern Ethiopia
(Nigussie and Alemayehu, 2013).

25 | P a g e
© Yimam JA, UNISA 2019
These three Woredas were selected using simple random sampling (in particular lottery method)
from 18 rural Woredas. Using the same fashion, a total of six Kebeles, 2 for each were selected
from the selected Woredas. The sampling procedure conducted in this research was a three-stage
sampling design. Woredas, Kebeles and households were the first, second and the third stage of
sampling, respectively.

Hence, there were three sample units in the first stage, and six in the second stage. For the third
stage, a complete list of household heads was obtained in each of the six selected Kebeles from
agriculture agent office of each Kebele. The determined sample was proportionally allocated for
each Woreda and then Kebele. Based on the allocated sample size, households included in the
sample were selected using systematic random sampling technique from those representative
Kebeles. Then a list of names of the sampled households was prepared for each Kebele.

Sample size determination is a very crucial task because a huge sample costs money and a small
sample reduces the power of estimation. Hence, during the determination of required sample size
issues/points one has to consider are objective of the research, design of the research, cost
constraint, degree of precision required for generalization, etc. Based the above information,
several sample size calculation formulas were developed that conform to different research

Z 2 p(1 − p)
situations. Accordingly, the sample size determination formula n = (Cochran, 2007)
d2
is adopted for this study since the target population is reasonably large. Where Z is the upper

2 points of standard normal distribution with  =0.05 significance level, which is Z =1.96.
The degree of precision d is taken to be 0.05. The parameter p represents proportion of food
security of household. P=0.60 is used in this study obtained from previous study in rural areas of
Guraghe Zone, Southern Ethiopia (Nigussie and Alemayehu, 2013). Accordingly, the sample
size using the given formula becomes n=369. Five percent of the sample size, which is 19, is
added to the determined sample size 369 to compensate for non-response rate and the sample
size becomes 388. Since the sampling design is multistage, 1.75 times of the sample size should
be taken to compensate the design effect. Therefore, the required sample size for the study
becomes n = 646. Next, based on these 646 farmer households, the following sample size

26 | P a g e
© Yimam JA, UNISA 2019
allocations were employed based on proportional allocation for the selected districts and then to
Kebeles as presented in Table 3.1.

Table 3. 1 Sample size allocations for the selected Woredas and then to selected Kebeles
with-in the respected Woredas

Total farmer Allocated Total farmer Allocated

Woredas households Sample size Kebeles households Sample size
Alansha (03) 1245 139
Kutaber 27, 443 210 Beshilo (06) 639 71
Kedida (07) 706 89
Kalu 31,693 235 Degan (019) 1150 146
Bededo (01) 1973 108
Tehuledere 27, 241 201 Jari (017) 1692 93
Total 87, 377 646 7405 646

The data collection process has three phases. Each phase has similar procedures to be
undertaken. Data collection was carried out using trained data collectors and data collection
supervisors under the direct supervision of the researcher who worked closely with them. All the
data collectors and data collection supervisors were trained on sample design, survey technique,
survey instruments, and confidentiality protocol both for the pre-test and main data collection.
This was necessary to ensure a common understanding of the whole survey in order to reduce
interviewer biases as much as possible. Both data collectors and data collection supervisors were
agricultural extension workers with a minimum of diploma for data collectors and Bachelor of
Science for supervisors who speak English and local language Amharic.

Before administering the questionnaires, the questionnaires were pre and pilot tested for the
purpose of insuring the questions were clear and understood by the study participants. The pilot
test fieldwork was conducted over half a day in one Kebele which is out of the selected Kebeles
from 50 household heads. The pilot test field staff and the investigator made thorough

27 | P a g e
© Yimam JA, UNISA 2019
discussions. After a while, based on lessons drawn from the pilot test exercise, the questionnaires
were modified.

During the data collection, the data collectors approached the sampled household heads and
requested their willingness to be take part of the study before starting the interview. Only
consented household heads were interviewed face-to-face that took place usually outside of the
house in the compound. But if the head was not present or available, then the spouse or an adult
household member aged 18 or more and live more than six months with the family to be
considered as a member of that household was interviewed with the same fashion as the
household head. Once data collection ended, the data collectors told the study participants before
leaving about the second phase data collection as they will come after six-month interval. The
third phase data collection proceeded like phases one and two. Each interview lasted on average
30 to 45 minutes. The surveys were carried out in a local language of the household head.

The supervisors were in charge verifying everyday what each data collectors had done (how the
questionnaire was filled in, omission and coherence of answers, and sometimes assisted in
interviews). This was very important because enumerators could quickly rectify any mistake that
had occurred by going back to the households to verify the information from their subjects when
it was necessary. The team (data collectors and supervisors) met the principal investigator every
morning for field feedback and every two days for logistical support. The data collection lasted
for 30 working days (exclusion of Sunday) and each interviewer had to administer seven to eight
questionnaires per day.

The identification code was prepared for each household head participated in the study. An
appointment abstraction form was also prepared to trace the name of the household heads and the
study participation code that serves for the six-month follow-up data collection process (for the
second and third round data collection processes). Once the data collection process is
accomplished, the follow-up appointment abstraction form detached from the data collection
questionnaire and placed in the separate place to secure confidentiality issues.

The current study was employed the longitudinal data collection approach. Three rounds of data
collection were employed at six months interval. The main harvest season in most of the study

28 | P a g e
© Yimam JA, UNISA 2019
locations is during the months of June and July. The first round of data collection that assessed
the food security experience for the last 12 months was carried out on June, 2014. The second-
round data collection that assessed the last six months food security experiences was carried out
on December, 2014. Lastly, the third round of data collection that assessed the last 6 months
food security experiences was employed on June, 2015.

In this thesis, different household food security measuring methods were employed with the hope
to have single index for availability, accessibility and utilisation. Moreover, a composite index
was also employed to assess the stability of household food security status over time.

Measuring Household Food Security Status of the Availability Dimension

Different scholars or organisations employed different methods and food security classification
for this dimension. The “Dietary Energy Intake” method is used to determine food security status
as food secured or in-secured (Coates, 2013, FAO., 2014). The “Months of Adequate Household
Food Provisioning (MAHFP)” is used also to determine least food insecure, moderately food and
most food insecure (Carletto et al., 2013, Moroda et al., 2018). The median score is used to
determine food security status, those below median score as food in-secured and above median
score as food secured (Kisi et al., 2018). Based on the recommendation obtained from (Capaldo
et al., 2010), we expanded the work of Kisi (Kisi et al., 2018) using the quarter score approach to
determine food security as “food secured”, “mildly food in-secured”, “moderately food in-
secured” and “severe food in-secured as follows.

Food availability at household level depends on own production or bought from the local
markets (Godfray et al., 2010). Coates (2013) used the total annual household production and
consumption of corn and bean per capita as a proxy measure of household food security status
for the availability dimension.

In the study area, foods like prepared from cereal crops, fruit and vegetables, milk and milk
products, and meat and meat products are more or less consumed from their own production or
from the local market. Moreover, food availability depends on foods provided by food aid
organisations. Twelve (12) questions were developed to assess the availability of the above food

29 | P a g e
© Yimam JA, UNISA 2019
groups from their own production and/or from local markets; and from food aid organisations to
measure the household food security status in terms of availability. We created a summative
scale using these questions and each answer was recoded as 1, 2, 3 and 4 where 1 stands for the
response “enough of the kinds of food we want to eat”, 2 for “enough but not always the kinds of
food we want”, 3 for “sometimes not enough to eat” and 4 for “often not enough to eat”.

Screening questions were used before asking the availability of food groups from their own
production. If a household is not produce a particular food group or some of them, he or she does
not ask about the availability of food group from own production. Similarly, a household obtains
“enough of the kinds of food he or want to eat” from own production is not asked about the
availability of that food group from local market. This implies that the number of question for
each household may not be equal and may be less than 12 for some of them.

Based on the above criteria, the item responses were summed to compute the score of household
food security status ranging between 12 and 48 points for those asked all of the 12 questions.
This range divided into four equal parts based on quartile score. The scores fall in the range 12-
20 grouped as “food secured”, 21-29 as “mildly food in-secured”, 30-38 “moderately food in-
secured” and 39-48 as “chronically food in-secured”. Similarly, for those asked 11 questions, the
range is 11 and 44 points, for those asked 10 questions, the range is 10 and 40 points and so on
for the other households asked less than 10 questions. The quartile square is applied for each
range to determine the household food security status.

Measuring Household Food Security Status of the Accessibility Dimension

Majority of food security studies relayed the Household Food Insecurity Access Scale (HFIAS)
for measuring household food security access status. Hence, we followed this scale for this thesis
to assess the household food insecurity status in terms of accessibility. The module consists of
nine items that measure the severity of a wide range of food hardships over the past 12 months.
The status of the households was classified into four as availability based on the criteria given in
the module. The cut-off points that serve to place households in a unique category of food
security status is given in Table 3.2.

30 | P a g e
© Yimam JA, UNISA 2019
 Table 3. 2: Cut-off points for household access scale

Frequencies
Questions Rare Sometimes Often
1
2
3
4
5
6
7
8
9
Food secure Moderately food insecure

Mildly food insecure Chronically food insecure

Measuring Household Food Security Status of the Utilisation Dimension

The composite score of the “Household Dietary Diversity Score (HDDS)” and the facilities in-
terms of access and use of water supply, sanitation, and hygiene (WASH) were suggested as a
measure using the three food security categories in this dimension (Carletto et al., 2013, Moroda
et al., 2018). The DDS score was used to determine food security based on the median score
those below the median score classified as food in-secured and above median as food secured
(Faber et al., 2009). Based on the recommendation obtained from Capaldo et al. (Capaldo et al.,
2010), we expanded the work of Faber et al. (Faber et al., 2009) using the quarter score approach
to determine food security as “food secured”, “mildly food in-secured”, “moderately food in-
secured”, and “severe food in-secured” as follows.

Utilisation is directly linked with safe and adequate diet; and access and use of water supply,
sanitation, and hygiene (WASH). “Household Dietary Diversity Score (HDDS)”, which is an
assessment of 12 food groups, can measure the safe and adequate diet. As a result, we developed

31 | P a g e
© Yimam JA, UNISA 2019
19 questions that can address these issues and each answer was recoded as 0 (no) and 1(yes). The
first 12 questions were the HDDS component and the last seven were the WASH component.

As in food availability, the item responses in utilisation were summed to compute the score of
household food security status ranging between 0 and 19 points. The scores fall in the range 15-
19 classified as “food secured”, 10-14 as “mildly food in-secured”, 5-9 as “moderately food in-
secured”, and 0-4 as “chronically food in-secured”.

Composite Food Security Index (CFSI)

The main objective of computing composite index was to determine the stability of household
food security over the successive time periods. Three rounds of data collections were made. In
each round, the household food security was measured for each dimension. To determine the
stability of household food security status over time the food security measure in each round
should be combined into one. We call this a composite food security index. Therefore, the
following approach was made to compute the composite food security index.

In each of the data collection phase, there are three food security measures that have four levels
namely “food secured”, “mildly food in-secured”, “moderately food in-secured”, and
“chronically food in-secured”.

We created a summative scale using the three measures of phase one and each answer was
recoded as 1, 2, 3 and 4 where 1 stands for the response “chronically food in-secured”, 2 for
“moderately food in-secured”, 3 for “mildly food in-secured” and 4 for “food secured”. The item
responses were summed to compute the score of household food security status ranging between
3 and 12 points. This range divided into four equal parts based on quartile score. The scores fall
in the range 3-5 grouped as “chronically food in-secured”, 6-7 as “moderately food in-secured”,
8-9 as “mildly food in-secured”, and 10-12 as “food secured”. The same procedure was made for
phase two and three to compute the combined food security status of the household. The
framework of this computation displayed in the following diagram.

32 | P a g e
© Yimam JA, UNISA 2019
 Phase one Phase Two Phase Three

Accessibilit Utilisation Availability Accessibility Utilisation Availability Accessibilit Utilisation

Availability
y y

Food security Food security Food security

status status status

Figure 3. 1: Framework of food security of the household at three follow-up time points
at six-month interval

Some of the determinant variables for food security are clarified as follows: Shortage of rain fall
is described by the length that the rain rains; Shortage of rainfall happens if the rain stops too late
and/or too early. Since the amount of rainfall depends on the nature of cultivable land, some
type of land may need a high amount of rainfall or some of the lands will need few amount of
rainfall. Hence, the farmers can declare the amount of rainfall happened in their village as per
their type of cultivable land.

The crops/vegetables disease can be described as any type of disease reported considered as
crops/vegetables disease occurred in that area. The type of weather conditions in the study area
(Hot, Medium and Cold) was replaced by type of agro-ecology of the study site.

Cultivation season is the amount of cultivation season per year in their majority cultivable land
(one time, two or more time per year). Moreover, the cultivable land of the study sites
categorised as less than or equal to half hectare and above half hectare. Since majority of the

33 | P a g e
© Yimam JA, UNISA 2019
farmers in the study site have less than half hectare cultivable land. Majority of other previous
studies categorized as we have conducted.

Fertility of cultivable land is categorised as fertile, medium fertile, and less fertile. The concept
of this question was to assess the fertility of majority of cultivable land. The fertility depends on
the nature of the land that cultivated in the rainy season or not, resist during the dry season and
the amount cultivated from that specific cultivable land. Moreover, the data collectors and
supervisors were diploma and BSc in agricultural science, respectively; working around the
farmers elaborated the categories in detail during data collection.

34 | P a g e
© Yimam JA, UNISA 2019
 Chapter Four

Methodology

For the long period of time, statisticians have been searching models that serve for measuring
relationships or associations both in the continuous and discrete analogs of multivariate
distribution. Mainly measuring relationships revolved on the continuous analogs whereas
association on the discrete using bivariate and tri-variate distribution functions with given
univariate margins. Sklar (1959), who develops new class of functions called copulas, gave the
concept of univariate margins for this case. The word copula originates from the Latin word
copulare, which means a link or connect or join. It was used by Sklar (1959) in the theorem for
the first time in a mathematical or a statistical context to describe multivariate distribution
functions that are constructed by joining together one-dimensional distribution functions (Sklar,
1959).

For the first time, copula was introduced to Encyclopaedia of Statistical Sciences in 1997 by
(Fisher, 1997). Fisher introduced the interest of copula to statisticians on the concept of
probability and statistics in this Encyclopaedia for two main reasons. The first reason is to
measure scale free dependence and the second reason is to construct the starting point of
bivariate distributions families (as cited by Nelson, 2007). “Copulas are multivariate distribution
functions whose one-dimensional margins are uniform on the interval (0, 1)”.

Various advancements of the copula functions have been introduced since its introduction for
many applications, especially, multivariate distributions. One of the recently developed copulas
for cascading the multivariate distribution into bivariate distribution with the great accuracy and
efficiency is the pair copula construction (PCC). Pair copula construction was first introduced for
continuous margins (Aas et al., 2009), then extended for discrete margins (Panagiotelis et al.,
2012). The current focus is the applicability of pair copula construction on multivariate
longitudinal ordinal outcomes with the hope of measuring the dependency between longitudinal
ordinal outcomes. This thesis further concerned itself on the implementation of PCC for
multivariate longitudinal ordinal outcomes using the frequentist paradigm because the Bayesian

35 | P a g e
© Yimam JA, UNISA 2019
paradigm requires intensive work in selecting the appropriate prior distribution for the marginal
model especially in ordinal outcomes setting.

In sum, this chapter addresses the basics of copula theory, the application of pair copula
construction for multivariate ordinal outcomes, for longitudinal ordinal outcomes, and for
multivariate longitudinal ordinal outcomes. The chapter also addresses parameter estimation of
all the three models stated above.

Definition and Properties of Copula Theory

Definition 4.1: A multivariate distribution function with uniformly distributed margins that
satisfies the following properties is called an m- dimensional copula C: [0, 1]𝑚 → [0, 1].

i. For every u in [0, 1]

a. C (u1 , u2 , ..., um ) = 0 , if any ui = 0

b. C (u1 , u2 , ..., um ) = 1 , if any ui = 1

ii. For any (a1 , a2 , ..., am ) and (b1 , b 2 , . . ., bm )  [0, 1]m if a j  b j , then

P(U1  [a1 , b1 ], ..., U m  [am , bm ])  0 , we have

2 2

...  (−1)i1 + .. +im C (u1,i1 ,..., um,im )  0

i1 =1 i m =1

where ui ,1 = a j , ui , 2 = b j and 𝑈𝑗 are the random numbers which have uniform margins.
The first property expresses the requirement of uniform marginal distributions whereas the
second property expresses the rectangle inequality. A copula characterizes through these two
properties; meaning that if a function C is fulfilled then it is a copula.

Sklar’s theorem summarized the importance of copula in the study of multivariate distribution
functions. The theorem shows how the copula coupled the univariate marginal distributions to
construct multivariate distributions.

36 | P a g e
© Yimam JA, UNISA 2019
Theorem 4.1-Sklar (1959): For m-dimensional random variables ( y1 , y 2 , . . ., y m ) with

corresponding margin F1 ( y1 ), F2 ( y 2 ), ..., Fm ( y m ) , their joint distribution function

F ( y1 , y2 ,..., ym ) can be expressed in-terms of an m-copula C functions as follows;

F ( y1 , y 2 ..., y m ) = C ( F1 ( y1 ), F2 ( y 2 ), ..., Fm ( y m )) . (4.1)

If all F j ( y j ) are continuous then unique function C is defined, if not, C is uniquely defined

within the product of the regions RanF1 ( y1 )  . . .  RanFm ( y m ) , where RanFi ( yi ) is the range of

the ith distribution function.

Equation (4.1) gives an expression for joint distribution functions in terms of a copula and
univariate distribution functions. But (4.1) can be inverted to express copulas in terms of a joint
distribution function and the “inverses” of the margins. However, if a margin is not strictly
increasing, then it does not possess an inverse in the usual sense. Therefore, we can use “quasi-
inverses” of distribution functions.

Corollary 4.1: Let F be an m-dimensional joint distribution function with margins F1 , F2 , ..., Fm ,

C be an m-copula and let F (−1) be the ith quasi-inverses of F . Then for any u in domain of C ,

( −1) ( −1)
C (u1 , ..., um ) = C ( F1 (u1 ), ..., Fm (um )) . (4.2)

This corollary is the unique copula satisfying equation (4.1). Given marginal and joint
cumulative distribution functions, the above result allows the direct construction of a copula.

In copula theory, there are special dependence structure functions. These are the “Frechet-
Hoeffding upper bound M named the comonotonicity copula”, the “Frechet-Hoeffding lower
bound W named the countermonotonicity copula”, and the “independence copula Π” (Nicklas,
2013). There expressions are given below respectively.

M (ui , ..., u m ) = C (ui , ..., u m ) = min(ui , ..., u m )

W (ui , ..., um ) = C (ui , ..., um ) = max( ui + ...,+um − m + 1, 0)
m
(u i , ..., u m ) = C (u i , ..., u m ) =  u i .
i =1

37 | P a g e
© Yimam JA, UNISA 2019
Note that in arbitrary dimensions, M and Π are copulas whereas W is a copula only in the
bivariate dimensions. Any copula functions are bounded point-wise by the Frechet-Hoeffding
bounds.

Proposition 4.1 For every u ∈ [0, 1]m in any copula function C, the following expression holds.

W (ui , ..., u m )  C (ui , ..., u m ) = M (ui , ..., u m ) (4.3)

In copula theories whenever M, Π, and W are copulas, they have a special interpretation as stated
in detail in (Nelsen, 2007).

Copula Density

The multivariate density f ( y1 , y 2 , ..., y m ) for the continuous case can be obtained through
both sides’ differentiation of equation 4.1 using the chain rule, we have

f ( y1 , y 2 , ..., y m ) = c( F1 ( y1 ), F2 ( y 2 ), ..., Fm ( y m ))  f1 ( y1 ) ... f m ( y m ) (4.4)

where f1 ( y1 ) ... f m ( ym ) are the marginal density of the jth margin and c(.), known as the copula

density, is the copula function differentiated with respect to each of its arguments.

Even if the copula function is not unique for discrete margins, parametric copulas may still be
used to model the dependence between discrete data which provides some evidence that discrete
data inherit dependence properties from a parametric copula like the continuous case. In contrast
with the continuous case, the probability mass function (pmf) for discrete data can be evaluated
by taking differences of the copula function. Without loss of generality assuming Y ∈ Nm (where
N is the set of natural numbers), the probability mass function of Y is given (Panagiotelis et al.,
2012, Nicklas, 2013, Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al., 2015);

 ...  (−1)
i1 +...+im
P(Y = y ) = P(Y1  y1 − i1 , ..., Ym  ym − im )
i1 =0 ,1 im =0 ,1
(4.5)
 ...  (−1) 1
i +...+im
= C ( F1 ( y1 − i1 ), ..., Fm ( ym − im )).
i1 =0 ,1 im =0 ,1

38 | P a g e
© Yimam JA, UNISA 2019
Equation (4.5) is the special class of copula distribution called multivariate copula function for
discrete data, which requires 2m evaluations of the pmf.

Alternately, the pmf for some copula functions that do not have a closed form can be evaluated
by integration over a rectangle. Hence, the probability mass function of the Gaussian copula with
discrete margins can be expressed as;

 1+  m+
P (Y1 = y1 , ..., Ym = y m ) =  ...   m ( 1 , ..., m ; )d 1 , ..., d m , (4.6)
 1−  m−

where  −j :=  −1 ( p(Y j = y j − 1)),  m (., ) and m (., ) respectively denote the cdf

and probability density function of an m-dimensional normal distribution with mean 0 and
variance matrix given by the correlation matrix Γ (Panagiotelis et al., 2012).

Both the multivariate and Gaussian copulas remain a highly challenging computational problem,
especially for higher dimensions. Furthermore, this computational challenge was resolved
through the introduction of vines pair copula construction (PCC) which requires 2m(m-1)
evaluation of the pmf less demanding than 2m of the former one (Panagiotelis et al., 2012, Huynh
et al., 2014, Sirisrisakulchai and Sriboonchitta, 2014). This is a copula-based framework that
effectively simplifies the computational cost of evaluating the pmf and also a large range of
dependence characteristics can be modelled. PCC will be discussed in detail later on section
[Link].

Families of Copulas

Several scholars in their literature have carried out the construction of copula families and their
properties. Here, we present few of the most popular in the literature and use for our purpose as
follows:

Elliptical Copulas

Copulas developed from elliptical distributions are elliptical copulas. The elliptical copulas can
be used to create new multivariate distribution functions by combining arbitrary margins. The

39 | P a g e
© Yimam JA, UNISA 2019
Gaussian and the t-copulas are the most commonly used elliptical copulas. The properties of
these copulas were presented below one by one.

Gaussian Copulas

Let the distribution function of the multivariate normal distribution with zero mean and
correlation matrix P is denoted by Φ1... m and the univariate standard normal distribution is by Φ.
Then the m-dimensional Gaussian copula is defined by

( −1) ( −1)
C Ga (u1 , . . ., u m ) = 1,..., m (1 (u1 ), . . .,  m (u m )) (4.7)

Even if we can express Gaussian copula as an integral, it does not have simple closed form
(Nicklas, 2013). In two dimensions, given that the covariance matrix is non-singular, we get

 −1 ( u1 )  −1 ( u2 ) 1  − ( x12 − 2 px1 x2 + x22 ) 

C (u1 , u 2 ) = 
Ga
 exp  dx1dx2 (4.8)
− −
2 (1 − p 2 )  2 (1 − p 2
) 

The dependence structure can be extracted from the multivariate normal distribution through this
Gaussian copula. We can obtain the independence copula from Gaussian copula if P = I m .

Similarly, the comonotonicity copula can be also obtained if P is an m×m matrix consisting
entirely of ones. In two dimensions, the Gaussian copula with ρ = −1 is equal to the
countermonotonicity copula. Hence, at least in two dimensions, the dependence structure that
interpolates between perfect positive and negative dependence can be thought as dependence in
the Gaussian copula.

40 | P a g e
© Yimam JA, UNISA 2019
The t copula

Let the distribution function of the m-dimensional t distribution with ν degrees of freedom, zero
mean vector and correlation matrix P denoted by 𝑡𝑣,1,2,…,𝑚 and the univariate t distribution by 𝑡𝑣 .
Then the m-dimensional t copula is defined by

Cvt (u1 , ..., u m ) = t v ,1, ..., m (t v−1 (u1 ),..., t v−1 (u m )), (4.9)

In t copula, there is an additional parameter v (degrees of freedom) (Nicklas, 2013). Owing to

this parameter t, copula becomes more suitable for financial application. This parameter controls
the dependence of the extreme events meaning that both extreme positive and extreme negative
events can be modelled equivalently.

Like the Gaussian copula, we can express t copula as an integral, and does not have simple
closed form. In two dimensions, the t copula with ν degrees of freedom has the following form:

v+2
−
t v−1 ( u1 ) t v−1 ( u 2 ) 1  x12 − 2 px1 x 2 + x 22  2
T p , v (u1 , u 2 ) = 
− 
−
1 +
2 (1 − p 2 )  v(1 − p 2 )


dx1 dx 2 .

(4.10)

As in the case of the Gaussian copula, the comonotonicity copula can be obtained if P is an m×m
matrix of ones. However, in contrast to the Gaussian copula, we do not obtain the independence
copula from t copula if P = I m . This is because uncorrelated multivariate t-distributed random

variables are not necessarily independent.

Archimedean Copulas

Among the copula families in parametric dependence modelling, the most popular on is the class
of Archimedean copula. Elliptical copulas have the advantage that simulating from them is easy.
However, they often do not have closed-form representations and they are all radially symmetric.
All Archimedean copulas are flexible in the types of dependence structures they can model, and
have closed-form expressions. Unlike the previously described copulas, Archimedean copulas
are not derived using marginal distributions and Sklar’s theorem, though they are still easy to
construct. The uniqueness of Archimedean copulas are defined by a generating function denoted
41 | P a g e
© Yimam JA, UNISA 2019
by  (Nicklas, 2013). Under a continuous, strictly decreasing, and convex function from I to [0, ∞] of

 with  (1) = 0 , then the Archimedean copula is given by

C (u , v) =  [ −1] ( (u ),  (v)) (4.11)

where  [−1] is the pseudo-inverse of  :

 [ −1] if 0  t   (0)
 [ −1]
(t ) =  (4.12)
0 if  (0)  t  

Since an Archimedean copula can be generated by any continuous, strictly decreasing, convex
function, a huge number of Archimedean copulas have the ability to model a wide range of
dependence structures. For a complete summary of families of one parameter, Archimedean
copulas, refer to (Nelsen, 2007). Moreover, hereunder we give the generator and the selected
bivariate one parametric Archimedean copula functions that we used only for our purpose in this
study.

Clayton copula (Clayton, 1978): given that the copula parameter   [−1, ] \ {0} , the Clayton
copula is given by

1
C (u, v) = (u − 1)
1
−
−
+v −  , with generator  (t ) = (t − − 1).


For  = 0 , we set C =  (Nicklas, 2013).

Gumbel copula (Gumbel, 1960): For the copula parameter θ ∈ [1, ∞), the Gumbel copula is
defined as


( ) 
1
C (u, v) = exp − (− log( u ) ) + (− log( v) )
    ,
 

= (− log(t ) )

With generator  (t ) (Nicklas, 2013).

42 | P a g e
© Yimam JA, UNISA 2019
Frank copula (Frank, 1979): For the copula parameter θ ∈ (−∞, ∞)\ {θ}(Aas et al.) the Frank
copula is defined as

1 (e −u − 1)(e −v − 1) 

C (u, v) = − log1 +  ,
  e − − 1 

 e −t − 1 
with generator  (t ) = − log
 e − − 1 .
 

Again, we set C = Π for θ = 0(Nicklas, 2013).

Ali-Mikhail-Haq Copula (Ali et al., 1978): For copula parameter θ ∈ [−1, 1], The Ali-Mikhail-
Haq is defined as

uv
C (u, v) =
1 −  (1 − u )(1 − v)

with generator  (t ) = ln[1 −  (1 − t )] / t

Note: Among the 22 Archimedean copulas, AMH copula is the only copula whose parameter
lies on [-1, 1] and measures both, positive and negative, dependence.

Pair Copula Construction

Another family of copula called a pair-copula construction (PCC) was developed using only
bivariate copulas to construct a general construction method for multivariate copulas. The classes
of multivariate copulas that we have discussed so far are limited in modelling various
dependence structures. Among the flexible multivariate dependence structures which are needed
especially in financial applications were the centre of the distribution and the tails (the upper and
lower tail) dependence parameters. The application of copula for modelling purpose has been
applied to many areas including actuarial sciences, finance, neuroscience, and weather research
as cited in Kim et al., (2013). Among the copula-based models which are parametric copula
families, elliptical and Archimedean copulas were commonly applicable for the application areas
stated above. These families are limited in some aspect like tail dependence and dependence

43 | P a g e
© Yimam JA, UNISA 2019
flexibility. For instance, from elliptical copulas, the Gaussian copula allows for an arbitrary
correlation matrix with zero tail dependence while the tail dependence parameter from the
multivariate t-copula were driven from only a single degree of freedom parameter (Fang et al.,
2002, Frahm et al., 2003).

Besides to fill the gaps of elliptical copulas, Archimedean copula classes, namely, fully and
partially nested Archimedean copulas, Hierarchical Archimedean copulas and Multivariate
Archimedean copulas were considered by several researchers (Joe, 1997, Nelsen, 2007, Savu and
Trede, 2006, Schirmacher and Schirmacher, 2008). However, these extensions require additional
parameter restrictions. These additional parameters reduced the flexibility of the extended copula
functions for modelling dependence structures. To overcome the limitations of existing copula-
based models, a vine copula-based model has been developed. This vine is called pair-copulas.
This copula-based model can express a multivariate copula by using a cascade of bivariate
copulas.

The first pair-copula construction of a multivariate copula for the continuous data were
introduced by Joe (1996) in terms of distribution functions while Bedford and Cooke (2001;
2002) expressed these constructions in terms of densities and graphical way involving a
sequence of nested trees, which they called regular vines. The two popular subclasses of PCC
models, which are called Drawable vines, or D-vines and Canonical vines or C-vines were also
identified by Bedford and Cooke. Aas et al. (2009) and Czado (2010) also conducted different
extensions in the continuous data. Even though Genest and Neslehova (2007) provided some
evidence that discrete data inherit dependence properties from a parametric copula in a similar
way to the continuous case, Panagiotelis et al. (2012) provided different PCC models for discrete
data.

To overcome these problems, pair copula have been developed first by (Joe, 1996) and extended
by different scholars like (Aas et al., 2009, Bedford and Cooke, 2001, Bedford and Cooke, 2002,
Czado, 2010) for continuous data and for discrete data (Panagiotelis et al., 2012). One of the
contributions of Pair Copula Constructions (PCCs) in the construction of multivariate copula was
to provide a highly flexible framework for constructing copulas exhibiting a wide range of
dependence characteristics. This flexibility arises since any combination of bivariate copulas can

44 | P a g e
© Yimam JA, UNISA 2019
be used to construct PCC models (Czado, 2010). Since PCC is the concern of the current study,
details of PCC is presented in the coming consecutive topics in sections 4.3, 4.4 and 4.5.

Dependency Measures

For understanding complicated dependence structure, measures of dependence are the most
commonly used instruments. Among commonly used measures of dependence, the most popular
is Pearson’s correlation coefficient. Under strictly increasing linear transformations, Pearson’s
correlation coefficient is invariant but not under non-linear transformations. It is also defined
only for pairs of random variables with finite variances, but this can bring problems when
working with heavy-tailed distributions. Therefore, measuring dependence by standard
correlation is adequate in the context of multivariate elliptical distributions. There are increasing
proportions of nonlinear risks like the non-normal behaviour of most financial time series. As a
result, other tools are needed since estimates of risk dependence via linear correlation neglects
nonlinearities and leads in most cases to underestimation of the global risk. Since copula is often
the key issue for numerous models in relation to the above limitations, it is very important to find
the copula that describes the complete dependence structure. For detail on these, we refer to
several researchers (Fan, 2009, Genest and Nešlehová, 2007, Nelsen, 2007, Pirktl, 2007, Nicklas,
2013).

Since the copula functions are invariant under strictly increasing transformations, it makes sense
to consider dependence measures which are also invariant under such transformations. Kendall’s
tau and Spearman’s rho are the most widely known scale-invariant measures of association. Both
measure the form of dependence known as concordance.

Measure of Concordance

Pair of random variables is said to be concordant, if large values of variable is associated with
large values of the other variable and small values of one with small values of the other. On the
other hand it is discordant, if large values of one variable are associated with small values of the
other. A more formal definition is the following:

45 | P a g e
© Yimam JA, UNISA 2019
Definition: Consider, (xi , x j ) and ( yi , y j ) two observations from a pair random vectors (X, Y).

We say that (x , x )
i j and (y , y )are
i j concordant if ( xi − x j )( yi − y j )  0 and discordant if

( xi − x j )( yi − y j )  0 .

Definition: Kendall’s tau for the pair random vectors (X, Y) and whose copula C is given by

 = P(( xi - x j )(y i - y j ) > 0) - P(( xi - x j )(yi - y j )  0)

= 4 C (u , v)dC (u , v) − 1
[ 0 ,1]2

Hence Kendall’s tau is the probability of concordance minus the probability of discordance
(Nelsen, 2007).

Definition: Spearman’s rho for the random vector (X, Y) and the copula C is given by

 ( X , Y ) = 3P (( xi - x j )(yi - y j ) > 0) - P (( xi - x j )(yi - y j )  0)

1 1
= 12   C(u, v)dudv − 3.
0 0

Kendall’s tau and Spearman’s rho have many common properties. They can measure the degree
of monotonic dependence between random variables. Both are taking values in [-1, 1] which
measures symmetric dependence. They have the value 1 when X and Y are “comonotonic” and
have -1 when they are “countermonotonic”, the value [-1, 1] does not necessarily imply that all
those values can actually be obtained by a particular copula.

Tail Dependence

The idea of Kendall’s tau and Spearman’s rho is to measure the dependence of the copula on the
event space (0, 1) (Nicklas, 2013). On the contrary, there are cases that measure the dependence
between the variables in the upper tail or the lower tail of the bivariate distributions. This is
called tail dependence, which measures the dependence of extreme events. Nelson (2007) defines
tail dependence for a copula as follows. A random variable X and Y have marginal distribution
functions FX (x) , FY ( y ) and the copula function C, the lower tail dependence coefficient is given
by

46 | P a g e
© Yimam JA, UNISA 2019
 L = lim Pr(Y  FY−1 (u ) | X  FX−1 (u ) )
u 0

C (u, v)
= lim
u 0 u
and dependence coefficient for the upper tail is given by
U = lim Pr(Y  FY−1 (u ) | X  FX−1 (u ) )
u 1

1 − C (u, v)
= 2 − lim
u 1 1− u

Note that C has lower tail dependence if L  (0, 1) and no lower tail dependence if  L = 0 .
Similarly, C has upper tail dependence if U  (0,1) and no upper tail dependence if U = 0 .

Dependencies Characteristics of Bivariate Copula Families

Previous sections have introduced Kendall’s tau as a measure of dependence. The parameters of
the copula and their values of Kendall’s tau have determined relationships as presented in Table
4.1. The table also includes both the upper and lower tail dependence properties for each copula
family.

The upper or lower tail dependence cannot be treated in the Gaussian as well as in the Frank
copula. Nevertheless, the t copula treats both cases and it is represented by U =  L . On the other

hand the Clayton can be used to model lower tail dependence whereas the Gumbel copula for
upper tail dependence.

47 | P a g e
© Yimam JA, UNISA 2019
 Table 4. 1: Kendall’s tau, upper and lower tail dependence for bivariate copula families
(Dissmann, 2010).

Copula Kendall’s tau Upper tail Lower tail

dependence dependence
Gauss 2 u = 0 L = 0
= arcsin e(  )

T 2  1−  
= arcsin e(  ) u = L = t v+1  − v + 1 
 1+  
 
Clayton  u = 0 L = 2 −1 / 
=
 +2
Gumbel  −1 U = 2 −1 /  L = 0
=

Frank 4 D1 ( ) u = 0 L = 0
 = 1− +4
 
Ali-Mikhail- 3 − 2 2(1 −  ) 2 ln(1 −  )
= −
Haq 3 3 2

 x /
D1 ( ) =  dx , being the Debye function.
0 exp(x) − 1

This section provides pair copula-based cumulative logit model for jointly modelling the
dependence between availability, accessibility and utilisation of food security dimensions and
their respective determinants. The quartile score computed for each dimension categorized the
food security status for each dimension as “severe food in-secured”, “mildly food in-secured”,
“moderately food in-secured”, and “food secured”. This computation resulted in three ordinal
dependent variables, namely, availability, accessibility and utilisation. Therefore, for assessing
and interpreting food security status and determinant factors, a well-defined conceptual
framework is crucial. As a result, modelling the determinant factors of household food insecurity
is the case of modelling multivariate ordinal data that can consider the dependency between the
dimensions.

48 | P a g e
© Yimam JA, UNISA 2019
A pair copula construction approach was proposed to determine both the dependence between
food security dimensions and their respective associated factors simultaneously. A nice feature
of the PCC approach in this setting is measuring the dependency of the food security dimensions
using the copula parameter and the associated determinant factors of household food security for
each dimension using the parameters of the marginal distributions. The pair copula construction
approach with D-vine is attractive since it allows pairwise positive dependence structures as the
presented conceptual framework by FAO (2008) and has closed form cumulative distribution
function (cdf). Moreover, no other copula family has both these properties. FAO (2008) indicates
that availability contributes to accessibility, accessibility contributes to utilisation and given that
accessibility, availability contributes to utilisation; this is what a D-vine assumes in PCC.

This section demonstrates how to model and estimate dependence and marginal parameters from
multivariate ordinal data using pair copula constructions via ordinal logistic regression to our
motivating problem. This thesis did not evaluate the performance of this approach through
simulation studies because it was evaluated via Bernoulli and Poisson discrete distributions by
(Panagiotelis et al., 2012) and found to be a good model. They have also implemented the model
for longitudinal ordinal data via probit model. However, the scale of the logistic is greater than
the normal and this made the interpretation easier for logistic version and popular in many fields
(Choi, 2012). As far as the researcher review of literature is concerned, no work has been
conducted on the ordinal logistic version so far. Hence this section concerned on implementing
the developed discrete PCC model via ordinal logistic regression for modelling household food
insecurity determinants.

Since the current study concerns on discrete aspect in particular multivariate ordinal data, we
now briefly review some key concepts for vine PCCs in the continuous case before introducing
discrete vine PCCs. The aim here is to highlight some important distinctions in modelling
discrete and continuous data via a copula approach, and to provide background for the
introduction of discrete D-vine PCC presented in detail in section 4.3.2.

PCCs in the Continuous Case

For a vector Y = (Y1 , ...,Ym ) of continuous random variables with joint density function

f ( y1 , ..., y m ) , a PCC is derived by starting with the following decomposition

49 | P a g e
© Yimam JA, UNISA 2019
 f ( y1 , ..., y m ) = f1|2, ...,m ( y1 | y2 , ..., y m ) f 2|3, ...,m ( y 2 | y3 , ..., y m )... f m ( ym ) (4.13)

Recalling equation (2.4), we can simplify the bivariate case to

f ( y1 , y2 ) = c12 ( F ( y1 ), F ( y2 )) f1 ( y1 ) f 2 ( y2 ) (4.14)

where c12 (., .) is the appropriate pair-copula density for the pair of transformed variables

F1 ( y1 ) and F2 ( y2 ) .

Any transformation using the factorization of Equation (4.13) and Equation (4.14) different
decomposition can be constructed. For example, the 3-dimensional case decomposition results in

f1, 2, 3 ( y1 , y 2 , y3 ) = f1|2, 3 ( y1 | y 2 , y3 ) f 2|3 ( y 2 | y3 ) f 3 ( y3 ) (4.15)

Basic calculations give the conditional density of Y2 and Y3

f 2 , 3 ( y 2 , y3 ) c23 ( F ( y 2 ), F ( y3 )) f 2 ( y 2 ) f 3 ( y3 )
f 2|3 ( y 2 | y3 ) = =
f 3 ( y3 ) f 3 ( y3 ) . (4.16)
= c23 ( F ( y 2 ), F ( y3 )) f 2 ( y 2 )

Similarly,

f13|2 ( y1 | y 2 , y3 | y 2 )
f1|2, 3 ( y1 | y 2 , y3 ) =
f 3|2 ( y3 | y 2 )
c13|2 ( F1|2 ( y1 | y 2 ), F3|2 ( y3 | y 2 )) f1|2 ( y1 | y 2 ) f 3|2 ( y3 | y 2 )
=
f 3|2 ( y3 | y 2 )
= c13|2 ( F1|2 ( y1 | y 2 ), F3|2 ( y3 | y 2 )) f1\ 2 ( y1 | y 2 )

= c13|2 ( F1|2 ( y1 | y2 ), F3|2 ( y3 | y2 )) f1|2 ( y1 | y2 )

(4.17)
= c13|2 ( F1|2 ( y1 | y2 ), F3|2 ( y3 | y2 )) . c12 ( F ( y1 ), F ( y2 )) f1 ( y1 )

Using Equation (4.15), (4.16) and (4.17), the following decomposition is appears

f1, 2, 3 ( y1 , y2 , y3 ) = c13|2 ( F1|2 ( y1 | y2 ), F3|2 ( y3 | y2 )). c12 ( F ( y1 ), F ( y2 )) f1 ( y1 ).

(4.18)
c23 ( F ( y2 ), F ( y3 )) . f 3 ( y3 ) . f 2 ( y2 ) . f1 ( y1 ).

50 | P a g e
© Yimam JA, UNISA 2019
This example illustrates the construction of a 3-dimensional density using the bivariate copula
and the corresponding marginal distributions. Any other factor in Equation (4.13), the same
procedure is possible using the general formula letting V h be any scalar element of V and V\ h its

complement, with Y j not an element of V (Panagiotelis et al., 2012, Nicklas, 2013,

Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al., 2015);

cY j ,Vh |V|h ( FY j |V|h , FVh |V|h ) fY j |V|h . f Vh |V|h

fY j |V =
fVh |V|h
(4.19)
= cY j ,Vh |V|h ( FY j |V|h , FVh |V|h ) fY j |V|h

where cY j , Vh |V|h denotes the pair copula density describing the dependence between Y j and V h

conditional on V|h = v|h . If we assume that the conditional copulas depend on the conditioning set

only through their arguments, the decomposition in Equation (4.19) can motivate a statistical
model. Typically, parametric bivariate copulas such as among the Archimedean families
(Clayton, Gumbel & Frank) and elliptical families (Gaussian and Student t) copulas can be
chosen to model the pair copulas.

The arguments of the pair copulas are conditional distribution functions and can be evaluated
using the following expression given by Joe (1996) and cited in (Panagiotelis et al., 2012,
Nicklas, 2013, Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al., 2015);

CY j,Vh |V|h ( FY j |V|h ( y j | V|h ), FVh |V|h (v j | V|h ))

FY j |Vh ,V|h ( y j | vh , V|h ) = . (4.20)
FVh |V|h (v j | V|h )

To compute the density of a PCC, one can be found the algorithms that recursively compute PCC
density (Aas et al., 2009).

In conclusion, under appropriate regularity conditions, a multivariate density can be expressed

as a product of m*(m − 1)/2 bivariate copulas, acting recursively on several different conditional
probability distributions using expression (4.19). This leads to a large number of possible pair-
copulas constructions. To organize all possible decompositions, a graphical model called a
regular vine has been introduced by Bedford and Cooke (2002). Regular vine decompositions

51 | P a g e
© Yimam JA, UNISA 2019
are concentrated only on the D-vines and C-vines, the special cases of regular vines. Each vine
gives a specific way of decomposing the density. These models can be specified as a nested set
of trees.

Vines of Continuous Case

A vine is characterized by m − 1 trees denoted T j for j = 1, ..., m − 1 . The j th tree is made up of

nodes, denoted N j and edges which join these nodes, denoted E j (Panagiotelis et al., 2012,

Nicklas, 2013, Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al., 2015). “A regular vine tree
is called D-vine tree if each node in T −1 has at most 2 edges whereas C-vine tree if each tree T j

has a unique node with m − j edges.” The node with m − 1 edges in tree T1 is called the root.
Figure 4.1 shows a D-vine decomposition for a 5-dimensional density function and Figure 4.2
shows a canonical-vine (Panagiotelis et al., 2012).

In Figure 4.1, in the first tree of a D-vine, the edges simply join adjacent nodes yielding
E1 = 12 = c12 , 23 = c23 , 34 = c34 , 45 = c45 . The edges on the first tree become the nodes on the

second tree and in general N j +1 = E j . The edges of trees T2 , ...,Tm−1 also connect adjacent nodes.

Any element shared by two nodes will be in the conditioning set of the edge joining them. For
example, the edge joining node 12 and 23 is 13|2 while the edge joining 24|3 and 35|4 will be
25|34 (Panagiotelis et al., 2012).

The pair copulas that make up the corresponding PCC are simply indicated by the edges of the
entire vine {E1 , ..., E m−1} , so that the density for a 5-dimensional PCC is given by

 5 
f ( y1 , ..., y5 ) =  f k ( yk ).c12 .c23.c34 .c45 .c13|2 .c24|3 .c35|4 .c14|23.c25|34 .c15|234
 k =1 

where the arguments of the pair copulas and density functions have been dropped for ease of
notation.

52 | P a g e
© Yimam JA, UNISA 2019
 12 23 34 45 T1
1 2 3 4 5

13|2 24|3 35|4 T2

12 23 34 45

14|23 25|34 T3
13|2 24|3 35|4

14|23 15|234 25|34 T4

Figure 4. 1: A D-vine tree representation for m = 5.

The density f ( y1 , ..., y m ) corresponding to a D-vine may be written using a general formula as

m m −1 m − j

 f ( y ) c
k =1
k
j =1 i =1
{F ( yi | yi +1 , ..., yi + j −1 ), F ( yi + j | yi +1 , ..., yi + j −1 )},
i , i + j |i +1, ...,i + j −1 (4.21)

where index j identifies the trees, while i runs over the edges in each tree. In a D-vine, no node in
any tree T j is connected to more than two edges.

Figure 4. 2: A C-vine tree representation for m = 5.

53 | P a g e
© Yimam JA, UNISA 2019
Similarly, in Figure 4.2, in the first tree of a C-vine, the edges simply join adjacent nodes
yielding E1 = 12,13,14,15 . The edges on the first tree become the nodes on the second tree and
in general N j +1 = E j . The edges of trees T2 , ...,Tm−1 also connect adjacent nodes. Any element

shared by two nodes will be in the conditioning set of the edge joining them. For example, the
edge joining node 12 and 13 is 23|1, while the edge joining 23|1 and 25|1, will be 35|12. The pair
copulas that make up the corresponding PCC are simply indicated by the edges of the entire vine
{E1 , ..., E m−1} , so that the density for a 5-dimensional PCC is given by

 5 
f ( y1 , ..., y5 ) =  f k ( yk ).c12 .c13.c14 .c15 .c23|1.c24|1.c25|4 .c34|12 .c35|12 .c45|123
 k =1 

where the arguments of the pair copulas and density functions have been dropped for ease of
notation. As a D-vine, a canonical vine for m-dimensional density is given by

n n −1 n − j

 f ( y ) c
k =1
k
j =1 i =1
{F ( y j | y1 , ..., y j −1 ), F ( yi + j | y1 , ..., y j −1 )},
j , j +i|1, ..., j −1 (4.22)

where index 𝑗 identifies the trees, while 𝑖 runs over the edges in each tree. In a canonical vine,
each tree 𝑇𝑗 has a unique node that is connected to 𝑛 − 𝑗 edges.

Regular Vine Parameter Estimation

For estimation of regular vine, different scholars proposed non-standard methods and standard
estimation methods. Stepwise and MLE, Inference Function for Margins (IFM) and Stepwise
Semi-parametric Estimator (SSP) are the common standard estimation methods. MLE were
considered for the first time by several researchers (Aas et al., 2009) IFM by (Joe, 1996), and
SSP were by (Haff, 2012). These methods were designed for continuous data. However, we will
not discuss here in detail since the current concern is on the discrete data. One can refer the
references cited here for more detail. Just we now go to the PCC in discrete data.

54 | P a g e
© Yimam JA, UNISA 2019
 PCCs in the Discrete Case

In the following sections, we are going to introduce vine PCCs for discrete margins that can be
applicable for ordinal data. First, the discrete analogues to some important equations introduced
in Section [Link]. Second, we discuss the D-vine decomposition with an illustration in full detail
using 3-dimensional vine because the D-vine has certain advantages in applications where some
intuitive ordering of the margins can be made and that it gives flexible models with parameters
that can be estimated in a computationally and statistically efficient manner (Panagiotelis et al.,
2012). Third, we discuss the selection of pair copula families of the D-vine and the parameter
estimation.

Discrete PCCs

The aim here is to decompose a general multivariate probability mass function (pmf) into
bivariate pair copula building blocks like continuous data. The joint pmf can be decomposed into
a product of conditional probabilities using equation (4.13) for m discrete random variables
Y1 , Y2 , ..., Ym as

Pr(Y1 = y1 , ...,Ym = ym ) = Pr(Y1 = y1 | Y2 = y2 , ...,Ym = ym )  Pr(Y2 = y2 | Y3 = y3 , ...,Ym = ym ) 

... Pr(Ym = ym ). (4.23)
Now, this expression has terms of the form Pr(Y j = y j | Y| j = y| j ) where Y| j is the vector of

random variables Y1 , Y2 , ..., Ym excluding Y j and y| j is the same vector for the realized values of

the random variables. Choosing another element from the vector of random variables, we can
rewrite the discrete joint probability in a similar fashion to the continuous case as following:

Pr(Y j = y j , Yh = yh | Y| j , h = y| j , h )
Pr(Y j = y j | Y| j = y| j ) = (4.24)
p(Yh = yh | Y| j , h = y| j , h )
Now, recalling the probability mass function and the multivariate copula function for discrete
data in Equation (4.5), the bivariate conditional probability in the numerator can be expressed in
terms of a copula giving

55 | P a g e
© Yimam JA, UNISA 2019
Pr(Y j = y j | Y| j = y|h )


i +ih
(−1) j Pr(Y j  y j − i j , Yh  yh − ih | Y| j .h = y| j .h )
i j =0 ,1 ih =0 ,1
=
Pr(Yh = yh | Y| j .h = y| j .h )

  (−1)
i j +ih
CY j .Yh | y| j . h ( FY j |Y| j .h ( y j − i j ), FY j |Y| j . h ( yh − ih ))
i j =0 ,1 ih =0 ,1
=
Pr(Yh = yh | Y| j .h = y| j .h ) (4.25)

The arguments in equation (4.25) of the copula functions are evaluated using the following
(Panagiotelis et al., 2012, Nicklas, 2013, Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al.,
2015);

FY j |Yh .Y| j . h ( y j | yh , y| j .h ) = [CY j .Yh|Y| j . h ( FY j \ |Y| j .h ( y j | y| j .h ), FYh|Y| j .h ( yh | y| j .h ))

− CY j .Yh|Y| j .h ( FY j \ |Y\ j . h ( y j | y| j .h ), FYh|Y| j .h ( yh − 1 | y| j .h ))]
/ Pr(Yh = yh | Y| j .h = y| j .h )
(4.26)

This vine PCC has nice feature than multivariate and Gaussian copulas functions in evaluating
the probability mass function because the PCC requires 2m(m − 1) evaluations whereas the

multivariate and Gaussian copulas require 2 m evaluations (Panagiotelis et al., 2012).

D-vine in Discrete Data

For illustration purposes, we present in detail the 3-dimensional case for instance in the food
security data, 𝑌1 = Availability, 𝑌2 = Accessibility and 𝑌3 = Utilisation. Therefore,

Pr(Y1 = y1 , Y2 = y2 , Y3 = y3 )
(4.27)
= Pr(Y1 = y1 | Y2 = y2 , Y3 = y3 )  Pr(Y3 = y3 | Y2 = y2 )  Pr(Y2 = y2 )

Utilizing Equation (4.25) the right hand side of the first conditional probability can be rewritten
as:

56 | P a g e
© Yimam JA, UNISA 2019
 Pr(Y1 = y1 | Y2 = y2 , Y3 = y3 )
  (−1)
i1 =0 ,1i3 =0 ,1
i1 +i3
C13|2 ( F1|2 ( y1 − i1 | y2 ), F3|2 ( y3 − i3 | y2 ))
(4.28)
=
Pr(Y3 = y3 | Y2 = y2 )

Similarly, utilizing Equation (4.25), the first argument of the copula function in the numerator of
Equation (3.16) is given by;

C12 ( F1 ( y1 − i1 ), F2 ( y2 )) − C12 ( F1 ( y1 − i1 ), F2 ( y2 − 1))

F1|2 ( y1 − i1 | y2 ) = , (4.29)
Pr(Y2 = y2 )

and the second argument can be expressed as

C23 ( F2 ( y2 ), F3 ( y3 − i3 )) − C23 ( F2 ( y2 − 1), F3 ( y3 − i3 ))

F3|2 ( y3 − i3 | y2 ) = (4.30)
Pr(Y2 = y2 )

By cancelling terms and substituting, the probability mass function of the full expression for the
3-dimensional discrete D-vine is given by

Pr(Y1 = y1 , Y2 = y2 , Y3 = y3 )
C12 ( F1 ( y1 − i1 ), F2 ( y2 )) − C12 ( F1 ( y1 − i1 ), F2 ( y2 − 1))
={  (−1) i1 +i3
C13|2 ( ,
i1 =0 ,1i3 =0 ,1 F2 ( y2 ) − F2 ( y2 − 1) (4.31)

C23 ( F2 ( y2 ), F3 ( y3 − i3 )) − C23 ( F2 ( y2 − 1), F3 ( y3 − i3 ))

)}[F2 ( y2 ) − F2 ( y2 − 1)].
F2 ( y2 ) − F2 ( y2 − 1)

Somewhat confusing to write the general D-vine structure; however, the general dimension
algorithm for computing the probability mass function of a D-vine was outlined by Panagiotelis
et al. (2012). It is evident both from this algorithm and the 3-dimensional example above that

each bivariate pair copula only needs to be evaluated 4 times, specifically CY j .Yh |Y| j .h must be

evaluated at (Panagiotelis et al., 2012, Nicklas, 2013, Sirisrisakulchai and Sriboonchitta, 2014,
Stöber et al., 2015);

57 | P a g e
© Yimam JA, UNISA 2019
 ( FY j \ |Y| j .h ( y j | y| j.h ), FYh |Y| j .h ( yh | y| j.h )) , ( FY j \ |Y| j .h ( y j − 1 | y| j.h ), FYh |Y| j .h ( y h | y| j.h )) ,

( FY j \ |Y| j .h ( y j | y| j.h ), FYh |Y| j .h ( y h − 1 | y| j.h )) and ( FY j \ |Y| j .h ( y j − 1 | y| j.h ), FYh |Y| j .h ( y h − 1 | y| j.h )) .

In general, the evaluation of the probability mass function requires 2m(m − 1) evaluations of
bivariate copula functions, even though the continuous vine PCC is composed of only m(m−1)/2
pair copulas. These vine PCCs have still greater potential in high-dimensional settings since the
computational burden of evaluating the pmf in the elliptical copulas is 2m. As pointed out by
Panagiotelis et al. (2012) among the major advantage of D-vine PCCs, a wide variety of
dependence structures can be modelled by selecting different copula families as building blocks.
Gaussian, t, Clayton, Frank and Gumbel copulas are the commonly used parametric copula
families as building blocks.

The marginal probabilities can be modelled by logistic, probit or Poisson models. If the discrete
data have binary outcomes, binary logistic or probit models can be used, if ordinal, ordinal
logistic or probit and if counted, Poisson regression model can be used. For the current study
since the data are ordinal, cumulative logit model is used. A detail of the cumulative logit model
is given below.

The marginal distribution of ordinal data via cumulative logistic regression model for single
ordinal response variable Y that has C categories and labelled 1, 2, ..., C − 1. is given by (Agresti,
2010);
Pr(Y  j )
Pr(Y  j ) =
1 − Pr(Y  j )
(4.32)
exp( j +  )
= , j = 1, 2, ..., C − 1.
1 + exp( j +  )

This is the cumulative probabilities that an observation fall in category j or below, for C
categories. Each of the cumulative logits is an ordinary binary logit indicating the probability of

an outcome falling into either the first 1... j categories or the j+ 1..., C categories. Similarly, j
the intercept for each cumulative probability of the c category and column vector ß of parameters
that describes the effects of the explanatory variables (Agresti, 2010).

58 | P a g e
© Yimam JA, UNISA 2019
Hence, the arguments in equation (4.31) can take the marginal distribution function given in
equation (4.32). Moreover, the joint probability mass function in question (4.27) is expressed in
terms of the pair copula functions and ordinal marginal distributions. The newly constructed joint
probability mass function can be called pair copula-based multivariate cumulative logit model.
Finally, this function can be estimated using the appropriate parameter estimation technique and
appropriate bivariate copula families as building blocks (Panagiotelis et al., 2012, Nicklas, 2013,
Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al., 2015).

[Link].1 Selection of Pair Copula Families and Parameter Estimation of the D-Vine

Before parameter estimations of the pair copula construction model, determining the order of the
D-vine and selection of appropriate bivariate copula families for the model is the first task to be
done.

For the structure of R-vine copula extensive works of several researchers (Aas et al., 2009,
Czado et al., 2013, Dissmann et al., 2013, Sutkoff, 2014) were conducted. In particular, “the
order in the trees corresponding to a D-vine copula Aas et. al (2009) put the strongest bivariate
dependencies in the first tree of the D-vine tree specification. Strongest bivariate dependencies
within the copula distribution might be measured by Kendall’s τ or the tail dependence
coefficient λ, which is a function of the chosen bivariate copula”. For the current study on food
security, the conceptual framework for food security dimension is given by FAO (2008) like
Availability-Accessibility-Utilisation. We can use this order for the structure of D-vine for
inference purposes.

Now the bivariate Copula Families of the vine distribution can be selected since the order of the
D-vine tree specification is chosen. This part can be discussed as follows.

i. Selection of Pair Copula Families

Accordingly, as described above, we need to select a copula family for every pair of variables.
Commonly used copula families that we consider in the later applications are Gaussian (N), t,
Clayton (C), Gumbel (G) and Frank (F). The Clayton and Gumbel copulas are applicable only to
model positive dependence. Hence, in case of negative dependence (i.e. negative values for
Kendall’s tau) we can reduce them. Further, if the degree of freedom of the MLE is higher than
30, we will not use a t copula.

59 | P a g e
© Yimam JA, UNISA 2019
After reducing the possible options further, we can decide which copula fits “best”. To select
jointly the vine structure and best fit copula families, (Panagiotelis et al., 2015) have developed
an algorithm through adaptation from the algorithm developed for continuous data by Dissmann
et al. (2013). We also customized this algorithm for our purpose to decide best fit copula families
only as presented in Algorism I.

Algorithm I

Consider discrete random variables 𝑌 = 𝑌1 , 𝑌2 , 𝑌3 , … . , 𝑌𝑚 with known marginal distribution

functions 𝐹𝑗 (. ) , the steps to select copula families is as follows.

+ +
1. Generate the `pseudo data'𝑢𝑖𝑗 : = 𝐹𝑗 (𝑦𝑖𝑗 ) and 𝑢𝑖𝑗 : = 𝐹𝑗 (𝑦𝑖𝑗 − 1) for 𝑗 = 1, 2, … , 𝑚 and 𝑖 =
1, 2, … , 𝑛, where 𝑦𝑖𝑗 the value of the response for the jth margin and the ith observation.
2. Consider a pair of two margins 𝐼1 𝑎𝑛𝑑 𝐼2 ∁ {1, 2, … , 𝑚}
𝑟
i. Fit the copula 𝐶 𝜃 using the pseudo data for the margins 𝐼1 𝑎𝑛𝑑 𝐼2 for each bivariate
copula families as follows
𝜃̂ 𝑟 = 𝑎𝑟𝑔𝑚𝑎𝑥𝑙𝑛𝐿𝑟𝐼1 ,𝐼2 (𝜃 𝑟 ) (4.33)
Where,
𝑛
𝑟
+ + 𝑟+ − 𝑟− + 𝑟 − −
𝑙𝑛𝐿𝑟𝐼1 ,𝐼2 (𝜃 𝑟 ) = ∑ ln (𝐶 𝜃 (𝑢𝑖𝐼1
, 𝑢𝑖𝐼2
) − 𝐶 𝜃 (𝑢𝑖𝐼1
, 𝑢𝑖𝐼2
) − 𝐶 𝜃 (𝑢𝑖𝐼1
, 𝑢𝑖𝐼2
) + 𝐶 𝜃 (𝑢𝑖𝐼1
, 𝑢𝑖𝐼2
)
𝑖=1

ii. Compute a modified Akaike Information Criterion (AIC), that removes the effect of the
margins, given by
𝑚𝐴𝐼𝐶 𝑟 = −2𝑙𝑛𝐿𝑟𝐼1 ,𝐼2 (𝜃 𝑟 ) − 𝑙𝑛𝐿𝑟𝐼1 − 𝑙𝑛𝐿𝑟𝐼2 + 2𝑞𝑟 (4.34)
Where 𝑞𝑟 is the dimension of𝜃 𝑟 ,𝑙𝑛𝐿𝑟𝐼1 = ∑𝑛𝑖=1 ln ( 𝑢𝑖𝐼
+
1
−
− 𝑢𝑖𝐼1
),
𝑛
+ −
𝑛𝐿𝑟𝐼2 = ∑ ln ( 𝑢𝑖𝐼2
− 𝑢𝑖𝐼2
).
𝑖=1

A smaller mAIC value indicates a better parametric model.

iii. Compute new pseudo data for tree 2 means that conditional pseudo data as given by
+ −
𝑢𝑖,ℎ1 |ℎ2
: = 𝐹ℎ1 |ℎ2 (𝑦𝑖 ℎ1 |𝑦𝑖 ℎ2 ), 𝑢𝑖,ℎ1 |ℎ2
: = 𝐹ℎ1 |ℎ2 (𝑦𝑖 ℎ1 − 1|𝑦𝑖 ℎ2 ),
+ +
𝑢𝑖,ℎ1 |ℎ2
: = 𝐹ℎ2 |ℎ1 (𝑦𝑖 ℎ2 |𝑦𝑖 ℎ1 ), 𝑢𝑖,ℎ2 −1|ℎ1
: = 𝐹ℎ2|ℎ1 (𝑦𝑖 ℎ2 |𝑦𝑖 ℎ1 ),

60 | P a g e
© Yimam JA, UNISA 2019
 3. Repeat step 2 for all pairs of the new pseudo data and corresponding pair copulas. Also
compute new pseudo data in a similar fashion as step 2(iii).
4. Iterate to select the pair copulas.
Once the copula families selected for each edges and consecutive trees, the next step is parameter
estimation.

ii. Parameter Estimation

For estimation of regular vine, stepwise and MLE were consider for the first time by Aas et al.
(2009), Inference function for margins (IFM) by Joe (1996) and Stepwise semiparametric
estimator (SSP) were by Haff (2013) in the continuous margins. Similarly, Panagiotelis et al.
(2012) conducted MLE and IFM for the discrete margins.

a. Maximum Likelihood (ML) Estimator

Since the 3 dimensional D-vine is derived, hence the log-likelihood function of a 3 dimensional
D-vine is given by

n
l (  , ; y ) =  log(Pr(Y1 = y1 , Y2 = y 2 , Y3 = y3 ;  , ))
i =1
3
C12 ( F1 ( y1 − i1 ; 1 ), F2 ( y 2 ;  2 );12 ) − C12 ( F1 ( y1 − i1 ; 1 ), F2 ( y 2 − 1;  2 );12 )
=  log({   (−1) i1 + i3
C13|2 ( ,
t =1 i1 = 0 ,1 i3 = 0 ,1 F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )
C 23 ( F2 ( y 2 ;  2 ), F3 ( y3 − i3 ;  3 ); 23 ) − C 23 ( F2 ( y 2 − 1;  2 ), F3 ( y3 − i3 ;  3 ); 23 )
;13|2 )}[F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )]) ( 4.35)
F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )
Here  = ( 1 ,  2 ,  3 ) and  = (12 ,  23 , 13|2 ) where the marginal and copula parameters

respectively. Then, the ML estimator ˆ ML is obtained by maximizing the above log-likelihood

function over all parameters,  and  , simultaneously.

For the general case, let the model for the jth margin imply a marginal distribution function
Fij ( yij ;  j ) , where  j are the marginal parameters and the subscript i denotes that we observe a

sample yi = ( yi1 , yi 2 , ..., yim ) for i = 1, 2, ..., n. Similarly, the copula parameters for m-

dimensional dependence were given by  i ,i + j|i +1,...,i + j −1 . Then, the ML estimator ˆ ML is obtained by
maximizing the log-likelihood function over all parameters,  j and  i ,i + j|i +1,...,i + j −1 , simultaneously.

61 | P a g e
© Yimam JA, UNISA 2019
In this optimization, good starting values are required. Starting values for the marginal
parameters are obtained by following the first step of the IFM approach (will be discussed next
to this). Moreover, starting values for the copula parameters can be found by computing
empirical Kendall’s τ of bivariate copula function of the first tree which act as ‘pseudo’ data and
then transformed back to the copula parameter using a known Bijection.

b. Inference Function for Margins (IFM) Estimator

Consider equation (3.18) the 3-dimensional marginal distribution and the general case, in the

first step, maximum likelihood estimates of the marginal parameters ̂

IFM
are estimated for all

one at a time, ignoring dependence with the other margins. The resulting estimates ̂
IFM
are
plugged into the arguments of the marginal in the bivariate copula functions to estimate the pair

copula parameter ˆ
IFM
. In the second step, the copula parameters are estimated by maximum

likelihood using ̂
IFM
as an argument of the marginal (Panagiotelis et al., 2012, Nicklas, 2013,
Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al., 2015).

Panagiotelis et al. (2012) point out that joint ML estimates are generally of a higher quality than
IFM, but only slightly so. On the other hand, IFM estimation is simpler and faster, particularly
for more complicated marginal models. Hence, in the current study, the marginal models are
only three; joint MLE can be implemented with the cumulative logit margin as discussed above.

In repeated or longitudinal outcomes, the dependency among outcomes must be accounted for in
order to make valid inference. In this study, the households were surveyed three times at six-
month interval for each of the three dimensions. A composite food security index that has four
levels was computed from these three dimensions for each of the three round of data collection.
The levels are “sever food in-secured”, “mildly food in-secured”, “moderately food in-secured”
and “food secured” for each round of data. Therefore, three composite food security indexes
were obtained from the three phases of data collection, resulted in longitudinal ordinal outcomes.
Hence, in this regard, modelling the household food security status is the case of modelling

62 | P a g e
© Yimam JA, UNISA 2019
longitudinal ordinal data that can take into consideration the dependency between consecutive
time points.

In modelling the stability and the determinants of household food insecurity, a PCC model was
proposed. A nice feature of the PCC approach in this setting is measuring the dependency of the
consecutive food security status of the households using the copula parameter and the respective
associated determinants using the parameters of the marginal distributions. The dependences of
the consecutive food security statuses are one-dimensional. Hence, the pair copula construction
approach with D-vine is attractive since it allows pairwise positive dependence structures and
has closed form cumulative distribution function (cdf), no other copula family has both these
properties.

This section demonstrated how to model and estimate dependence and marginal parameters from
longitudinal ordinal data using pair copula constructions via ordinal logistic regression to our
motivating problem. The thesis did not evaluate the performance of this approach through
simulation studies because it was evaluated via Bernoulli and Poisson discrete distributions by
(Panagiotelis et al., 2012) and found to be a good model. They have also implemented the model
for longitudinal ordinal data via probit model. However, the scale of the logistic is greater than
the normal and this made the interpretation easier for logistic version and popular in many fields
(Choi, 2012). As far as the researcher review of literature is concerned, no work has been
conducted on the ordinal logistic version so far. Hence, this thesis in this section implemented
the developed discrete PCC model via ordinal logistic regression for modelling the stability and
determinants of household food insecurity status.

Since the current study concerns on discrete aspect in particular longitudinal ordinal data, first
we briefly review some key concepts for vine PCCs in the continuous longitudinal cases before
introducing discrete vine PCCs for longitudinal ordinal margins.

Pair Copula Construction for Longitudinal Continuous Data

A continues univariate random variable repeatedly measured for 𝑇 time points given by
Y = (Y1 , ..., YT ) , the joint density function f ( y1 , ..., yT ) is decomposed as follows;

63 | P a g e
© Yimam JA, UNISA 2019
 T
f ( y1 , ..., yT ) = f ( yt | yt −1 , ..., y1 ) * f ( yt −1 | yt −2 , ..., y1 ) * ... * f ( y1 ) =  f ( yt | yt −1 , ..., y1 ) * f ( y1 )
t =2 . (4.36)

In the Sklar’s theorem the conditional bivariate densities, 𝑓(𝑦𝑡 |𝑦𝑡−1 , 𝑦𝑡−2 , … , 𝑦1 ) in equation
(4.36) for 𝑡 > 𝑠 is given by

f ( yt , y s | yt −1 , ..., y s +1 )
f ( yt | yt −1 , ..., y1 ) =
f ( y s | yt −1 , ..., y s +1 ) , (4.37)
= ct , s |t −1, t − 2 , ...,s +1 ( F ( yt | yt −1 , ..., y s +1 ), F ( y s | yt −1 , ..., y s +1 )) f ( yt | yt −1 , ..., y s +1 )

where f (• | • ) and F (• | •) denotes the conditional density and cumulative density functions,
respectively and t and s be any arbitrary distinct indices.

By setting s=1, the bivariate conditional density in equation (4.37) yields the following
decomposition,

f ( y t | y t −1 , ..., y1 ) = ct ,1 |t −1, t −2, ...,2 ( F ( y t | yt −1 , ..., y 2 ), F ( y s | y t −1 , ..., y 2 )) f ( yt | yt −1 , ..., y 2 )

. (4.38)

Repeatedly, setting s = 2, 3, …, t-1, the conditional density in (4.3) leads the following
conditional density decomposition,

t −1
f ( yt | yt −1 , ..., y1 ) =  ct ,s |t −1, t − 2 , ...,s +1 ( F ( yt | yt −1 , ..., y s +1 ), F ( y s | yt −1 , ..., y s +1 ))* f ( yt )
s =1

(4.39)

Replacing equation (4.39) in equation (4.36), then the joint distribution function becomes

 t −1 
f(y1 , ..., yT ) =   ct,s|t −1, t − 2 , ...,s+1(F(yt |yt −1 , ..., y s +1 ), F(ys |yt −1 , ..., y s +1 ))*f(yt )*f ( y1 )
T
(4.40)
t = 2  s =1 

Equation (4.40) is a product of 𝑇(𝑇 − 1)/2 bivariate pair copula densities and 𝑇 marginal
densities (Ruscone and Osmetti, 2017, Smith et al., 2010). This leads to a large number of
possible pair-copulas constructions. To organize all possible decompositions, a graphical model
called a regular vine has been introduced by (Bedford and Cooke, 2002). Regular vine
decompositions are concentrated only on the D-vines and C-vines, the special cases of regular

64 | P a g e
© Yimam JA, UNISA 2019
vines. Equation (4.40) can be recognized as D-vine model. Detail of D-vine construction was
displayed earlier in section 4.3.1.

D-Vine Parameter Estimation

For estimation of regular vine, different scholars proposed non-standard methods and standard
estimation methods. Stepwise and MLE (MLE), Inference Function for Margins (IFM) and
Stepwise Semi-parametric Estimator (SSP) are the common standard estimation methods. MLE
were considered for the first time by (Aas et al., 2009), IFM by (Joe, 1996), and SSP were by
(Haff, 2012). These methods are designed for continuous data. We will not discuss here in detail
since the current concern is on the discrete data. One can refer the referees cited here for more
detail. Just we now go to the PCC in discrete data.

Pair Copula Construction for Longitudinal Discrete Data

The aim here is to decompose the probability mass function (pmf) of longitudinal discrete data
into bivariate pair copula building blocks like the decomposition of longitudinal continuous data.
For 𝑇 time ordered discrete random variables given by 𝑌1 , 𝑌2 , … , 𝑌𝑇 , the joint pmf can be
decomposed into a product of conditional probabilities as

Pr(Y1 = y1 , ...,YT = yT ) = Pr(Yt = yt | Yt −1 = yt −1 , ...,Y1 = y1 )  Pr(Yt −1 = yt −1 | Yt −2 = yt −2 , ...,Y1 = y1 ) ... Pr(Y1 = y1 )

T
=  Pr(Yt = yt | Yt −1 = yt −1 , ...,Y1 = y1 )* Pr(Y1 = y1 ). (4.41)
t =2

Now, the expression 𝑃𝑟(𝑌𝑡 = 𝑦𝑡 |𝑌𝑡−1 = 𝑦𝑡−1 , 𝑌𝑡−2, = 𝑦𝑡−2 , … , 𝑌1 = 𝑦1 ) can be written as the
form 𝑃𝑟(𝑌𝑡 = 𝑦𝑡 |𝑌|𝑡 = 𝑦|𝑡 ), where 𝑌|𝑡 is the vector of random variables 𝑌1 , 𝑌2 , … , 𝑌𝑇 excluding 𝑌𝑡
and 𝑦|𝑡 is the same vector for the realized values of the random variables. Choosing another
element s from the vector of random variables, we can rewrite the discrete joint probability as
following:
Pr(Yt = yt , Ys = ys | Y|t , s = y|t , s )
Pr(Yt = yt | Y|t = y|t ) = (4.42)
Pr(Ys = ys | Y|t , s = y|t , s )
Recalling equation (4.5), the bivariate conditional probability in the numerator of equation (4.42)
can be expressed in terms of a copula giving,

65 | P a g e
© Yimam JA, UNISA 2019
   (−1)
it = 0 ,1 i s = 0 ,1
it + i s
Pr(Yt  yt − it , Ys  ys − is | Y|t .s = y|t .s )
Pr(Yt = yt | Y|t = y|t ) =
Pr(Ys = ys | Y| j .s = y| j .s )

  (−1)
it = 0 ,1 i s = 0 ,1
it + i s
CYt .Ys | y|t . s ( FYt |Y|t . s ( yt − it ), FYs |Y|t . s ( ys − is )) (4.43)

=
Pr(Ys = ys | Y|t .s = y|t .s )

Inserting equation (4.43) in equation (4.41), now we get the following decomposed joint
probability mass function for longitudinal discrete data.

   (−1) it +is CYt .Ys | y|t . s ( FYt |Y|t . s ( yt − it ), FYt |Y|t . s ( y s − is )) 

 i =0,1i =0,1
T

Pr(Y1 = y1 , ..., YT = yT ) =   t s  * Pr(Y1 = y1 ). (4.44)
t =2  Pr(Ys = y s | Y|t .s = y|t .s ) 
 
Equation (4.44) can be recognized as general D-vine pair copula construction model. This vine
PCC requires 2T (T − 1) evaluations for evaluating the probability mass function whereas the

multivariate and Gaussian copulas require 2T evaluations (Panagiotelis et al., 2012).

For illustration purposes, we present in detail the 3-dimensional longitudinal case. Therefore,

Pr(Y1 = y1 , Y2 = y2 , Y3 = y3 )
= Pr(Y3 = y3 | Y1 = y1 , Y2 = y2 )  Pr(Y2 = y2 | Y1 = y1 )  Pr(Y1 = y1 )
= Pr(Y3 = y3 | Y1 = y1 , Y2 = y2 )  Pr(Y1 = y1 | Y2 = y2 )  Pr(Y2 = y2 ) (4.45)

Utilizing Equation (4.43) the conditional probability, 𝑃𝑟(𝑌3 = 𝑦3 |𝑌1 = 𝑦1 , 𝑌2 = 𝑦2 ) can be

rewritten as:

  (−1)
i1 = 0 ,1 i3 = 0 , 1
i1 + i3
C13| 2 ( F ( y1 − i1 | y2 ), F ( y3 − i3 | y2 ))
Pr(Y3 = y3 | Y1 = y1 , Y2 = y2 ) = (4.46)
Pr(Y3 = y3 | Y2 = y2 )

Similarly utilizing Equation (4.8), the first argument of the copula function in the numerator of
Equation (4.45), 𝐹(𝑦1 − 𝑖1 |𝑦2 ) is given by

C12 ( F ( y1 − i1 ), F ( y2 )) − C12 ( F ( y1 − i1 ), F ( y2 − 1))

F ( y1 − i1 | y2 ) = . (4.47)
Pr(Y2 = y2 )

66 | P a g e
© Yimam JA, UNISA 2019
 And the second argument, F ( y3 − i3 | y2 ) can be expressed as

C23 ( F ( y2 ), F ( y3 − i3 )) − C23 ( F ( y2 − 1), F ( y3 − i3 ))

F ( y3 − i3 | y2 ) = . (4.49)
Pr(Y2 = y2 )

By cancelling similar terms and substituting, the probability mass function of the full expression
for the 3-dimensional longitudinal discrete D-vine is given by

Pr(Y1 = y1 , Y2 = y2 , Y3 = y3 )
C12 ( F ( y1 − i1 ), F ( y2 )) − C12 ( F ( y1 − i1 ), F ( y2 − 1))
={   (−1) i1 + i3
C13|2 ( , (4.50)
i1 = 0 ,1 i3 = 0 ,1 F ( y2 ) − F ( y2 − 1)
C23 ( F ( y2 ), F ( y3 − i3 )) − C23 ( F ( y2 − 1), F ( y3 − i3 ))
)}[F ( y2 ) − F ( y2 − 1)].
F ( y2 ) − F ( y2 − 1)

Equation (4.50) is a D-vine model for 3-dimentional longitudinal discrete case. It is evident from
the 3-dimensional example above that each bivariate pair copula only needs to be evaluated four

times, specifically CYt .Ys |Y|t . s must be evaluated at

( F ( yt | y|t .s ), F ( ys | y|t .s )) , ( F ( yt − 1 | y|t .s ), F ( ys | y|t .s )) , ( F ( yt | y|t .s ), F ( ys − 1 | y|t .s )) and

( F ( yt − 1 | y|t .s ), F ( ys − 1 | y|t .s )) .

In general, the evaluation of the probability mass function requires 2T(T − 1) evaluations of
bivariate copula functions, even though the continuous vine PCC is composed of only T(T−1)/2
pair copulas. These vine PCCs have still greater potential in high-dimensional settings since the
computational burden of evaluating the pmf in the elliptical copulas is 2T. As pointed out by
Panagiotelis et al. (2012) among the major advantage of D-vine PCCs, a wide variety of
dependence structures can be modelled by selecting different copula families as building blocks.
Among these, Gaussian, t, AMH Clayton, Frank and Gumbel copulas are the commonly used
parametric copula families as building blocks.

The marginal probabilities in equation (4.50) can be modelled either of among discrete
probability distributions. For the current study, since the data are ordinal, cumulative logit
model is used. Details of the cumulative logit model were given in equation (4.32).

67 | P a g e
© Yimam JA, UNISA 2019
Hence, the arguments in equation (4.44 or 4.50) can take the marginal distribution function given
in equation (4.32). Moreover, the joint probability mass function in question (4.45) is expressed
in terms of the pair copula functions and ordinal marginal distributions. The newly constructed
joint probability mass function can be called pair copula-based longitudinal cumulative logit
model. Finally, this function can be estimated using the appropriate parameter estimation
technique and appropriate bivariate copula families as building blocks.

Selection of Pair Copula Families and Parameter Estimation of the D-Vine

Before the estimation of the parameters for the pair copula construction model determining the
order of the D vine and choose appropriate bivariate copula families for the model is the first
task to be done.

For the structure of R-vine, copula extensive approaches were reviewed in chapter three of this
paper. Moreover, the current study concerned on time ordered or longitudinal case, the time
order by itself can be taken as D-vine structure for inference purposes being stars from newest to
the oldest. Now the bivariate Copula Families of the vine distribution can be selected since the
order of the D-vine tree specification is chosen. This part can be discussed as follows.

Selection of Pair Copula Families

Accordingly as described above, we need to select a copula family for every pair of variables.
Commonly used copula families that we consider in the later applications are Gaussian (N), Ali-
Mikhail-Haq (AMH), Clayton (C), Gumbel (G) and Frank (F). The Clayton and Gumbel copulas
are applicable only to model positive dependence. Hence, in case of negative dependence (i.e.,
negative values for Kendall’s tau), we can reduce them. Further, if the degree of freedom of the
MLE is higher than 30, we will not use a t copula.

After reducing the possible options further, we can decide which copula fits “best”. To select
jointly the vine structure and best fit copula families, (Panagiotelis et al., 2015) have developed
an algorithm through adaptation from the algorithm developed for continuous data by Dissmann
et al. (2013). We also customized this algorithm for our purpose to decide best fit copula families
only as presented in Algorism II of this chapter.

68 | P a g e
© Yimam JA, UNISA 2019
Algorithm II

Consider 𝑇 time ordered discrete random variables 𝑌 = 𝑌1 , 𝑌2 , 𝑌3 , … . , 𝑌𝑇 with known marginal

distribution functions 𝐹𝑡 (. ) , the steps to select copula families is as follows.

+ +
1. Generate the `pseudo data' 𝑢𝑖𝑡 : = 𝐹𝑡 (𝑦𝑖𝑡 ) and 𝑢𝑖𝑡 : = 𝐹𝑡 (𝑦𝑖𝑡 − 1) for 𝑡 = 1, 2, … , 𝑇 and 𝑖 =
1, 2, … , 𝑛, where 𝑦𝑖𝑡 the value of the response for the tth time ordered margin and the ith
observation.
2. Consider a pair of two time ordered margins 𝐼1 𝑎𝑛𝑑 𝐼2 ∁ {1, 2, … , 𝑇}
𝑟
i. Fit the copula 𝐶 𝜃 using the pseudo data for the margins 𝐼1 𝑎𝑛𝑑 𝐼2 for each bivariate
copula families as follows:
𝜃̂ 𝑟 = 𝑎𝑟𝑔𝑚𝑎𝑥𝑙𝑛𝐿𝑟𝐼1 ,𝐼2 (𝜃 𝑟 ) (4.51)
Where,
𝑟 𝑟 𝑟
𝑙𝑛𝐿𝑟𝐼1 ,𝐼2 (𝜃 𝑟 ) = ∑𝑛𝑖=1 ln (𝐶 𝜃 (𝑢𝑖𝐼
+
1
+
, 𝑢𝑖𝐼2
+
) − 𝐶 𝜃 (𝑢𝑖𝐼1
−
, 𝑢𝑖𝐼2
−
) − 𝐶 𝜃 (𝑢𝑖𝐼1
+
, 𝑢𝑖𝐼2
)+
𝑟 − −
𝐶 𝜃 (𝑢𝑖𝐼1
, 𝑢𝑖𝐼2
))
ii. Compute a modified Akaike Information Criterion (AIC), that removes the effect of the
margins, given by
𝑚𝐴𝐼𝐶 𝑟 = −2𝑙𝑛𝐿𝑟𝐼1 ,𝐼2 (𝜃 𝑟 ) − 𝑙𝑛𝐿𝑟𝐼1 − 𝑙𝑛𝐿𝑟𝐼2 + 2𝑞𝑟 (4.52)
Where 𝑞𝑟 is the dimension of 𝜃 𝑟 ,𝑙𝑛𝐿𝑟𝐼1 = ∑𝑛𝑖=1 ln ( 𝑢𝑖𝐼
+
1
−
− 𝑢𝑖𝐼1
),
𝑙𝑛𝐿𝑟𝐼2 = ∑𝑛𝑖=1 ln ( 𝑢𝑖𝐼
+
2
−
− 𝑢𝑖𝐼2
).
A smaller mAIC value indicates a better parametric model.
iii. Compute new pseudo data for tree 2 means that conditional pseudo data as given by
+ −
𝑢𝑖,ℎ1 |ℎ2
: = 𝐹ℎ1 |ℎ2 (𝑦𝑖 ℎ1 |𝑦𝑖 ℎ2 ), 𝑢𝑖,ℎ1 |ℎ2
: = 𝐹ℎ1 |ℎ2 (𝑦𝑖 ℎ1 − 1|𝑦𝑖 ℎ2 ),
+ +
𝑢𝑖,ℎ1 |ℎ2
: = 𝐹ℎ2 |ℎ1 (𝑦𝑖 ℎ2 |𝑦𝑖 ℎ1 ), 𝑢𝑖,ℎ2 −1|ℎ1
: = 𝐹ℎ2|ℎ1 (𝑦𝑖 ℎ2 |𝑦𝑖 ℎ1 ),
3. Repeat step 2 for all pairs of the new pseudo data and corresponding pair copulas. Also
compute new pseudo data in a similar fashion as step 2(iii).
4. Iterate to select the pair copulas.
Once the copula families selected for each edges and consecutive trees, the next step is parameter
estimation.

69 | P a g e
© Yimam JA, UNISA 2019
 Parameter Estimation

For estimation of regular vine, details of different estimation techniques were assessed in
Chapter 3. Here, we only concerned on MLE that will be applied for this application area.

Since the 3-dimensional D-vine is derived, hence the log-likelihood function of a 3 dimensional
D-vine is given by

n
l (  , ; y ) =  log(Pr(Y1 = y1 , Y2 = y 2 , Y3 = y3 ;  , ))
i =1
3
C12 ( F1 ( y1 − i1 ; 1 ), F2 ( y 2 ;  2 );12 ) − C12 ( F1 ( y1 − i1 ; 1 ), F2 ( y 2 − 1;  2 );12 )
=  log({   (−1) i1 + i3
C13|2 ( ,
n =1 i1 = 0 ,1 i3 =0 ,1 F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )
C 23 ( F2 ( y 2 ;  2 ), F3 ( y3 − i3 ;  3 ); 23 ) − C 23 ( F2 ( y 2 − 1;  2 ), F3 ( y3 − i3 ;  3 ); 23 )
;13|2 )}[F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )]) (4.53)
F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )
Here  = ( 1 ,  2 ,  3 ) and  = (12 ,  23 , 13|2 ) where the marginal and copula parameters

respectively. Then, the ML estimator ˆ ML is obtained by maximizing the above log-likelihood

function over all parameters,  and  , simultaneously.

For the general case, let the model for the tth time ordered margin imply a marginal distribution
function Fit ( yit ; t ) , where  t are the marginal parameters and the subscript i denotes that we

observe a sample yi = ( yi1 , yi 2 , ..., yiT ) for i = 1, 2, ..., n. Similarly, the copula parameters for t-

dimensional dependence are given by  i ,i + t |i +1, ...,i + t −1 . Then, the ML estimator ˆ ML is obtained by

maximizing the log-likelihood function over all parameters,  t and  i ,i + t |i +1, ...,i + t −1 , simultaneously.

In this optimization, good starting values are required. Starting values for the marginal
parameters are obtained by maximum likelihood estimates of the marginal parameters estimated
for all one at a time. And starting values for the copula parameters can be found by computing
empirical Kendall’s τ of bivariate copula function of the first tree which act as ‘pseudo’ data and
then transformed back to the copula parameter using a known Bijection.

In many fields of specializations including clinical trials, medicine, public health, social sciences,
education, economics, psychometric and pharmacokinetics, multiple outcomes measured
70 | P a g e
© Yimam JA, UNISA 2019
repeatedly over time from the same sets of study participants to analyse the changes over time,
resulting in multivariate longitudinal data. The statistical analysis of this type of data are for
studying changes across time by reducing the dimension of the multivariate longitudinal data to
univariate longitudinal data using some kind of summary measures (Asar and İlk, 2014, Laffont
et al., 2014), or jointly addressing the associations/dependencies across multivariate covariates
and the changes across time points (Abegaz et al., 2015, Jiang, 2012, Verbeke et al., 2017). In
multivariate longitudinal outcomes, the dependence among outcomes and changes over time
must be accounted for in order to make valid inference. Our motivated example was household
food security which has similar feature with multivariate longitudinal outcomes.

As we have discussed so far in the previous section of this chapter, the availability, accessibility
and utilisation dimensions took ordinal levels based on the quartile score. For incorporating the
fourth dimension, the stability of the three dimensions over time, the dimensions were repeatedly
measured three times at six months interval. Hence, three longitudinal ordinal outcomes were
obtained, resulting in multivariate longitudinal ordinal outcomes. The dimensions have pair-wise
dependence between them at the same time point and each have dependence between
consecutive time points (Capaldo et al., 2010, FAO, 2008). Therefore, modelling the stability
and determinants of household food insecurity is the case of modelling multivariate longitudinal
ordinal data that can consider the dependence between the dimensions and the dependence of
each dimension between consecutive time points.

In modelling the stability and determinants of each household food security dimensions, the PCC
model was proposed. A nice feature of the PCC approach in this setting is measuring the
dependency of the three dimensions using the copula parameters, the parameter of the
consecutive food security status of the households and the associated determinants of household
food security for each dimension using the parameters of the marginal distributions. Hence, the
pair copula construction approach with D-vine is attractive since it allows pairwise positive
dependence structures and has closed form cumulative distribution function (cdf), no other
copula family has both these properties.

In sum, this section proposes a model for estimating dependence and marginal parameters from
multivariate longitudinal ordinal data using pair copula constructions via marginal model of the
longitudinal ordinal logistic regression to our motivating problem. The performance of pair

71 | P a g e
© Yimam JA, UNISA 2019
copula construction for discrete data was evaluated via Bernoulli and Poisson discrete
distributions by (Panagiotelis et al., 2012) and found to be a good model. In our work, we follow
their approach of constructing pair copula construction to model multivariate longitudinal ordinal
data. As far as the researcher review of literature is concerned, no work has been conducted on
multivariate longitudinal ordinal outcomes in the ordinal logistic version so far using pair copula
construction. Hence, this study is concerned on implementing the developed discrete PCC model
via marginal model of ordinal logistic regression for modelling household food insecurity status
and determinant factors.

Even thought our aim is concerned on multivariate longitudinal ordinal data, let us review the
continuous case and will continue to discrete one.

Pair Copula Construction for Multivariate Longitudinal Continuous Data

Consider an M-dimension multivariate continues random variables repeatedly measured for 𝑇

time points of the 𝑖 𝑡ℎ individual given by 𝑌𝑗𝑖 = (𝑌𝑗𝑖1 , 𝑌𝑗𝑖2 , … , 𝑌𝑗𝑖𝑇 ) where 𝑗 = 1, 2, … , 𝑀 𝑎𝑛𝑑 𝑖 =
1, 2, … , 𝑛. Smith (2015) re-ordered the observations of the multivariate series into the univariate
outcomes of dimensions 𝑁 = 𝑇 ∗ 𝑛 given by 𝑌 = (𝑌1 , 𝑌2 , … , 𝑌𝑀 ), where 𝑌1 = (𝑦11 , 𝑦21 , … , 𝑦𝑁1 )′ ,
𝑌2 = (𝑦12 , 𝑦22 , … , 𝑦𝑁2 )′ 𝑎𝑛𝑑 𝑌𝑀 = (𝑦1𝑚 , 𝑦2𝑚 , . , 𝑦𝑁𝑚 )′ (Smith, 2015).

Hence, the joint density function 𝑓(𝑦1 , 𝑦2 , … , 𝑦𝑚 ) is decomposed as follows:

f ( y1 , ..., y m ) = f1|2, ...,m ( y1 | y2 , ..., y m ) f 2|3, ...,m ( y 2 | y3 , ..., y m )... f m ( ym ) . (4.56)

Recalling equations (4.4 and 4.12), we can simplify the bivariate case to

f ( y1 , y2 ) = c12 ( F ( y1 ), F ( y2 )) f1 ( y1 ) f 2 ( y2 ) (4.57)
where c12 (., .) is the appropriate pair-copula density for the pair of transformed variables

F1 ( y1 ) and F2 ( y2 ) .

Any transformation using the factorization of Equation (4.56), Equation (4.57) and Equation
(4.4) different decomposition can be constructed. For example, the 3-dimensional case
decomposition results in

f1, 2, 3 ( y1 , y 2 , y3 ) = f1|2, 3 ( y1 | y 2 , y3 ) f 2|3 ( y 2 | y3 ) f 3 ( y3 )

(4.58)

72 | P a g e
© Yimam JA, UNISA 2019
Using the Sklar’s theorem, the conditional density of 𝑌2 𝑎𝑛𝑑 𝑌3 in equation (4.58) is given by:

f 2 , 3 ( y 2 , y3 ) c23 ( F ( y 2 ), F ( y3 )) f 2 ( y 2 ) f 3 ( y3 )
f 2|3 ( y 2 | y3 ) = =
f 3 ( y3 ) f 3 ( y3 )
= c23 ( F ( y 2 ), F ( y3 )) f 2 ( y 2 ) . (4.59)

Similarly,

f13|2 ( y1 | y 2 , y3 | y 2 )
f1|2, 3 ( y1 | y 2 , y3 ) =
f 3|2 ( y3 | y 2 )
c13|2 ( F1|2 ( y1 | y 2 ), F3|2 ( y3 | y 2 )) f1|2 ( y1 | y 2 ) f 3|2 ( y3 | y 2 )
=
f 3|2 ( y3 | y 2 )
= c13|2 ( F1|2 ( y1 | y 2 ), F3|2 ( y3 | y 2 )) f1\ 2 ( y1 | y 2 )

= c13|2 ( F1|2 ( y1 | y2 ), F3|2 ( y3 | y2 )) f1|2 ( y1 | y2 )

(4.60)
= c13|2 ( F1|2 ( y1 | y2 ), F3|2 ( y3 | y2 )) . c12 ( F ( y1 ), F ( y2 )) f1 ( y1 )

Using Equation (4.58), (4.59) and (4.60), the following decomposition appears

f1, 2, 3 ( y1 , y2 , y3 ) = c13|2 ( F1|2 ( y1 | y2 ), F3|2 ( y3 | y2 )). c12 ( F ( y1 ), F ( y2 )) f1 ( y1 ).

c23 ( F ( y2 ), F ( y3 )) . f 3 ( y3 ) . f 2 ( y2 ) . f1 ( y1 ). (4.61)

This example illustrates the construction of a 3-dimensional density using the bivariate copula
and the corresponding marginal distributions.

Similarly, for any other factor in Equation (4.56), the same procedure is possible using the
general formula as follows:

𝑓𝑖|𝑗𝑘 (𝑦𝑖 |𝑦𝑗,𝑘 ) = 𝑐𝑗𝑘|𝑘 (𝐹𝑖|𝑘 (𝑦𝑖 |𝑦𝑘 ), 𝐹𝑖|𝑘 (𝑦𝑖 |𝑦𝑘 )) 𝑓𝑖|𝑘 (𝑦𝑖 |𝑦𝑘 ), ( 4.62)

where k can be empty, a single index or multiple indices (Lennon, 2016).

For instance, the joint density of a four-dimension varieties can be decomposed into bivariate
pair copulas using (4.62) as,

𝑓(𝑦1 , 𝑦2 , 𝑦3 , 𝑦4 ) = 𝑓4|321 (𝑦4 |𝑦3 , 𝑦2 , 𝑦1 )𝑓3|21 (𝑦3 |𝑦2 , 𝑦1 )𝑓2|1 (𝑦2 |𝑦1 )𝑓1 (𝑦1 ),
= 𝑐14|23 𝑓4|23 (𝑦4 |𝑦2 , 𝑦3 ). 𝑐13|2 𝑓3|2 (𝑦3 |𝑦2 ). 𝑐12 𝑓2 (𝑦2 )𝑓1 (𝑦1 ),

73 | P a g e
© Yimam JA, UNISA 2019
 = 𝑐14|23 𝑐24|3 𝑓4|2 (𝑦4 |𝑦2 )𝑐13|2 𝑐23 𝑓3 (𝑦3 )𝑐12 𝑓2 (𝑦2 )𝑓1 (𝑦1 ),
= 𝑐14|23 𝑐24|3 𝑐24 𝑓4 (𝑦4 )𝑐13|2 𝑐23 𝑓3 (𝑦3 )𝑐12 𝑓2 (𝑦2 )𝑓1 (𝑦1 ),
= 𝑐14|23 𝑐13|2 𝑐24|3 𝑐34 𝑐23 𝑐12 𝑓4 (𝑦4 )𝑓3 (𝑦3 )𝑓2 (𝑦2 )𝑓1 (𝑦1 ), (4.63)
Hence, the decomposition in (4.63) can be written with full expression as:
𝑓(𝑦1 , 𝑦2 , 𝑦3 , 𝑦4 ) = 𝑐12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 ))𝑐23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 ))𝑐34 (𝐹3 (𝑦3 ), 𝐹4 (𝑦4 ))

× 𝑐13|2 (𝐹1|2 (𝑦1 |𝑦2 ), 𝐹3|2 (𝑦3 |𝑦2 )) 𝑐14|2 (𝐹1|2 (𝑦1 |𝑦2 ), 𝐹4|2 (𝑦4 |𝑦2 ))

× 𝑐34|12 (𝐹3|12 (𝑦3 |𝑦1 , 𝑦2 ), 𝐹4|12 (𝑦4 |𝑦1 , 𝑦2 )) 𝑓4 (𝑦4 )𝑓3 (𝑦3 )𝑓2 (𝑦2 )𝑓1 (𝑦1 )

Hence, based on the general form of equation 4.62, the decomposition of 𝑓(𝑦1 , 𝑦2 , … , 𝑦𝑚 )
according to the D-vine pair copula construction can be written as

m m−1 m− j
f (y1 , y 2 ,…, y m ) =  f ( yk ) ci , i + j|i +1, ...,i + j −1{F ( yi | yi +1 , ..., yi + j −1 ), F ( yi + j | yi +1 , ..., yi + j −1 )}, (4.64)
k =1 j =1 i =1

Equation (4.64) is a product of 𝑚 ∗ (𝑚 − 1)/2 bivariate pair copula densities and m marginal
densities (Aas et al., 2009, Czado, 2010, Lennon, 2016, Ruscone and Osmetti, 2017, Smith et al.,
2010, Smith, 2015). This leads to a large number of possible pair-copulas constructions. To
organize all possible decompositions, a graphical model called a regular vine has been
introduced by (Bedford and Cooke, 2002). Regular vine decompositions are concentrated only
on the D-vines and C-vines, the special cases of regular vines. Detail of D-vine construction was
displayed earlier in section 4.2.3.

D-Vine Parameter Estimation

For estimation of regular vine, different scholars proposed non-standard methods and standard
estimation methods. Stepwise and MLE, IFM and SSP are the common standard estimation
methods. MLE were considered for the first time by (Aas et al., 2009), IFM by (Joe, 1996), and
SSP were by (Haff, 2012). These methods are designed for continuous data. We will not discuss
here in detail since the current concern is on the discrete data. One can refer the references cited
here for more detail. Just we now go to the PCC in discrete data.

74 | P a g e
© Yimam JA, UNISA 2019
 Pair Copula Construction for Multivariate Longitudinal Discrete Data

Like the continuous case, consider an M-dimension multivariate discrete random variable
repeatedly measured for 𝑇 time points of the 𝑖 𝑡ℎ individual given by 𝑌𝑗𝑖 = (𝑌𝑗𝑖1 , 𝑌𝑗𝑖2 , … , 𝑌𝑗𝑖𝑇 )
where 𝑗 = 1, 2, … , 𝑀 𝑎𝑛𝑑 𝑖 = 1, 2, … , 𝑛. We re-ordered the observations of the multivariate
series into the univariate outcomes of dimensions 𝑁 = 𝑇 ∗ 𝑛 given by 𝑌 = (𝑌1 , 𝑌2 , … , 𝑌𝑀 ), where
𝑌1 = (𝑦11 , 𝑦21 , … , 𝑦𝑁1 )′ , 𝑌2 = (𝑦12 , 𝑦22 , … , 𝑦𝑁2 )′ 𝑎𝑛𝑑 𝑌𝑀 = (𝑦1𝑚 , 𝑦2𝑚 , . , 𝑦𝑁𝑚 )′ .
Hence, the joint probability mass function 𝑃𝑟(𝑌1 = 𝑦1 , 𝑌2 = 𝑦2 , … , 𝑌𝑚 = 𝑦𝑚 ) is decomposed as
follows:
Pr(𝑌1 , 𝑌2 , … , 𝑌𝑚 ) = Pr(𝑌1 = 𝑦1 |𝑌2 = 𝑦2 , … , 𝑌𝑚 = 𝑦𝑚 ) × Pr(𝑌2 = 𝑦2 |𝑌3 = 𝑦3 , … , 𝑌𝑚 = 𝑦𝑚 )
× … × Pr(𝑌𝑚 = 𝑦𝑚 ) (4.65)
The bivariate cumulative distribution function of (𝑌1 , 𝑌2 ) is given by the standard notation as
𝐹𝑖𝑗 (𝑦𝑖 , 𝑦𝑗 ) = Pr (𝑌𝑖 ≤ 𝑦𝑖 , 𝑌𝑗 ≤ 𝑦𝑗 ) (4.66)
Similarly, the conditional cumulative distribution is given as follows:
𝐹𝑖𝑗|𝑘 (𝑦𝑖 , 𝑦𝑗 |𝑦𝑘 ) = Pr (𝑌𝑖 ≤ 𝑦𝑖 , 𝑌𝑗 ≤ 𝑦𝑗 |𝑌𝑘 = 𝑦𝑘 )
(4.67)
The expression in equation (4.65) has terms of the form Pr(Y j = y j | Y| j = y| j ) where Y| j is the

vector of random variables Y1 , Y2 , ..., Ym excluding Y j and y| j is the same vector for the realized

values of the random variables. Choosing another element ℎ from the vector of random
variables, we can re-write the discrete joint probability in a similar fashion to the continuous case
as following:
Pr(Y j = y j , Yh = yh | Y| j , h = y| j , h )
Pr(Y j = y j | Y| j = y| j ) = (4.68)
Pr(Yh = y h | Y| j , h = y| j , h )

Now, recalling the probability mass function and the multivariate copula function for discrete
data in Equation (4.5), the bivariate conditional probability in the numerator can be expressed in
terms of a copula giving

75 | P a g e
© Yimam JA, UNISA 2019
Pr(Y j = y j | Y| j = y|h )


i +ih
(−1) j Pr(Y j  y j − i j , Yh  yh − ih | Y| j .h = y| j .h )
i j =0 ,1 ih =0 ,1
=
Pr(Yh = yh | Y| j .h = y| j .h )

  (−1)
i j +ih
CY j .Yh | y| j . h ( FY j |Y| j .h ( y j − i j ), FY j |Y| j . h ( yh − ih ))
i j =0 ,1 ih =0 ,1
=
Pr(Yh = yh | Y| j .h = y| j .h ) (4.69)
The arguments in equation (4.69) of the copula functions are evaluated using the following
(Nicklas, 2013, Panagiotelis et al., 2012, Sirisrisakulchai and Sriboonchitta, 2014, Stöber et al.,
2015);

FY j |Yh .Y| j . h ( y j | yh , y| j .h ) = [CY j .Yh|Y| j . h ( FY j \ |Y| j .h ( y j | y| j .h ), FYh|Y| j . h ( yh | y| j .h ))

− CY j .Yh|Y| j .h ( FY j \ |Y\ j . h ( y j | y| j .h ), FYh|Y| j .h ( yh − 1 | y| j .h ))]
/ Pr(Yh = yh | Y| j .h = y| j .h )
(4.70)

This vine PCC has nice feature than multivariate as well as Gaussian copulas functions in
evaluating the probability mass function because the PCC requires 2m(m − 1) evaluations

whereas the multivariate and Gaussian copulas require 2 m evaluations (Panagiotelis et al.,
2012).

D vine in Multivariate Longitudinal Discrete Data

For illustration purposes, we present in detail the 3-dimensional case. Therefore,

Pr(Y1 = y1 , Y2 = y2 , Y3 = y3 )
(4.71)
= Pr(Y1 = y1 | Y2 = y2 , Y3 = y3 )  Pr(Y3 = y3 | Y2 = y2 )  Pr(Y2 = y2 )

Utilizing Equation (4.71) the right-hand side of the first conditional probability can be rewritten
as:

76 | P a g e
© Yimam JA, UNISA 2019
 Pr(Y1 = y1 | Y2 = y2 , Y3 = y3 )
  (−1)
i1 =0 ,1i3 =0 ,1
i1 +i3
C13|2 ( F1|2 ( y1 − i1 | y2 ), F3|2 ( y3 − i3 | y2 ))
(4.72)
=
Pr(Y3 = y3 | Y2 = y2 )

Similarly utilizing Equation (4.69), the first argument of the copula function in the numerator of
Equation (4.72) is given by

C12 ( F1 ( y1 − i1 ), F2 ( y2 )) − C12 ( F1 ( y1 − i1 ), F2 ( y2 − 1))

F1|2 ( y1 − i1 | y2 ) = , (4.73)
Pr(Y2 = y2 )

and the second argument can be expressed as

C23 ( F2 ( y2 ), F3 ( y3 − i3 )) − C23 ( F2 ( y2 − 1), F3 ( y3 − i3 ))

F3|2 ( y3 − i3 | y2 ) = (4.74)
Pr(Y2 = y2 )

By cancelling terms and substituting, the probability mass function of the full expression for the
3-dimensional discrete D-vine is given by

Pr(Y1 = y1 , Y2 = y2 , Y3 = y3 )
C12 ( F1 ( y1 − i1 ), F2 ( y2 )) − C12 ( F1 ( y1 − i1 ), F2 ( y2 − 1))
={  (−1) i1 +i3
C13|2 ( ,
i1 =0 ,1i3 =0 ,1 F2 ( y2 ) − F2 ( y2 − 1) (4.75)

C23 ( F2 ( y2 ), F3 ( y3 − i3 )) − C23 ( F2 ( y2 − 1), F3 ( y3 − i3 ))

)}[F2 ( y2 ) − F2 ( y2 − 1)].
F2 ( y2 ) − F2 ( y2 − 1)

Somewhat confusing to write the general D-vine structure; however, the general dimension
algorithm for computing the probability mass function of a D-vine was outlined by Panagiotelis
et al. (2012). It is evident both from this algorithm and the 3-dimensional example above that

each bivariate pair copula only needs to be evaluated 4 times, specifically CY j .Yh |Y| j .h must be

evaluated (Panagiotelis et al., 2012, Nicklas, 2013, Sirisrisakulchai and Sriboonchitta, 2014,
Stöber et al., 2015);

( FY j \ |Y| j .h ( y j | y| j.h ), FYh |Y| j .h ( yh | y| j.h )) , ( FY j \ |Y| j .h ( y j − 1 | y| j.h ), FYh |Y| j .h ( y h | y| j.h )) ,

( FY j \ |Y| j . h ( y j | y| j.h ), FYh |Y| j .h ( y h − 1 | y| j .h )) and ( FY j \ |Y| j .h ( y j − 1 | y| j .h ), FYh |Y| j . h ( y h − 1 | y| j .h )) .

77 | P a g e
© Yimam JA, UNISA 2019
In general, evaluation of the probability mass function requires 2m(m − 1) evaluations of
bivariate copula functions, even though the continuous vine PCC is composed of only m(m−1)/2
pair copulas. These vine PCCs have still greater potential in high-dimensional settings since the
computational burden of evaluating the pmf in the elliptical copulas is 2m. As pointed out by
Panagiotelis et al. (2012) among the major advantage of D-vine PCCs, a wide variety of
dependence structures can be modelled by selecting different copula families as building blocks.
Among these, Gaussian, t, AMH, Clayton, Frank and Gumbel copulas are the commonly used
parametric copula families as building blocks.

The marginal probabilities in equation (4.75) can be modelled either of among discrete
probability distributions. For the current study since the data are multivariate longitudinal ordinal
data, the marginal model for univariate longitudinal ordinal data via cumulative logit model is
used. Details of this model are given below.

Marginal Model of longitudinal ordinal outcomes

Let 𝑌 denotes ordinal response variable observed over 𝑇 time points such that the response
variable has C categories and labelled (1,2, … , 𝐶 − 1) and 𝑌𝑖𝑡 is the 𝑖 𝑡ℎ individual at time 𝑡 for
𝑖 = 1, 2, … , 𝑛 and 𝑡 = 1, 2, … , 𝑇. The marginal model for univariate longitudinal ordinal
outcome via the cumulative logit model is given by:

𝑙𝑜𝑔𝑖𝑡𝑃(𝑌𝑖𝑡 ≤ 𝑐|𝑋, 𝛼𝑐, 𝛽) = 𝛼𝑐 + 𝑋 ′ 𝛽 𝑓𝑜𝑟 𝑐 = 1, 2, … , 𝐶 − 1 (4.76)

where 𝑋 is a vector of fixed or time varying covariates and 𝛽 is a vector of unknown regression
coefficients (Abegaz et al., 2015). Equation 4.76 implies

exp(𝛼𝑐 + 𝑋 ′ 𝛽) exp(𝛼𝑐−1 + 𝑋 ′ 𝛽)
𝑙𝑜𝑔𝑖𝑡𝑷(𝑌𝑖𝑡 ≤ 𝑐|𝑋, 𝛼𝑐, 𝛽) = − (4.77)
1 + exp(𝛼𝑐 + 𝑋 ′ 𝛽) 1 + exp(𝛼𝑐−1 + 𝑋 ′ 𝛽)

Equation 4.76 or 4.77 is the marginal model of the proportional odds univariate ordinal logistic
regression model. The parameters of this model are obtained by maximizing the likelihood
function defined in 4.77. The regression coefficients in this model have simple interpretations in
the population-average interpretation fashion in-terms of odds ratios. Equation (4.77) serves for

78 | P a g e
© Yimam JA, UNISA 2019
the pair copula construction as a marginal distribution of the multivariate longitudinal ordinal
cumulative logistic regression model.

Hence, the arguments in equation (4.75) can take the marginal distribution function given in
equation (4.77). Moreover, the joint probability mass function in question (4.71) is expressed in
terms of the pair copula functions and longitudinal ordinal marginal distributions. The newly
constructed joint probability mass function can be called pair copula based multivariate
longitudinal (marginal) cumulative logit model. Finally, this function can be estimated using the
appropriate parameter estimation technique and appropriate bivariate copula families as building
blocks.

Selection of Pair Copula Families

Before parameter estimations of the pair copula construction model, determining the order of the
D-vine and the appropriate bivariate copula families are required. For the current study on food
security, the conceptual framework for food security dimensions is given by FAO (2008) like
Availability-Accessibility-Utilisation as it was discussed in section 4.3 of this thesis. We can use
this order for the structure of D-vine for inference purpose. The difference in this chapter is that
each dimension of food security consists of 𝑁 = 𝑇 ∗ 𝑛 length of dataset. However, in section 4.3
the length of each dimension was only 𝑛, where 𝑇 𝑎𝑛𝑑 𝑛 are the number of data collection
phases and sample size, respectively. The reason that the length of the dataset became 𝑁 = 𝑇 ∗ 𝑛
was because of reordering the multivariate time series based on the time order into one vector for
each dimension.

The next step was selection of bivariate copula families of the vine distribution using the D-vine
for every pair of variables. In Chapter 3, we customized the algorithm (Algorism I) developed
for discrete pair copula bivariate copula family selection by Panagiotelis et al. (2015). We
followed the same fashion of this algorithm for this chapter. For this chapter we have used
marginal model via cumulative logit model to compute the “Pseudo” data. However, in Chapter
3, we have used simply cumulative logit model. “Pseudo” data were computed from the marginal
model via cumulative logit model using Algorithm I to select appropriate bivariate pair families
that fit the model best which have the smallest modified Akaike Information Criterion (mAIC).

79 | P a g e
© Yimam JA, UNISA 2019
 Parameter Estimation of the D-Vine

For estimation of regular vine, details of different estimation techniques were assessed in
Chapter 3. Here, we only concerned on MLE that will be applied for this application area.

Since the 3-dimensional D-vine is derived, hence the log-likelihood function of a 3-dimensional
D-vine is given by

n
l (  , ; y ) =  log(Pr(Y1 = y1 , Y2 = y 2 , Y3 = y3 ;  , ))
i =1
3
C12 ( F1 ( y1 − i1 ; 1 ), F2 ( y 2 ;  2 );12 ) − C12 ( F1 ( y1 − i1 ; 1 ), F2 ( y 2 − 1;  2 );12 )
=  log({   (−1) i1 + i3
C13|2 ( ,
n =1 i1 = 0 ,1 i3 = 0 ,1 F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )
C 23 ( F2 ( y 2 ;  2 ), F3 ( y3 − i3 ;  3 ); 23 ) − C 23 ( F2 ( y 2 − 1;  2 ), F3 ( y3 − i3 ;  3 ); 23 )
;13|2 )}[F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )]) (4.78)
F2 ( y 2 ;  2 ) − F2 ( y 2 − 1;  2 )
Here  = ( 1 ,  2 ,  3 ) and  = (12 ,  23 , 13|2 ) where the marginal and copula parameters

respectively. Then, the ML estimator ˆ ML is obtained by maximizing the above log-likelihood

function over all parameters,  and  , simultaneously.

For the general case, let the model for the jth margin of the tth time order imply a marginal
distribution function 𝐹𝑖𝑗 (𝑦𝑖𝑗 , 𝛽𝑗 ), where 𝛽𝑗 are the marginal parameters and the subscript 𝑖
denotes that we observe a sample 𝑦𝑖 = (𝑦𝑖1 , 𝑦𝑖2 , . . . , 𝑦𝑖𝑀 )′ for 𝑖 = 1, 2, … , 𝑇 ∗ 𝑛. Similarly, the
copula parameters for M-dimensional dependence are given by 𝜃𝑖,𝑖+𝑗|𝑖+1,… ,𝑖+𝑚−1 . Then, the ML

estimator ˆ ML is obtained by maximizing the log-likelihood function over all parameters, 𝛽𝑗

and 𝜃𝑖,𝑖+𝑗|𝑖+1,… ,𝑖+𝑚−1, simultaneously.

In this optimization, good starting values are required. Starting values for the marginal
parameters are obtained by maximum likelihood estimates of the marginal parameters estimated
from the marginal model via cumulative logit model for all dimensions. In addition, starting
values for the copula parameters can be found by computing empirical Kendall’s τ of bivariate
copula function of the first tree which act as ‘pseudo’ data and then transformed back to the
copula parameter using a known Bijection like we have computed for sections 4.3 and 4.4.

80 | P a g e
© Yimam JA, UNISA 2019
“VineCopula” R package was used to compute the modified Akaike Information Criterion
(mAIC) for selecting the best fitted bivariate copula families (Schepsmeier et al., 2015). Since
the package did not include for t-bivariate copula owing to computational challenges, the t-
copula is not proposed as a candidate in this thesis.

An “Alabama” R package was implemented to jointly estimate the marginal and copula
parameters and their respective standard errors. Alabama is Augmented Lagrangian
Minimization Algorithm for optimising smooth nonlinear objective functions with constraints. It
allows both for Linear or nonlinear equality and inequality constraints (Varadhan and
Grothendieck, 2011). It optimises using the nonlinear optimization with constraints. The package
addressed this issue using “auglag” optimization. We wrote our own R code using “auglag”
optimisation R package to estimate the parameters of the copula and the marginal distribution
functions. Details of the R code were displayed in Appendix C I-III for all the three models,
respectively.

Variables that had p-value less than 0.2 in the preliminary analysis of the univariate analysis
were incorporated for the final model of the univariate model. The significant variables in the
final model were selected using forward wald method with significance level of 0.05. Variables
that were significance with 0.05 p-values were incorporated for the multivariate, longitudinal and
multivariate longitudinal ordinal models.

All statistical significant variables in each of the marginal model via the cumulative logit model
were incorporated in the model for each dimension that hopes to be helpful as additional
information with the existing knowledge in this area. In some of the tables of the result section,
the blank space indicated that particular variable was not statistically significant for that
dimension in the marginal model and hence that particular variable was not included only for
that particular dimension during the estimation of the final joint model.

81 | P a g e
© Yimam JA, UNISA 2019
 Chapter Five

Analysis of the Household Food Data

Internal consistency of the data collection instrument for household food security in terms of the
three dimensions was assessed using Cronbach’s Alpha. The Cronbach’s Alpha value computed
as 0.735 which is in the acceptable internal consistence level since it is above 0.7. The
Cronbach’s Alpha value of a particular item deleted was assessed for each item. The detail
summary statistics was displayed in appendix B.

Three phases of data collection were conducted from the same household head at six months
interval. The study included a total of 630 households after the removal of 2.5 % (16) of the
respondents because they have dropped out at least one of the data collection phases. Since the
study was longitudinal and have both time varying and fixed covariates, for the time varying
variables we used all responses from the three follow-up interviews and arranged the data into a
vector according to the time points while for fixed covariates we used the response of one time
point. The summary statistics is displayed in Tables 5.1 and 5.2 for the fixed and time-varying
covariates, respectively.

Household Head Characteristics

Table 5.1 showed that husbands headed majorities of the household heads (more than 85%).
Almost 69% of the households did not have formal education of which 38% were unable to read
and write and 31% can read and write. Only 31% of the household heads attended formal
education of which 6% was secondary school or above completion. Two thirds of the household
heads were married and living together. Twelve percent of them were cohabitating but not
married and 7% were widowed.

Table 5.1 also showed majorities of the households consisted of four to six family members
which accounts 65% of the total respondents. Nineteen percent of the respondents had more than
six family members whereas 18% of them have less than four family members.

82 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 1: Summary measures of non-time varying variables of the households

Variables Frequency Percent

Household Head (HH)
Husband 548 86.9
Wife 71 11.3
Son/daughter 11 1.7
Highest level of education attained by HH
Unable to read and write 241 38.2
Can read and write 192 30.5
Regular Primary education (1-8) 157 24.9
Secondary education and above 40 6.4
HH current marital status
Never married 13 2
Cohabiting 74 11.7
Married 475 75.4
Divorced 23 3.7
Widowed 45 7.2
Family Size of HH
Three and Less 110 17.5
Four to Six 400 63.5
Seven and Above 120 19.0
Study Site
Kutaber 204 32.5
Kalu 234 37.2
Tehuledere 192 30.5
Total farmland size in Hectare
<=0.5 Hectare 383 60.8
> 0.5 Hectare 247 39.2

83 | P a g e
© Yimam JA, UNISA 2019
Table 5. 1: Summary measure of non-time varying variables of the households … continued
Variables Frequency Percent
Quality or fertility of land ploughed
Fertile 75 11.9
Medium fertile 422 66.9
Less fertile 133 21
HH Income Source
Only Farming 261 41.4
Both Farming and Off-farming 369 58.6
Types of Cereal Crops Cultivated
One type 169 26.8
two type 187 29.7
Three and more types 274 43.5
Types of livestock
One and less type 156 24.8
Two to Three types 275 43.7
Four or more types 198 31.5
Agro-ecology of study site
Hot (Kolla) 146 23.2
Medium (Weinadega) 340 54.0
Cold (Dega) 144 22.9

Income Source of the Household

As Table 5.1 showed, 44% of the households obtained their income from farming only whereas
59% of them practice off-farming activities to get their income. About 73% of the households
had harvested more than one type of cereal crops. In addition to farming, the households
participated in different agricultural activities, among these two thirds of the households
participated in more than one type of livestock activities.

84 | P a g e
© Yimam JA, UNISA 2019
Land Size and Fertility Characteristics

Based on the information in Table 5.1, around 60% of the households ploughed not greater than
0.5 hectares and 67% of them had medium fertile land for cultivation purposes. Only 11%
households had fertile land for plough and 21% less fertile. Among the total households in this
sample, 74% of them cultivated once per year whereas 26% of them cultivated twice or triple per
year.

Table 5. 2: Summary measures of the time-varying variables of the households

Data collection phases

Characteristic variables Phase one Phase two Phase three
Frequency Percent Frequency Percent Frequency Percent
Shortage of rainfall
No 117 18.6 378 60.0 92 14.6
Yes 513 81.4 252 40.0 538 85.4
Crop Disease
No 422 67.0 328 52.1 421 66.8
Yes 208 33.0 302 47.9 209 33.2
Increase in market price
No 385 61.1 476 75.6 452 71.7
Yes 245 38.9 154 24.4 178 28.3
Use of Pesticides
No 501 79.5 332 52.7 501 79.5
Yes 129 20.5 298 47.3 129 20.5
Presence of Pests
No 259 41.1 239 37.9 285 45.2
Yes 371 58.9 391 62.1 345 54.8

Agro-Ecology of the Study Site

The study was conducted in three selected South Wollo Zone Woredas, namely, Kutaber, Kalu
and Tehuledere. As Table 5.1 revealed that around 33, 37 and 31 percent of the households were
from Kutaber, Kalu and Tehuledere Woredas respectively. Moreover, almost half (54%) of the
85 | P a g e
© Yimam JA, UNISA 2019
study areas have medium (Weinadega) agro-ecology and the rest proportion was covered by hot
(Kola) and cold (Dega) agro-ecology with almost equal proportion.

Time-Varying Characteristics

Table 5.2 shows the distribution of some of the time-varying covariates during the three-phase
data collection. Based on Table 5.2, the presence of rainfall, disease of the cultivation, presence
of market price increase, use of pesticides, and presence of pests vary with time in the study area.
More than 80% of rainfall shortage was observed in the first and third phases, and smallest 40%
during the second phase.

Moreover, Table 5.2 showed that 48% of cultivation disease was reported in the second phase
but 33% each in the first and third phase. The highest market price increase was observed in
phase one followed by phase three and two, 39%, 28%, and 24% respectively. The presence of
pests had highest proportion in phase two followed by phase one and three, 62, 59 and 55
percent, respectively.

The household food security status was computed for each dimension at each data collection
phase using the quartile score as presented in Table 5.3. Table 5.3 showed that food secured
households was smaller in phase three for food availability and utilisation dimensions than
phases one and two. Conversely, the proportion of food secured households was higher in phase
three for accessibility dimension than the other phases.

Based on the information in Table 5.3, the proportion of chronically food in-secured households
decreased from 19% to 2% and from 3% to 0% in accessibility and utilisation dimensions
respectively when one goes from phases one and two to phase three. However, in availability
dimension, the proportion of chronically food in-secured was almost similarly distributed
between the three data collection phases.

Table 5.3 also showed the highest proportion of mildly food in-secured (60%) was found in food
utilisation during the third phase followed by 42% in accessibility. The highest proportion (51%)
and (50%) of moderately food in-secured households were observed in utilisation of phase one
and two respectively followed by 45% in availability of phase three.
86 | P a g e
© Yimam JA, UNISA 2019
Furthermore, the household food security status was determined using the composite index
method combining the household food security status of the three data collection phases into
one vector for each dimension. The result of this computation was summarized in Table 5.4.

Table 5. 3: Household food security statuses of the three food security dimensions in all
data collection rounds (f stands for frequency and % for percent)

Data collection phases

Food Household food security Phase I Phase II Phase III
Security Status f % F % F %
Dimensions
Chronically food in-secured 32 5.1 33 5.2 41 6.5
Availability Moderately food in-secured 215 34.1 218 34.6 284 45.1
Mildly food in-secured 254 40.3 254 40.3 209 33.2
Food secured 129 20.5 125 19.8 96 15.2
Total 630 100.0 630 100.0 630 100.0
Chronically food in-secured 119 18.9 118 18.7 14 2.2
Accessibility Moderately food in-secured 219 34.8 216 34.3 230 36.5
Mildly food in-secured 222 35.2 223 35.4 263 41.7
Food secured 70 11.1 73 11.6 123 19.5
Total 630 100.0 630 100.0 630 100.0
Chronically food in-secured 20 3.2 20 3.2 1 .2
Utilisation Moderately food in-secured 318 50.5 314 49.8 228 36.2
Mildly food in-secured 213 33.8 216 34.3 379 60.2
Food secured 79 12.5 80 12.7 22 3.5
Total 630 100.0 630 100.0 630 100.0

Based on the composite index summary statistics presented in Table 5.4, the highest proportion
of chronically food in-secured was observed in accessibility dimension (13.3%) followed by in
availability and utilisation (5.6% and 2.2%) respectively. Similarly, in-terms of moderately food
in-secured and mildly food in-secured classification, the highest proportion were observed in
utilisation followed by availability and then accessibility. Moreover, the highest proportion of

87 | P a g e
© Yimam JA, UNISA 2019
food secured households was observed in availability followed by accessibility and then
utilisation.

Table 5. 4: The Composite Food Security Status of households in the three data collection rounds

Availability Accessibility Utilisation

Food security status Frequency Percent Frequency Percent Frequency Percent
Chronically food in-secured 106 5.6 251 13.3 41 2.2
Moderately food in-secured 717 37.9 665 35.2 860 45.5
Mildly food in-secured 717 37.9 708 37.5 808 42.8
Food secured 350 18.5 266 14.1 181 9.6
Total 1890 100.0 1890 100.0 1890 100.0

Kendall's tau Correlation Coefficient

The bivariate correlation coefficients between the food security dimensions were computed
through the non-parametric correlations called “Kendall's tau_b” using the raw discrete food
security data and presented in Table 5.5.

Similarly, the Kendall’s tau was also computed from the pseudo data of food security dimensions
(Availability, Accessibility and Utilisation) using cumulative logit marginal distributions as
presented in Table 5.6 and this result is further used for preliminary analysis to select appropriate
copula families for the copula cumulative logit model.

Table 5. 5: Nonparametric correlation coefficients and their significant α values in

brackets of the household food security status of the dimensions

Availability Accessibility Utilisation

Availability 1 -0.051 (0.128) .092 (0.011)
Accessibility 1 -.199 (.000)
Utilisation 1

88 | P a g e
© Yimam JA, UNISA 2019
Based on the results in Table 5.5 and 5.6, one can observe that almost all similar and consistent
results were found and conclude that food utilisation had positive and significant correlation with
food availability but negative with accessibility.

Table 5. 6: The Kendall’s tau for the pseudo data computed from the cumulative logit
marginal distribution for the application data availability, accessibility and utilisation.

Dimensions Availability Accessibility Utilisation

Availability 1.000000 -0.05132 0.091865
Accessibility -0.05132 1.000000 -0.19932
Utilisation 0.091865 -0.19932 1.000000

PCC Selection

The preliminary analysis computed in Table 5.7 reduces the number of bivariate copula families
and the result showed that the Clayton and Gumbel copulas are not applicable for measuring the
dependence between availability, accessibility and accessibility and utilisation because their
Kendall’s tau showed negative dependence.

For those satisfied the preliminary analysis, Algorithm I was employed and their corresponding
modified Akaike Information Criterion (mAIC) were computed as presented in Table 5.7. The
D-vine structure of household food security in the first tree was (availability, accessibility) and
(accessibility, utilisation). Similarly, in the second tree the structure is (availability | accessibility
and utilisation | accessibility). From Table 5.7, one can deduce that the best fitted copula families
on the first tree were AMH for edge 1 since the mAIC is smaller than the other copula families.
However, for edge 2 the Gaussian and Frank copulas have negligible mAIC difference between
them. As a result, we set additional criterion in-terms of parsimonious that the model provides.
The Gaussian copula models are a natural choice for integer-valued covariates with interpretable
parameters (Lennon, 2016). A Frank copula can capture a wide range of dependence including
positive and negative dependence and belong to the Archimedean family with a closed form of
distribution functions and benefits of easy computation (Yang et al., 2020). Frank bivariate
copula is the ideal candidate in this study. Hence, AMH and Frank copula families were selected
as best fit bivariate copula that serves for parameter estimation in the full maximum likelihood
parameter estimation for the corresponding vine structure they fitted best.
89 | P a g e
© Yimam JA, UNISA 2019
Table 5. 7: The summary of copula families by applying Algorithm 1 to the data and estimating
modified Akaike Information Criterion (mAIC) to select the best fit copula families.

Modified Akaki Information Criterion (mAIC)

Copula Av, Ac Ac, Ut Av |Ac, Ut |Ac
Gaussian 4736.193 4465.663 4281.878
Clayton NA NA 4290.001
Gumbel NA NA 4281.459
Frank 4736.951 4465.641 4280.713
AMH 4736.076 4469.087 4282.067
Independent 4739.37 4498.055 4285.219
where Av = Availability, Ac = Accessibility and Ut = Utilisation

Estimation of the Copula and Marginal Parameters

Table 5.7 showed the dependence between availability and accessibility was expressed by AMH
copula. On the other hand, Frank expressed the dependence between accessibility and utilisation
dimensions. Similarly, like the pervious one the dependence between availability given
accessibility and utilisation given accessibility was also expressed by Frank copula. Here, the
conditioning of accessibility is assumed to be not affecting the dependence between availability
and utilisation.

Our main purpose in this chapter was to estimate jointly the dependence between food security
dimensions and their corresponding predictor factors at household level. We applied MLE using
the selected bivariate copula families and the cumulative logit marginal distribution functions.
Derivation of the likelihood and log-likelihood function was made as well (see Appendix B I).

We wrote our own R code using “auglag” optimization R package to estimate the parameters of
the copula and the marginal distribution functions. Details of the R code were displayed in
Appendix C I. The estimated values of the parameters and the corresponding standard errors for
the copula functions were displayed in Table 5.8 and for marginal distributions in Table 5.9.

90 | P a g e
© Yimam JA, UNISA 2019
Copula Parameter

Table 5.8 summarizes the results of the estimated dependence parameters of the selected
bivariate copula families. The result showed positive dependence was observed between all
dimensions of household food security statuses. AMH copula measures the dependence between
availability and accessibility household food security status, Frank measures the accessibility and
utilisation and Frank availability and utilisation given that the accessibility dimension has been
happened and found that positive dependence was observed. Moreover, these dependences were
observed as statistically significant.

Table 5. 8: The estimates of dependence parameters using the selected pair copula for the
application data of multivariate ordinal household food security status.

Tree Copula Estimated Estimated SE Bijection tau

family Parameter
I AMH 0.999 0.2612 0.3333267

Frank 1.4053 0.2425 0.1531605

II Frank 1.18358 0.3152 0.1297094

Marginal Parameter

Several variables were incorporated in the model that hopes to be helpful as additional
information with the existing knowledge in this area. Among the incorporated variables, the
presence of crops/vegetables/fruits disease, shortage of rainfall, cultivating once a year and small
land size cultivated were identified the potential statistically significant variables that lead the
household to be chronically, mildly and moderately food in-secured in all dimensions.

On the other hand, Table 5.9 showed that study site contributes on the status of household food
insecurity in availability and utilisation dimensions. Moreover, cold (dega) agro-ecology leads
household to be chronically, mildly, and moderately food in-secured in availability and
accessibility dimensions. Similarly, the presence of market price increase and household head
headed by son/daughter were more likely to be chronically, mildly and moderately food in-
secured than headed by husband at availability and utilisation dimensions respectively.

91 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 9: Summary results of pair copula based cumulative logit model parameters
estimates for the household food security data.

. Availability Accessibility Utilisation

Variable Categories Estimates S.E Estimates S.E Estimates S.E

1|2 -4.3413* 1.123 -4.6699* 1.085 -1.58125 1.189

Intercept 2|3 -1.4085 1.105 -2.6982* 1.075 2.16749 1.198
3|4 0.901668 1.104 -0.61065 1.070 4.2333* 1.202
Household Head (HH)
Husband -0.41414 .608 -0.64717 .594 -1.2361* .606
Wife -0.69952 .638 -0.73125 .621 -0.9262 .633
Sibling
Education Level (HH)
Regularly educated -0.16271 .184 -0.16329 .176 -0.17596 .185
Regularly uneducated
Marital Status (HH)
Never married or Cohabiting -1.0803* .370 0.67283* .253 -0.56257 .367
Married -0.14025 .315 0.380646 .301 -0.3404 .310
Divorced or Widowed
Study Site (Woredas)
Kutaber -1.2178* .476 -0.13779 .457 1.04154* .427
Tehuledere -0.39089 .452 0.183894 .437 0.81636* .408
Kalu
Land Size Cultivated
Less than 0.5 Hectare 0.45521* .161 0.06726* .029 0.24435* .110
Above 0.5 Hectare
Land Fertility
Fertile -0.49066 .311 -0.0784 .297 -0.58645 .322
Medium fertile -0.4183 .208 -0.07936 .197 -0.08722 .205
Less fertile
* indicates significant at 5% level of significance

92 | P a g e
© Yimam JA, UNISA 2019
Table 5.9 continued……………. (* indicates significant at 5% level of significance)
Availability Accessibility Utilisation
Variable Categories Estimates S.E Estimates S.E Estimates S.E
Cultivate time
Yearly 0.94121* .438 -1.3246* .427 1.3303* .496
Biannual and more
Shortage of Rainfall
Yes 0.64373* .235 0.56948* .225 0.34071* .144
No
Crops/Vegetables Disease
Yes 1.6343* .187 0.20649* .068 0.24712* .076
No
Market Price Increase
Yes -0.5452* .171 0.217151 .163 -0.06202 .172
No
Agro-ecology of study site
Cold (Dega) -1.5768* .356 -0.8357* .337 -0.04717 .351
Medium (Wenadega) -1.1473* .285 -0.48742 .271 0.08621 .282
Hot (Kolla)

Effects of PCC on the Univariate Cumulative Logit Model

To assess the effect of the PCC model on the usual univariate cumulative logit model, the
univariate estimates for each dimension was displayed in Table 5.10. The following comparisons
were made between the marginal parameters of the PCC-Based cumulative logit model presented
in Table 5.9 and univariate cumulative logit model presented in Table 5.10. The PCC- based
cumulative logit model identified more significant determinants for households to be food in-
secured in all of the three dimensions over the univariate cumulative logit model.

In availability dimension, both PCC and univariate cumulative logit models identified marital
status classified under never married or cohabitated, study site, small cultivable land, shortage of

93 | P a g e
© Yimam JA, UNISA 2019
rainfall, presences of crop disease, presences of market price increase, and clod agro-ecology as
positive predictors for households to be severe to mildly food in-secured compared with their
counterparts. In addition to these determinants, the PCC model identified medium agro-ecology
and yearly once cultivation season as positive predictors for household to be food in-secured.
Conversely, the univariate model identified less fertile cultivable land as positive predictor for
food insecurity.

A yearly based cultivation of agricultural activities and cold agro-ecology were the determinants
of household food insecurity in the accessibility dimension both in the PCC and univariate
cumulative logit model. However, the PCC model identified additional determinants for
households to be food in-secured include never married or cohabitated household headed, small
cultivable land, shortage of rainfall, and occurrences of crop disease.

The study site, small cultivable land and yearly once cultivation activities were identified as
significant determents for households to be severe to mildly food in-secured in the utilisation
dimension both in the two models. Like accessibility dimension, the PCC identified marital
status grouped under never married or cohabitated and occurrences of cultivation diseases as
additional determinants for households to be food in-secured.

The majority of the estimates obtained through PCC model were overestimated in availability
and accessibility dimension. However, in utilisation dimension, the PCC model underestimated
almost all of the determinants of household food insecurity. Since PCC identified more
significant determinants in all of the three food security dimensions, interpretation and
discussion of determinants were made using the estimates of the PCC model throughout this
topic.

94 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 10: Summary results of the univariate cumulative logit model parameters
estimates for the household food security data.

. Availability Accessibility Utilisation

Variable Categories Estimates S.E Estimates S.E Estimates S.E

1|2 -5.861* .810 -1.037 .761 -6.562* 0.893

Intercept 2|3 -3.057* .785 0.763 .763 -2.793* 0.837
3|4 -0.696 .774 2.842* .768 -0.896 0.833
Household Head (HH)
Husband -.514 .590 -0.865 .584 -1.065 .615
Wife -.698 .619 -0.751 .611 0.908 645
Sibling
Education Level (HH)
Regularly uneducated .192 .183 0.242 .177 0.112 .185
Regularly educated
Marital Status (HH)
Never married or Cohabiting -1.090* .374 0.647 .359 -0.544 .370
Married -.231 .311 0.185 .299 -0.184 .313
Divorced or Widowed .
Study Site (Woredas)
Kutaber -1.258* .467 -0.223 .447 1.117* .532
Kalu -.445 .443 0.058 .426 -0.910 .508
Tehuledere
Land Size Cultivated
Less than 0.5 Hectare .432* .161 -0.050 .154 0.349* .161
Above 0.5 Hectare
Land Fertility
Fertile -.492 .307 -0.086 .309 -0.582 .316
Medium fertile -.503* .210 -0.267 .200 0.066 .203
Less fertile
* indicates significant at 5% level of significance

95 | P a g e
© Yimam JA, UNISA 2019
Table 5.10 continued……………. (* indicates significant at 5% level of significance)
Availability Accessibility Utilisation
Variable Categories Estimates S.E Estimates S.E Estimates S.E
Cultivate time
Yearly .850 .432 -1.586* 0.421 1.542* 0.495
Biannual and more .
Shortage of Rainfall
Yes .586* .232 0.388 0.227 0.492* 0.241
No
Crop/vegetable Disease
Yes 1.650* .187 0.159 0.170 0.290 0.174
No
Market Price Increase
Yes -.578* .174 0.132 0.165 -0.010 0.170
No
Agro-ecology of study site
Cold (Dega) 1.622* .235 0.920* 0.340 -0.012 0.354
Medium (Wenadeg) .4308 .360 0.331 0.214 0.158 0.224
Hot (Kolla)

Kendall's tau Correlation Coefficient

The bivariate correlation coefficients between the successive time point’s food security statuses
were computed through the non-parametric correlations called “Kendall's tau_b” using the raw
discrete food security data and presented in Table 5.11.

The Kendall’s tau was computed from the pseudo data of food security statuses of the three-
phase using cumulative logit marginal distributions as presented in Table 5.12 and this result

96 | P a g e
© Yimam JA, UNISA 2019
further used for preliminary analysis to select appropriate copula families for the copula
cumulative logit model.

Table 5. 11: Nonparametric correlation coefficients of the household food security states
of the three data collection phases

Phase I Phase II Phase III

Phase I 1 .486** .214**

Phase II .486** 1 .253**

Phase III .214** .253** 1

** indicates significant at 0.01 level of significance.

Table 5. 12: The Kendall’s tau for the pseudo data computed from the cumulative logit marginal
distribution for the application of three-phase longitudinal data.

Phase I Phase II Phase III

Phase I 1.000 0.487 0.211
Phase II 0.487 1.000 .250
Phase III 0.211 .250 1.000

Based on the results in Tables 5.11 and 5.12, one can observe that similar and consistent results
were found, and it can be said that positive and significant correlation were observed between the
successive food security phases.

PCC Selection

The preliminary analysis computed in Table 5.12 showed that the Kendall’s tau of the pseudo
data computed from the cumulative logit model was positively correlated. The correlation
between successive time points of food security status was positive so that all bivariate copula
families listed in this thesis can be the candidate for measuring the dependence between
successive food security statuses of household. Algorithm II was used to select the best fit
bivariate copula for this application data. For the time being, for minimizing computational
challenges, t copula is not used. The result of pair copula selection process is presented in Table
97 | P a g e
© Yimam JA, UNISA 2019
5.13. Algorithm II was employed and their corresponding modified Akaike Information Criterion
(mAIC) were computed as presented in Table 5.13.

Table 5. 13; Summary of copula families by applying Algorithm II to the data and estimated
modified Akaike Information Criterion (mAIC) to select the best fit copula families.

Modified Akaike Information Criterion (mAIC)

Copula (𝑌1 , 𝑌2 ) (𝑌2 , 𝑌3 ) (𝑌1|𝑌2 , 𝑌3 |𝑌2 )
Gaussian 4011.59 4140.945 3720.277
Clayton 4098.848 4140.945 3721.848
Gumbel 3995.368 4132.143 3715.798
Frank 4009.531 4158.39 3720.474
AMH 4227.585 4180.113 3720.715
Independent 4406.431 4213.82 3733.431
Where; 𝑌1 = household food security status at the first 12 months considered as
baseline food security status
𝑌2 = household food security status at the middle six months

𝑌3 = household food security status at the last six months

The D-vine structure of the longitudinal household food security status in the first tree was (𝑌1 ,
𝑌2 ) and (𝑌2 , 𝑌3 ). Similarly, in the second tree the structure is (𝑌1|𝑌2 , 𝑌3 |𝑌2 ). From Table 5.13, one
can deduce that the best fitted copula families on the first tree were Clayton and Gumbel for edge
1 and 2 respectively while the second tree has AMH copula family since the mAIC is smaller
than the other copula families. Hence, Clayton, Gumbel and AMH copula families were selected
as best fit bivariate copula that serves for parameter estimation in the full maximum likelihood
parameter estimation for the corresponding vine structure they fitted best.

Estimation of the Copula and Marginal Parameters

The best fitted bivariate copulas were selected using Algorithm II as presented in Table 5.13.
Gumbel bivariate copula was selected to measure the dependence between first and second
household food security status, and second and third phase. Similarly, it was also selected for the

98 | P a g e
© Yimam JA, UNISA 2019
first and the third phase given that the second household food security status. The corresponding
marginal distribution for the application data was cumulative logit model.

The main purpose of this study was to address the dependence of household food security status
over time and their corresponding predictor factors at household level. MLE approach was used
to compute jointly the dependence parameters using the selected bivariate copula families and
the marginal parameters using cumulative logit model. Derivation of the likelihood and log-
likelihood function was made and presented in Appendix B II.

We wrote our own R code using “auglag” optimization R package to estimate the parameters of
the copula and the marginal distribution functions. Details of the R code were displayed in
Appendix B II. The estimated values of the parameters and the corresponding standard errors
both for the copula and marginal distributions were displayed in Table 5.14 and 5.15
respectively.

Copula Parameter

Table 5.14 summarizes the results of the estimated dependence parameters of the selected
bivariate copula families. The result shows positive dependence was observed between all phases
of household food security statuses. Gumbel copula measures the pairwise dependence between
all of the three phases of the individual household food security status. Moreover, these
dependences were observed as statistically significant. As a result, this leads to the conclusion
that individual household food security status varies with time. Therefore, the household food
security status in the study area is not stable over time.

Table 5. 14: The estimates of dependence parameters using the selected pair copula for
the application data of longitudinal household food security status.

Tree Copula Estimated Estimated SE Bijection tau

family Parameter
I Gumbel 1.2511 0.4452 0.2007034
Gumbel 1.1987 0.4625 0.1657629
II Gumbel 1.057081 0.4531 0.0539987

99 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 15: Summary results of the marginal parameters of pair copula based
longitudinal cumulative logit model for the household food security data.

𝒀𝟏 𝒀𝟐 𝒀𝟑
Variable Categories Estimates S.E Estimates S.E Estimates S.E
1|2 -2.55073 0.601 -2.1034 0.59882 -3.24248 0.6156
Intercept 2|3 0.166815 0.583 -0.06289 0.58211 -0.15301 0.57222
3|4 2.93809 0.599 3.29915 0.60151 2.42231 0.57948
Times of cultivate within a year
Yearly 0.018707 0.186 -0.5623* 0.1868 -0.01574 0.18434
Biannual and more
Crop disease
Yes 0.748778* 0.177 0.596795* 0.17873 0.731031* 0.17284
No
Increase in market price
Yes -0.33019* 0.169 -0.61814* 0.16923 -0.53993* 0.16799
No
Weathering condition of the village
Cold (Dega) 0.21021 0.212 -0.04505 0.16923 -0.4704 0.21033
Medium (Weinadega) 0.761084* 0.293 0.557457* 0.2133 -0.70407* 0.28844
Hot (Kolla)
Availability of rain
Little -0.39891 0.533 -0.40582 0.53077 0.446841 0.52183
Enough -0.53664 0.544 -0.73143 0.54211 0.960943 0.53419
High

Marginal Parameter

Five time-vary covariates were included in this study to assess the effect of these variables on the
household food security status over time. Summary statistics were computed in Table 5.15. The
result shows a statistically significant difference in the marginal parameter between the presence
and the absence of crop disease, the presence of market price increase and not, and hot
weathering condition and medium weathering condition in all time points of the household food
100 | P a g e
© Yimam JA, UNISA 2019
security status. Areas where crop disease happened are more likely to lead households to be
chronically, moderately and mildly food in-secured than areas not crop disease happened. In
addition, hot weather conditions are more likely to lead households to be chronically, moderately
and mildly food in-secured than medium weather conditions in all time points. Likewise,
increased market price is another factor to lead households chronically, moderately and mildly
food in-secured than stable market price in all phases household data.

Moreover, households cultivating once a year are more likely chronically, moderately and mildly
food in-secured than those cultivating two or more time a year in the second phase of household
data. In contrast, in this study, the only time-varying covariate that did not affect the household
food security status in all time points was availability of rainfall.

Effects of PCC Model on the Univariate Cumulative Logit

To compare the effects of PCC model on the cumulative logit model in fitting longitudinal
ordinal data, the finding of both the PCC model and univariate model in the cumulative version
were presented in Table 5.15 and 5.16 respectively. Both the PCC and univariate models
identified the presence of cultivation disease and cold agro-ecology as positive and significant
determinants for households to be food in-secured in the first round. Furthermore, the PCC
model identified the presence of market price as determinant of household food insecurity.
Likewise, both models identified cultivation once a year, crop disease and cold agro-ecology
determinants for households to be food in-secured in the second round. Moreover, the PCC
model identified the presence of market price as determinant of household food insecurity.

In the third round, the univariate model identified one more predictor for household food
insecurity than the PCC model. Crop disease, market price increase and hot and medium agro-
ecology were identified as predictors of household food insecurity through both the PCC and
univariate models. The PCC model drops the yearly once cultivation season while it was
significant in the univariate model.

101 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 16: Summary results of the marginal parameters of the univariate cumulative
logit model for the household food security data

𝒀𝟏 𝒀𝟐 𝒀𝟑
Variable Categories Estimates S.E Estimates S.E Estimates S.E
1|2 -3.328* 0.628 -1.594* 0.410 -4.677* 0.699
Intercept 2|3 -0.717 0.606 0.962* 0.394 -1.442* 0.653
3|4 2.122* 0.613 3.792* 0.426 1.098 0.652
Times of cultivate within a year
Yearly 0.284 0.193 -0.423* 0.178 0.789* 0.176
Biannual and more
Crop disease
Yes 1.033* 0.177 0.426* 0.190 0.989* 0.199
No
Increase in market price
Yes -0.257 0.169 -0.213 0.201 -0.830* 0.195
No
Agro-Ecology of Study Site
Cold (Dega) 0.901* 0.300 0.277 0.266 -0.620* 0.226
Medium (Weinadega) 0.088 0.208 0.433 0.222 -0.668* 0.289
Hot (Kolla)
Availability of rain
Little -0.657 0.551 1.448* 0.414 -0.150 0.613
Enough -0.537 0.544 0.573 0.306 0.310 0.601
High

In all of the significant determinants in the availability dimension, the PCC model
underestimated the parameters of the marginal model. In the accessibility dimension, 50% of the
predictors were overestimated and the rest of them were underestimated. Similarly, in the
utilisation dimension, some of the significant predictors were overestimated and some of them
were underestimated. The interpretation and discussion of the finding were made using the
finding of the PCC model.

102 | P a g e
© Yimam JA, UNISA 2019
 Kendall's tau Correlation Coefficient

The bivariate correlation coefficients between the successive food security dimensions for the
combined data were computed through the non-parametric correlations called “Kendall's tau_b”
using the raw discrete food security data and presented in Table 5.17.

The Kendall’s tau was also computed from the pseudo data of food security statuses of the three
dimensions using marginal model of the cumulative logit marginal distributions presented in
section 4.5 equation (4.77) as presented in Table 5.18. This result is further used for preliminary
analysis to select appropriate copula families for the copula marginal model via cumulative logit
model.

Table 5. 17: Nonparametric correlation coefficients of the household food security states
of the three combined dimensions

Availability Accessibility Utilisation

Availability 1 0.0234(0.24) 082** (0.000)
Accessibility 0.0234(0.24) 1 -.096** (0.000)
Utilisation 082** (0.000) -.096** (0.000) 1
** indicates significant at 0.01 level of significance.

Table 5. 18: The Kendall’s tau for the pseudo data computed from the marginal model of
the cumulative logit marginal distribution for the multivariate longitudinal data.

Availability Accessibility Utilisation

Availability 1.00000000 0.02336231 0.08231387
Accessibility 0.02336231 1.00000000 -0.09637995
Utilisation 0.08231387 -0.09637995 1.00000000

Based on the results in Tables 5.17 and 5.18, one can observe that similar and consistent results
were found. It can be said that negative and significant correlation was observed between food

103 | P a g e
© Yimam JA, UNISA 2019
access and utilisation. In contrast, positive but not statistically significant correlation was
observed between food access and availability.

PCC Selection

The preliminary analysis computed in Table 5.18 reduces the number of bivariate copula families
and the result showed that the Clayton and Gumbel copulas are not applicable for measuring the
dependence between accessibility and utilisation because their Kendall’s tau shows negative
dependence. Moreover, the Archimedean and AMH copulas are preferable for longitudinal data
to incorporate determinant factors in the sense that the dataset is large. Hence, we removed the
Gaussian and t copula from the lists of bivariate copulas. For those that satisfied the preliminary
analysis, Algorithm I was employed and their corresponding modified Akaike Information
Criterion (mAIC) were computed as presented in Table 5.19.

Table 5. 19: Summary of copula families using Algorithm I to the data and estimated
modified Akaike Information Criterion (mAIC) to select the best fit copula families.

Modified Akaike Information Criterion (mAIC)

Bivariate Copula (𝑌1 , 𝑌2 ) (𝑌2 , 𝑌3 ) (𝑌1|𝑌2 , 𝑌3 |𝑌2 )
Clayton 14115.8 NA NA
Gumbel 14121.48 NA NA
Frank 14117.95 13078.27 12667.74
AMH 14117.78 13077.34 12668.16
Independent 14119.07 13100.01 12663.28
where 𝑌1 = Availability, 𝑌2 = Accessibility and 𝑌3 = Utilisation

The D-vine structure of household food security in the first tree was (availability (𝑌1 ),
accessibility(𝑌2)) and (accessibility (𝑌2 ), utilisation (𝑌3 )). Similarly, in the second tree, the
structure is (𝑌1 |𝑌2 and 𝑌3 |𝑌2 ). The bivariate correlation between 𝑌1 |𝑌2 and 𝑌3 |𝑌2 was computed
as negative. Hence, the Clayton and Gumbel copulas are not still applicable for measuring the
dependence between them. From Table 5.19, one can deduce that the best fitted copula families
on the first tree were Clayton and AMH for edge 1 and 2 respectively while the second tree has
independent copula family since the mAIC is smaller than the other copula families. Hence,

104 | P a g e
© Yimam JA, UNISA 2019
Clayton, AMH and Independent copula families were selected as best fit bivariate copula that
serves for parameter estimation in the full maximum likelihood parameter estimation for the
corresponding vine structure they fitted best.

Estimation of the Copula and Marginal Parameters

The best fitted bivariate copulas were selected using Algorithm I as presented in Table 5.19.
Clayton bivariate copula was selected to measure the dependence of household food security
statuses between availability and accessibility dimensions. Similarly, to measure between
accessibility and utilisation, AMH was selected. Moreover, independent copula was for
availability and utilisation given that accessibility household food security status has already
appeared. The corresponding marginal distribution for the application data was marginal model
via cumulative logit model.

The likelihood function was computed using these selected bivariate copula families and the
marginal distribution. MLE approach was used to jointly compute the dependence parameters
using the selected bivariate copula families and the marginal parameters using marginal model of
the cumulative logit model. Derivation of the likelihood and log-likelihood function was
computed and presented in Appendix B III.

We wrote an R code using “auglag” optimization R package to estimate the parameters of the
copula and the marginal distribution functions. Detail of the R code is displayed in Appendix C
III. The estimated values of the parameters and the corresponding standard errors both for the
copula and marginal distributions were shown in Table 5.20 and 5.21 respectively.

Copula Parameter

Table 5.20 summarises the results of the estimated dependence parameters of the selected
bivariate copula families. Clayton copula measures the dependence between availability and
accessibility household food security status, AMH copula measures the accessibility and the
utilisation and independent the availability and the utilisation given that accessibility happened
and found that positive dependence were observed. Moreover, these dependences were observed
as statistically significant. This leads to the conclusion that household food security status
dependences to each other.

105 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 20: The estimates of dependence parameters using the selected pair copula for
the application data of multivariate longitudinal household food security data.

Tree Copula Estimated Estimated SE Bijection tau

family Parameter
I Clayton 1.5 0.6351 0.4285714
AMH 0.99999 0.3811 0.3333267
II Independent

Marginal Parameter

All statistically significant variables in each of the marginal model via the cumulative logit
model were incorporated in the model for each dimension that hopes to be helpful as additional
information with the existing knowledge in this area. Summary statistics of the final model is
displayed in Table 5.21. In Table 5.21, the blank space indicated that particular variable was not
statistically significant for that dimension in the marginal model and hence, that particular
variable was not included only for that particular dimension during the estimation of the final
joint model.

Among the variables incorporated in the model, the follow-up time point and total farmland size
in hectare were identified as the potential statistically significant variables for the household food
in-security status in all dimensions. In each dimension, household food security status varies
with time. Therefore, the household food security status in each dimension in the study area is
not stable over time. Small land size ploughed (<= 0.5 hectare) is more likely to lead households
to be chronically, moderately and mildly food in-secured than those ploughed greater than 0.5
hectare.

106 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 21: Summary results of pair copula based marginal model via cumulative logit
model parameters estimates for the household food security data.

Availability Accessibility Utilisation

Variable Categories Estimates S.E Estimates S.E Estimates S.E
1|2 -4.10884 0.244 -2.76063 .585 -5.02558 .626
Intercept 2|3 -1.06692 0.221 -0.60402 .583 -1.06226 .606
3|4 1.20519 0.221 1.56259 .583 1.33108 .606
Time 0.127028 0.01 0.043073 .009 0.101924 .009
Household Head (HH)
Husband -1.06369 .564 -1.095 .590
Wife -1.19779 .575 -1.41323 .598
Son/daughter
Marital Status (HH)
Never married/Cohabiting -0.33469 0.16 -0.42594 .199
Married -0.30143 0.146 0.03492 .189
Divorced/Widowed
Study Site (Woredas)
Kutaber -0.98267 0.131 0.24066 .124
Tehuledere -1.36491 0.133 -0.46358 .123
Kalu
Total farmland size in Hectare
<=0.5 Hectare 0.39055 0.092 0.276432 .087 0.362342 .091
> 0.5 Hectare
Types of Cereal Crops cultivated
One type 0.355914 0.129
two type 0.452018 0.116
Three and more types
Time of cultivate within a year
Yearly -0.71082 0.117
Biannual and more

107 | P a g e
© Yimam JA, UNISA 2019
 Table 5.21 Summary results of pair copula based marginal model via cumulative logit
model parameters estimates for the household food security data … continued
Availability Accessibility Utilisation
Variable Categories Estimates S.E Estimates S.E Estimates S.E
Types of livestock
One and less type 0.29618 .121
Two to Three types 0.118352 .104
Four or more types
Presence of Pests
Yes -0.17848 .092
No
Shortage of rainfall
Yes 0.33973 0.1 0.444659 .098
No
Crop Disease
Yes 1.00354 0.103
No
Increase in market price
Yes 0.47612 0.102 0.239897 .097 -0.148 .099
No
Use of Pesticides
Yes 0.838011 0.113
No
Agro-Ecology of Study Site
Hot (Kolla) -0.74647 .179
Medium (Weinadega) -0.59454 .125
Cold (Dega)

Similarly, among the fixed covariates, study site was identified as a significant influencing
variable for household food security status both in availability and accessibility dimensions.
Households cultivated less than three types of cereal crops and cultivating once a year were more
likely to be chronically, moderately and mildly food in-secured than those who cultivated more
108 | P a g e
© Yimam JA, UNISA 2019
than two types and two or more times per year in availability dimension respectively. Moreover,
households headed by women who have widowed or divorced marital status are more likely to
be chronically, moderately and mildly food in-secured than headed by son/daughter and never
married or cohabiting in utilisation dimensions.

On the other hand, among time varying covariates, the shortage of rainfall is identified as
significant variables that lead households to be chronically, moderately and mildly food in-
secured both in availability and utilisation dimensions in all aggregated time points. Areas used
pesticides, market price increase and crop disease happened are more likely to lead households to
be chronically, moderately and mildly food in-secured than areas not used pesticides and crop
disease happened in availability dimension in all time points. Similarly, hot (Kolla) or medium
(Weinadega) agro-ecology, less than two types of livestock and stability of market price are
more likely to lead households to be chronically, moderately and mildly food in-secured than
cold (Dega), more than three types of livestock and the absence of market price increase in
accessibility dimension in all aggregated time points.

Effects of PCC Model on the Univariate Marginal Cumulative Logit Model

The univariate and PCC population-average cumulative logit model identified almost equal
significant predictors for household food insecurity in the availability and accessibility
dimensions. However, in the utilisation dimension, the univariate model identified more
significant predictors for household food insecurity over the PCC model. For comparison
purpose, the univariate marginal model outputs both for fixed and time varying covariates were
displayed in Table 5.22. Similarly, the PCC population-average cumulative model outputs both
for fixed and time-varying covariates were presented as well in Tables 5.21.

109 | P a g e
© Yimam JA, UNISA 2019
 Table 5. 22: Summary results of the univariate marginal model via cumulative logit
model parameters estimates for the household food security data.
Availability Accessibility Utilisation
Variable Categories Estimates S.E Estimates S.E Estimates S.E
1|2 -3.923* 0.235 -2.827* .575 -5.392* .675
Intercept 2|3 -1.054* 0.209 -0.929 .572 -1.571* .656
3|4 1.182* 0.210 1.006 .571 0.853 .656
Time -0.063* 0.010 0.073 .009 0.039* .009
Household Head (HH)
Husband -1.095 .552 -1.228 .638
Wife -1.207* .564 -1.429* .645
Son/daughter
Marital Status (HH)
Never married/Cohabiting -0.412* 0.158 -0.539* .203
Married -0.337 * 0.144 -0.056 .193
Divorced/Widowed
Study Site (Woredas)
Kutaber -0.804* 0.139 0.268* .124
Kalu -1.070* 0.141 0.442* .123
Tehuledere
Total farmland size in Hectare
<=0.5 Hectare -0.414* 0.091 0.250* .088 -0.332 .091
> 0.5 Hectare
Types of Cereal Crops cultivated
One type 0.539* 0.123
two type 0.670* 0.110
Three and more types
Time of cultivate within a year
Yearly -0.209 0.110
Biannual and more
*significant at 5% level of significance

110 | P a g e
© Yimam JA, UNISA 2019
Table 5.22 Summary results of the univariate marginal model via cumulative logit model
parameters estimates for the household food security data …Continued
Availability Accessibility Utilisation
Variable Categories Estimates S.E Estimates S.E Estimates S.E
Types of livestock
One and less type 0.294* .121
Two to Three types 0.091 .104
Four or more types
Presence of Pests
Yes -0.341* .091
No
Shortage of rainfall
Yes 0.323* 0.10 0.362* .092
No
Disease of cultivation
Yes 1.060* 0.104
No
Increase in market price
Yes 0.651* 0.101 0.274 .097 -0.363* .099
No
Use of Pesticides
Yes 0.898* 0.114
No
Weathering condition of the cite
Hot (Kolla) -0.766* .177
Medium (Weinadega) -0.634* .125
Cold (Dega)
*significant at 5% level of significance

111 | P a g e
© Yimam JA, UNISA 2019
Based on the finding presented in Tables 5.21 and 5.22, among the fixed covariates, the PCC
population-average cumulative model identified a yearly once cultivation activity was the
positive determinant for households to be severe to mildly food in-secured compared with
counterparts while it was not significant in the univariate model in the availability dimension.
Conversely, the univariate model identified the marital status of the household head and types of
household head as a predictor for household food insecurity in availability and accessibility
dimensions respectively while the PCC model has dropped them out.

The PCC model underestimated the almost all significant predictors of household food insecurity
in the availability dimension. On the other hand, the PCC model overestimated in almost all
significant predictors of household food insecurity in the accessibility dimension. Similarly, the
PCC model in the utilisation dimension dropped out predictors like presences of pests and
market price increase while these were significant determinants for household food insecurity in
the univariate population-average cumulative model. The PCC model overestimated all of the
significant predictors in the utilisation dimensions. The interpretation and discussions for the
finding were intended on the findings of the PCC population-based cumulative logit model.

112 | P a g e
© Yimam JA, UNISA 2019
 Chapter Six

Discussions

In this study, pair copula construction (PCC) approach was implemented for analysing
multivariate, longitudinal and multivariate longitudinal ordinal data applied to household food
insecurity collected from selected Woredas of South Wollo Zone. The practical implementation
of PCC model for the three types of data was separately discussed in the subsequent sections.

In this section, we applied a pair copula construction based cumulative logit regression model to
jointly determine the dependence between the three food security dimensions and the respective
determinants (Olaomi and Yimam, 2019). Prior to the application, we conducted selection of
appropriate bivariate copula families that best fit to examine the dependence of food security
dimensions and estimate their corresponding parameter using the pseudo data. For this purpose,
the algorithm relevant for bivariate copula selection was developed. Among the candidate
bivariate copula families, the Frank copula selected as best fitted to express the dependence
between accessibility and utilisation, and availability | accessibility and utilisation | accessibility.
Moreover, AMH copula was the best fit for measuring the dependence of availability and
accessibility. Furthermore, the cumulative logit model was used as marginal distribution to
estimate the marginal parameters. Finally, full MLE method was implemented for jointly
estimate the dependence between the three dimensions and their respective determinants.

Overall, this model provided a very good description of the data and estimated all the
dependence parameter and marginal parameters as needed. The nice feature of this model in this
setting was that it allows for the estimation of the effect of the covariates both on the marginal
parameters and on the dependence of the outcomes over the other multivariate ordinal models.
This model depicted clearly the effects of the covariates on the dependence parameter of the
three dimensions. The Kendall’s tau for the pseudo data computed from the cumulative logit and
Bijection tau computed from the copula parameter showed large difference. The Bijection tau
was computed from the copula parameter, which was estimated incorporating the effect of
covariates in the model. The Bijection tau more relied with the literature in the food security

113 | P a g e
© Yimam JA, UNISA 2019
analysis. Furthermore, to assess the effects of PCC model on the cumulative logit model, the
univariate cumulative logit estimates were also fitted for each dimension. The PCC model
identified additional significant determinants of household food insecurity in all dimensions. On
the other hand, the PCC model overestimated majority determinants of household food insecurity
in availability and accessibility dimensions while underestimated in the utilisation dimension.

In PCC, the copula parameters captured the pair-wise non-normal relation between the food
security dimensions. Food availability, accessibility and utilisation have pair-wise positive
relationship. Moreover, the D-vine PCC determined the direction of the relationship as
availability contributes to accessibility, accessibility contributes to utilisation and given that
accessibility, availability contributes to utilisation. This finding is consistent with the framework
developed by FAO (FAO, 2008). The finding implies that the households being food secured in
availability dimension, the likelihood of food insecurity trap declines in food accessibility and
then in utilisation dimensions. Likewise, the households that are food secured in accessibility
dimension, the likelihood of food insecurity trap declines in food utilisation dimension.

Determinants of Availability: The findings of marginal parameters revealed that households

with higher agro-ecology (study site), less ploughed land, shortage of rainfall, cultivating once a
year, market price increases, hot agro-ecology and presence of disease on the cultivated land
were more likely to be chronically to mildly food in-secured. In contrast, households headed by
divorced or widowed marital status were less likely to be chronically to mildly food in-secured
(Olaomi and Yimam, 2019).

The study site with lower agro-ecology has positive effects on food in-security compared with
the study site with higher agro-ecology. This finding may relate with hot agro-ecology of the
study site because one of the factors that positively affect food insecurity was hot agro-ecology.
This finding is consistent with the result of meta-analysis conducted in Ethiopia by (Bashir and
Schilizzi, 2012). Similarly, a household with small land ownership is more likely to be food in-
secured and this finding consistent with the study conducted in Ethiopia by (Bashir and Schilizzi,
2012). Households’ agricultural activities have also a positive effect on the households’ food
insecurity status. Households cultivated agricultural produced once on the yearly base were more
likely to be food in-secured than those obtained twice or more per year.

114 | P a g e
© Yimam JA, UNISA 2019
Among climate change related factors, the availability of limited amount of annual rainfall has
positive effect on the household food insecurity. A similar finding was observed in a research
conducted in Ethiopia by (Abegaz, 2017). Moreover, recurrent disease that occurred on the
cultivated land was positively affects the household food insecurity status. Market price increase
is also another factor that positively influences the status of household food insecurity. This
finding is similar with findings in the meta-analysis conducted by (Bashir and Schilizzi, 2012) in
Ethiopia and a research conducted by (Ahmed et al., 2017) in Pakistan.

Moreover, among household characteristics household headed by never married or cohabiting

marital status were less likely to be food in-secured compared to households headed by divorced
or widowed.

Determinants of Accessibility: The findings of the marginal parameter of the pair copula based
cumulative logit model revealed that households headed by never married or cohabiting marital
status, small farmland size, shortage of rain fall, cultivating once a year, hot weathering
condition and presence of disease on the cultivated land contribute to making the households to
be chronically to mildly food in-secured.

Among demographic factors for household food insecurity status, households with never married
or cohabiting were more likely food in-secured compared with divorced or widowed households.
This finding is in contrast with finding in availability dimension of this study. On the other hand
this finding is similar with the finding pointed out by Mensah et al. (2013) in Sekyere-Afram
Plains District of Ghana. However, the finding by (Magaña-Lemus et al., 2016) in Mexico
household headed by single, widowed or divorced women were more likely to be food in-
secured compared with the married one.

Moreover, the amount of land ownership has effects on the household food security status.
Households with small farmland size were more likely to be food in-secured than who have large
farm size. The results of this study provides consistent result with the finding in some part of
Ethiopia by (Astemir, 2015), (Shone et al., 2017), (Feyisa, 2018) and (Moroda et al., 2018) and
in part of Ghana by Mensah et al. (2013). Similarly, agricultural activities have also a positive
effect on the households’ food insecurity status. Households those harvested once yearly were
more likely to be food in-secured than those harvested twice or more per year.

115 | P a g e
© Yimam JA, UNISA 2019
On the other hand, among climate change and environmental factors, the availability of limited
amount of annual rainfall has a positive effect on the household food insecurity and this result
resonates with finding in the research conducted by (Abafita and Kim, 2014) and (Abegaz, 2017)
in Ethiopia and in rural Zambia by (Wineman, 2016). Households living in hot agro-ecology
were more exposed to chronic to moderate food in-secured compared with households living in
cold agro-ecology. Moreover, recurrent disease occurred on the cultivated land was also
positively affects the household food insecurity status.

Determinants of Utilisation: The marginal estimates of the pair copula based cumulative logit
model of the utilisation dimensions revealed that household headed by sibling, higher agro-
ecology (study site), small farmland size, shortage of rainfall, cultivating once a year, and
presence of disease on the cultivated land were positively affect households to be chronically to
mildly food in-secured.

Among the demographic variables statistically significant factors for food insecurity status is the
household headed by siblings. Households headed by siblings were more likely to be food in-
secured than headed by husbands. Households living in higher agro-ecology environment were
more likely to be food insecurity than living in lower agro-ecology environment. This finding is
in contrast with the finding in the availability dimension of this study. Moreover, households that
cultivated large farmland size have the potential to be food secured than cultivated small
farmland size. This is because households have the potential to harvested different food groups
by their own other than purchase from the local market. This finding is in line with the finding
carried out by (Moroda et al., 2018) in Ethiopia. Households cultivating once per year were more
likely to be food in-secured than cultivated twice or more per year.

Among climate change and environmental variables, household obtained small amount of annual
rainfall were more likely to be food in-secured. This finding is consistent with the finding carried
out by (Moroda et al., 2018) in Ethiopia. Moreover, recurrent disease that occurred on the
cultivated land was also positively affects the household food insecurity status. This finding is
also harmony with the finds by (Abegaz, 2017).

116 | P a g e
© Yimam JA, UNISA 2019
Among the variables incorporated in the final model, four of them were obtained as common
determinant factors for all of the three food insecurity dimensions. These are small farmland size,
shortage of annual rainfall, cultivating once a year and presence of disease on the cultivated land.

In this section, we applied a pair copula construction based cumulative logit regression model to
jointly determine the stability of household food security over time and the respective
determinants (Olaomi and Yimam, 2019). Prior to the application, we conducted the selection of
appropriate bivariate copula families that best fits to examine the dependence of food security
statuses over time and estimate their corresponding parameters using the pseudo data. For this
purpose, the algorithm relevant for bivariate copula selection was developed. Among the
bivariate copula families, the Gumbel copula selected as best fitted to express the dependence
between the first and second, the second and the third time points. Moreover, the Gumbel copula
was the best fit for measuring the dependence of first and third phase food security status given
that the second phase has already happened. Furthermore, the cumulative logit model was used
as marginal distribution to estimate the marginal parameters. Finally, full MLE method was
implemented for jointly estimate the dependence between the three consecutive time point
household food security status and their respective determinants.

Overall, this model provided a very good description of the data and estimated all the
dependence parameter and marginal parameter as needed. The PCC model estimated the
dependence between food security status among the successive time points and the effects of the
time varying covariates on the dependence of food security statuses using the estimated bivariate
copula parameters and the effect of the time varying covariates on each food security status of
the successive time points using the estimated parameter of the cumulative logit model. The nice
feature of this model in this setting was that it allows estimation of the effect of the covariates
both on the marginal parameters and on the dependence of the outcomes over the other
longitudinal ordinal models. On the top of this, the model depicted also the recurrent covariates
that affect the household food security status over time the other longitudinal ordinal models did
not have. Furthermore, to assess the effects of PCC model on the cumulative logit model, the
univariate cumulative logit estimates was also fitted for each time points. The PCC model

117 | P a g e
© Yimam JA, UNISA 2019
identified additional significant determinants of household food insecurity in the first and second
rounds. The univariate model identified more significant predictors than the PCC model in third
rounds. Moreover, the PCC model underestimated majority of determinants of household food
insecurity in the first round. In the second and third rounds in some of the predictors, the PCC
model underestimated and in some of them overestimated.

The findings in the copula parameter showed that there were statistically significant differences
between the pair-wise dependence in all successive time points. The copula parameters showed
positive dependence between successive time points. In the longitudinal analysis, positive
correlation is expected between successive time points of the individual response. This model is
concordant with the expectation of longitudinal analysis. Again, in the longitudinal analysis,
strong correlation resulted in stability over time. Likewise, large dependence copula parameter
resulted in strong correlation and subsequently stability over time. Therefore, the current study
indicates that significant but small dependence parameter leads to instability over time. Hence,
the individual household food security status is not stable over time. This means that household
food security status varied from time-to-time.

The findings in the marginal parameter showed that presence of crop disease, market price
increase and medium (Weinadega) agro-ecology were significant and recurrent factors for
households to be chronically to mildly food in-secured. The findings of this study indicate that
availability of adequate amount of annual rainfall is crucial for household food security. This
finding is consistent with the study conducted in Ethiopia by (Abafita and Kim, 2014) and
(Mbolanyi et al., 2017) and in Uganda by (Mbolanyi et al., 2017). An increase of market price
positively influences the status of household food insecurity. This finding resonates with the
finding conducted in Pakistan by (Ahmed et al., 2017) and a meta-analysis in Ethiopia by (Bashir
and Schilizzi, 2012). Moreover, the fluctuation of agro-ecology over time affects the status of
household food security status. The finding is similar with finding conducted Ethiopia by
(Mbolanyi et al., 2017).

One-time cultivation per year is the significant covariate that leads households to be chronically
to mildly food in-secured in the second time points compared with those cultivated more than
one cultivation season. For the time being, in this study, the availability of rainfall was not
significant factor for household food security.

118 | P a g e
© Yimam JA, UNISA 2019
In this study, the pair copula construction approach was extended for multivariate longitudinal
ordinal data via the marginal model of the cumulative logit model. Marginal model is one of the
statistical models commonly used in the univariate longitudinal ordinal data analysis using the
cumulative logit version. Therefore, we proposed population-average based PCC model for
multivariate longitudinal ordinal data. This model has the potential to accommodate jointly the
dependence between multivariate ordinal outcomes, the covariate and follow-up time effects of
the ordinal outcomes both on the dependence measures and the marginal probabilities. However,
previous recent works on this area lacked to accommodate the aforementioned important
information jointly in a single model (Abegaz et al., 2015, Laffont et al., 2014). Our model filed
the population-average gap of the random effect models developed by (Laffont et al., 2014) and
the computational challenge of population-average multivariate t-copula models developed by
(Abegaz et al., 2015). Furthermore, the additional nice feature of this model is it allows
estimation of the effect of the covariates and follow-up time points both on the marginal
parameters and the dependency of the outcomes and allows estimation of the dependence
between multivariate ordinal outcomes.

Our model was applied in household food security data. The dependence between food security
status among the three dimensions and the effects of the covariates and time components on the
dependence of food security status of the three dimensions was computed using the estimated
bivariate copula parameters while the effect of the covariates and the follow-up time components
on each food security dimension status was computed using the estimated parameter of the
marginal model of the cumulative logit model.

Prior to the application, we conducted the selection of appropriate bivariate copula families that
fit best to examine the dependence of food security statuses over time in each of the three
dimensions and estimate their corresponding parameter using the pseudo data. For this purpose,
the Algorithm relevant for bivariate copula selection was developed. Among the bivariate copula
families, the Clayton copula selected as best to express the dependence between the availability
and accessibility, the AMH copula for measuring the dependence of accessibility and utilisation
food security status in all of the aggregated time points for the particular application data.

119 | P a g e
© Yimam JA, UNISA 2019
Moreover, the independent copula was the best fit for measuring the dependence of availability
and utilisation food security status given that the accessibility dimension has already conducted.

Based on the selected bivariate copula families, we fitted the pair copula multivariate
longitudinal cumulative logit model to our data. Moreover, we focused on estimation of the
dependence between food security dimensions using the copula parameter, the stability over time
using the significance of the time component and their predictor variables using the marginal
distribution parameters of the marginal model. Ten covariates for availability, seven for
accessibility and eight for utilisation were incorporated in this model, which are significant at 5
percent significance level in the marginal model. The full MLE method was employed to
estimate both the copula and marginal parameters simultaneously.

In line with the PCC estimation, the univariate population based cumulative logit was fitted for
each dimension to see the effect of the PCC model on the univariate one. Both the univariate and
PCC population-average cumulative logit model identified almost equal significant predictors for
household food insecurity in the availability and accessibility dimensions. However, in the
utilisation dimension, the univariate model identified more significant predictors for household
food insecurity over the PCC model. The PCC model underestimated almost all the significant
predictors of household food insecurity in the availability dimension. The PCC model
overestimated almost all of the significant predictors in the accessibility and utilisation
dimensions.

The findings of the copula parameter showed that the copula parameters captured the pair-wise
dependence between the food security dimensions. Food availability, accessibility and utilisation
have pair-wise positive dependence. Moreover, the D-vine PCC determined the direction of the
relationship as availability contributes to accessibility, accessibility contributes to utilisation and
given that accessibility, availability contributes to utilisation. This finding is consistent with the
framework developed by FAO (FAO, 2008). The finding implies that the households being food
secured in availability, the likelihood of food insecurity trap declines in food accessibility and
then in utilisation. Likewise, the households being food secured in accessibility, the likelihood of
food insecurity trap declines in food utilisation. Hence, food security dimension specific
intervention might reduce the likelihood of food insecurity at household level.

120 | P a g e
© Yimam JA, UNISA 2019
Determinants of availability: The marginal estimate of the population-average cumulative
model in the pair copula-based model carried out so far produced different significant associated
determinant factors for household food security status in the availability dimension. The finding
of this study revealed that the household food security status changes over time. Likewise,
population-average household food security status is not stable over time. Similarly, the finding
showed that lower agro-ecology, shortage of rainfall, the presence of cultivation disease,
increased market price, use of pesticides, cultivating smaller types of cereal crops and cultivating
once per year positively affect the household food in-security (Olaomi and Yimam, 2019).

Households living in lower agro-ecology are more likely to be chronically to mildly food in-
secured compared with living in higher agro-ecology. In contrast, the findings were observed in
the research conducted by (Motbainor et al., 2016) in Ethiopia. Moreover, households who
ploughed smaller farmland size (less or equal to half hectare) were more likely to be chronically
to mildly food in-secured than ploughed greater than half hectare. (Bashir and Schilizzi, 2012)
identified similar findings in their meta-analysis in Ethiopia. Harvesting different types of cereal
crops also affects the status of household food security. Households harvesting fewer types of
cereal crops (less than three types) were more likely chronically to mildly food in-secured than
harvesting three or more types. Moreover, households living in a village that has one period of
cultivation season were more suspected to be food in-secured than cultivating two or more times
in one-year period.

Among the time varying covariates, during the three consecutive follow-up interviews, the
presence of small amount of annual rainfall positively affects household food insecurity status.
This finding resonates with that of the studies conducted by (Wineman, 2016) in rural Zambia
and in Ethiopia by (Abafita and Kim, 2014) and (Abegaz, 2017). Similarly, the instability of
market price positively affects the household food insecurity status and the finding is similar
with finding in Pakistan carried out by (Ahmed et al., 2017). Moreover, households affected by
cultivation disease and used pesticides were more likely to be food in-secured compared with
those not affected by cultivation disease and not used pesticide. This is owing to either the
disease destroying much number of products or the households may have invested much amount
of money for pesticides. This automatically leads the households to be food in-secured.

121 | P a g e
© Yimam JA, UNISA 2019
Determinants of accessibility: The marginal estimate of the population-average cumulative
model in the pair copula-based model carried out so far produced different significant associated
determinant factors for household food security status in the accessibility dimension. The finding
of this study revealed that the household food security status changes over time. Likewise,
population-average household food security status is not stable over time. Similarly, the finding
showed that lower agro-ecology, increased market price, herbing small number of livestock, hot
agro-ecology and small farmland size positively affected the household food in-security (Olaomi
and Yimam, 2019).

Like in the availability dimensions, households living in lower agro-ecology are more likely to
be chronically to mildly food in-secured compared with those living in medium agro-ecology.
This is may be owing to hot agro-ecology of the study site because cold agro-ecology is one of
the factors that positively affect household food insecurity status. In contrast, the findings were
observed in the research conducted by (Motbainor et al., 2016) in Ethiopia. Moreover,
households which ploughed smaller farmland size (less or equal to half hectare) were more likely
to be chronically to mildly food in-secured than ploughed greater than half hectare. (Bashir and
Schilizzi, 2012) and (Shone et al., 2017) in Ethiopia and (Mensah et al., 2013) in Pakistan
identified similar findings. Households herding fewer types of livestock (less than two types)
were more likely to be chronically to mildly food in-secured than herding two or more types.
(Motbainor et al., 2016) and (Habyarimana, 2015) also discovered similar findings in Ethiopia
and in Rwanda respectively.

Among the time varying covariates, during the three consecutive follow-up interviews, the
presence of small amount of annual rainfall positively affects household food insecurity status.
Similar findings were observed in Ethiopia conducted by (Abafita and Kim, 2014), (Abegaz,
2017) and (Agidew and Singh, 2018) and also in rural Zambia by (Wineman, 2016). Similarly,
the instability of market price positively affects the household food insecurity status and the
finding is similar with finding of a study conducted in Pakistan carried out by (Ahmed et al.,
2017).

Determinants of utilisation: Like in availability and accessibility, the marginal estimate of the
population-average cumulative model in the pair copula-based model carried out so far produced
different significant associated determinant factors for household food security status in the

122 | P a g e
© Yimam JA, UNISA 2019
utilisation dimension. The findings of this study revealed that the household food security status
changes over time. Likewise, population-average household food security status is not stable
over time. Similarly, the results showed that households headed by a woman, divorced/widowed
marital status of the household head, shortage of rainfall and small farmland size positively
affected the household food in-security.

The findings suggested that households headed by women positively affect the utilisation food
insecurity status of households. Moreover, divorced/widowed households were more likely food
in-secured compared to single or cohabitation household. On the other hand, households which
ploughed smaller farmland size (less or equal to half hectare) were more likely to be chronically
to mildly food in-secured than those that ploughed greater than half hectare. The finding is
similar to (Moroda et al., 2018) study conducted in Ethiopia is similar with the current study.

Among the time varying covariates, during the three consecutive follow-up interviews, the
presence of small amount of annual rainfall positively affects household food insecurity status.
Similar findings were observed in Ethiopia conducted by (Abafita and Kim, 2014) and (Abegaz,
2017) and also in rural Zambia by (Wineman, 2016).

This study revealed the determinants for all of the three dimensions and tried to compare with the
work of the others. However, owing to lack of available literature in the availability and
utilisation dimensions for some of the determinants the current study limited to compare with
other works.

Household food security dimensions are correlated to each other. This implies that for the
households being food secured in availability, the likelihood of food insecurity trap declines in
food accessibility and then in utilisation. Likewise, the households being food secured in
accessibility, the likelihood of food insecurity trap declines in food utilisation. This type of
modelling assists the food aid agents, planners or policy makers that in which dimension a
household is highly affected and which dimension let household’s food in-secured. This implies
food security dimension specific intervention might reduce the likelihood of food insecurity at
household level.

123 | P a g e
© Yimam JA, UNISA 2019
Hot agro-ecology areas were highly vulnerable to food insecurity. This implies that agro-ecology
or area specific intervention could alleviate the risk of food in-secured in the rural households.
This suggests that adaption strategies like harvest in hot agro-ecology. In response to this
strategy, farmers can use seeds that are resistant to short rain fall season and conserves water in a
hot area to increase the productivity.

Small cultivable land size increases the likelihood of households being food in-secured. This
implies that households with large land ownership could produce more food or may generate
income from it to purchase food for consumption. Hence, alternative income generating
mechanism should be set for rural households to reduce the pressure of cultivable land in
addition to encourage maximum yield from a given holdings through investing in land
improvements and soil conservation. Alternatively, to increase cultivable land size, a strong
policy intervention may be needed to relocate the population where settlement is densely
populated to the sites where it is not. This enables, at least some group, to share from land
holdings and any entitlements to resources which can lead to ensuring food security.

The number of cultivation season increases on the yearly base, the likelihood of households
being food in-secured trap declines. This implies that households cultivating two or three times
per year could produce more food or getting more income to purchase food for consumption than
cultivation once per year. Hence, ways should be sought through promoting irrigation activity in
order to increase cultivation season.

Households which experienced lower rainfall level were more likely to remain food in-secured.
The majority of the Ethiopian rural households are rain-dependent for their agricultural
production, resulting in persistent food insecurity. Hence, careful promotion of investment in
infrastructure to support irrigation and water resources development is one aspect worth
considering. On the other hand, climatic adaption strategies should be insight like selecting
appropriate crop varieties that can be planted in low amount of rainfall.

Crop disease positively affects the household food insecurity status. Hence, the development
agents (DAs) working in the area should provide immediate response either by providing
medicines or consulting the community for alleviating the problem in the early stage. The

124 | P a g e
© Yimam JA, UNISA 2019
government should also advance the agricultural strategies in a way that provides immediate
responses for this and other related agricultural problems.

The price of food increases, purchasing power goes down, dietary quality and total energy intake
are reduced, the likelihood of household being food in-secured increased. Hence, strong market
price policy to make stable the market price as well as increase the supply of food for
consumption is worth considering.

More livestock was kept by households that were more food secured. As the households’
livestock possessions were increased, their food security status would inevitably also respond
positive. More importantly, livestock possession enables the households to be food secure either
through the income earned or by direct consumption. This implies that the availability of greater
number of livestock permit households enhances their economic wellbeing in general and their
food entitlement in particular. Hence, careful promotion of investment in livestock project
supported by scientific methods is one aspect worth considering.

125 | P a g e
© Yimam JA, UNISA 2019
 Chapter Seven

Conclusions, Recommendations and Future Works

The pair copula based multivariate ordinal model with the cumulative logit version successfully
captured the non-normal relationship between ordinal outcomes and their respective
determinants simultaneously. Allowing the estimation of the effect of the covariates both on the
marginal parameters and on the non-normal correlation of the ordinal outcomes strengthens the
estimation performance of this model over previous multivariate ordinal models. The copula
parameters in the food security data revealed that pair-wise positive dependence was observed
between food availability, accessibility and utilisation dimensions. The marginal parameters of
this model depicted that small cultivable land, shortage of rainfall, cultivating once a year and
presence crop disease were positively influences household food insecurity in all the three
dimensions. Moreover, lower agro-ecology and market price increase positively affects
household food insecurity in availability dimension. Similarly, hot agro-ecology positively
affects household food security in accessibility. Moreover, lower agro-ecology positively affects
household food security in utilisation.

A pair copula based longitudinal ordinal model with cumulative logit version jointly estimated
the stability of ordinal outcome over time and the respective determinants. The result of the
copula parameter in this model pointed out the food security status at household level is not
stable over time. Estimating determinants for each longitudinal ordinal outcome broadens the
inclusion of recurrent determinants over the longitudinal periods does the previous longitudinal
models lacked. The marginal parameters of this model revealed that the presence of crop disease,
market price increase, and medium (Weinadega) agro-ecology were significant and recurrent
factors for households’ food insecurity over the three time periods.

Population-average based pair copula multivariate longitudinal ordinal model with cumulative
logit version jointly estimated the dependence between multivariate ordinal outcomes, and the
covariate and follow-up time effects of the ordinal outcomes both on the dependence measures
and the marginal probabilities using the full MLE method. This model successfully reduced the
population-average lack of the random effects model and the computational challenge of the

126 | P a g e
© Yimam JA, UNISA 2019
multivariate copula models of the multivariate longitudinal ordinal data analysis. This study
provides a good measure of dependence between food security dimensions using the copula
parameter and also the stability over time and the determinants of household food using the
marginal model parameters for all dimensions simultaneously. The findings of the copula
parameter showed that positive and statistically significant dependence were observed between
availability and accessibility and accessibility and utilisation. The marginal model of the
cumulative logit model was used to measure the parameters of the marginal distributions. The
findings of the model reveal that household food security was unstable over time for each
dimension. Small land size and shortage of rainfall were the common predictors of household
food insecurity in all dimensions. Moreover, lower agro-ecology and instability of market price
were the common predictors of household food insecurity in availability and accessibility
dimensions.

Food security dimensions depended to each other. The rap of one dimension affects the other
dimensions. Therefore, it is critically important to consider the common factors to provide
immediate intervention for severely food in-secured households. Moreover, great attention also
required to lookup which dimension is leading households to food in-secured.

Households’ food security status either in the individual food security dimensions or in the
composite food security is not stable over time. So great attention is required for granting
households to be food secured taking valuable intervention for the identified recurrent
determinants as well as other climate change and environment factors. Moreover, climatic
adaption strategies should be insight like selecting appropriate crop varieties that can be resistant
to short rain fall season and conserves water in a hot area to increase the productivity.

Likewise, systematic investment in infrastructure to support irrigation and water resources

development to increase cultivation season and alternative income generating mechanism to
reduce the pressure of cultivable land from a given holdings through investing in land
improvements and soil conservation are critical viewpoints to reduce the likelihood of food
insecurity at household level.

127 | P a g e
© Yimam JA, UNISA 2019
The government should design strong market price policy to make stable the market price as well
as increase the supply of food for consumption; careful promotion of investment in livestock
project supported by scientific methods; and advance the agricultural strategies in a way that
provides immediate responses for crop disease.

The pair copula based multivariate and longitudinal ordinal model provided easily interpretable
and understandable outputs. Therefore, we suggest the model for any multivariate and
longitudinal discrete data analysis.

The population-average based pair copula multivariate longitudinal ordinal model addressed all
of the food security dimensions simultaneously and the model found computationally effective
for not large set of data. Therefore, we suggest this model to apply for other application areas for
not extremely large number of outcomes and covariates.

In this thesis, the applied pair copula model for multivariate longitudinal ordinal data used a
three-stage sampling procedure to get the application data. In three-stage sampling, each random
selection may introduce a random effect. It is due to the contributions of the different stages to
the variance of an estimator. However, the current model didn’t take into account this during
parameter estimation.

The second limitation is, in the notion of the concept of food security we used three round data
collection to address the stability of the other three dimensions over time. Three rounds of data
may not be providing a realistic estimator to oversee the entire stability. It would have been
great if measurements were obtained for three to five seasons. However, we believe that some
kind of longitudinal data is better than cross-sectional data to study the household food security
situation.

The study tried to address all the four food security dimensions. Due to limited data at the
national level to address all the dimensions, the study forced to conduct primary data from
selected Woredas of one region of Amhara, Ethiopia. Hence the study did not represent the food
security situation of the Amhara region.

128 | P a g e
© Yimam JA, UNISA 2019
The pair copula-based regression models have applied throughout this thesis allowed specifying
the effect of covariates of the marginal distributions on the dependence and marginal structure.
We proposed a population-average based pair copula construction models for multivariate
longitudinal ordinal outcomes using marginal model of the cumulative logit marginal
distribution. Although we feel that our contribution is a major step forward in the modeling of
multivariate longitudinal ordinal outcomes via pair copula construction, we discuss three
important open questions in more detail.

The first open question is in the multivariate longitudinal ordinal outcomes, we have M ordinal
outcomes repeatedly measured T times. During our model development we re-ordered the
observations of the multivariate series into the univariate outcomes of dimensions 𝑁 = 𝑇∗𝑀
given by 𝑌 = (𝑌1 , 𝑌2 , … , 𝑌𝑀 ), where 𝑌1 = (𝑦11 , 𝑦21 , … , 𝑦𝑁1 )′ , 𝑌2 = (𝑦12 , 𝑦22 , … , 𝑦𝑁2 )′ 𝑎𝑛𝑑 𝑌𝑀 =
(𝑦1𝑚 , 𝑦2𝑚 , . , 𝑦𝑁𝑚 )′ . As a result, the joint probability mass function 𝑃𝑟(𝑌1 = 𝑦1 , 𝑌2 =
𝑦2 , … , 𝑌𝑚 = 𝑦𝑚 ) is decomposed as follows:

Pr(𝑌1 , 𝑌2 , … , 𝑌𝑚 ) = Pr(𝑌1 = 𝑦1|𝑌2 = 𝑦2 , … , 𝑌𝑚 = 𝑦𝑚 ) × Pr(𝑌2 = 𝑦2 |𝑌3 = 𝑦3 , … , 𝑌𝑚 = 𝑦𝑚 ) ×

… × Pr(𝑌𝑚 = 𝑦𝑚 ).

The parameters of the joint distribution of the entire dependence were estimated by a D-vine
copula of dimension T ∗ M. This model reduced the multivariate longitudinal dimensions into a
univariate longitudinal series based on the time point for each multivariate ordinal outcome. This
model loses the dependence between successive time points and the effect of covariates on the
dependence measure of the successive time point outcomes. This is not the issue of population-
average longitudinal data analysis but for other applications to consider the individual change
over time one can extend this model for multivariate longitudinal data (T repeated observations
of M dimensional vectors for the sample of n subjects) by using a different D-vine copula
approach as follows.

𝒀 = (𝒀𝟏 , 𝒀𝟐 , … , 𝒀𝑴 ), where 𝒀𝟏 = (𝑌1 , 𝑌2 , … , 𝑌𝑇 ), 𝒀𝟐 = (𝑌1 , 𝑌2 , … , 𝑌𝑀 ), … , 𝒀𝑴 = (𝑌1 , 𝑌2 , … , 𝑌𝑇 ),

𝑌1 = (𝑦11 , 𝑦21 , … , 𝑦𝑡1 )′ , 𝑌2 = (𝑦12 , 𝑦22 , … , 𝑦𝑡2 )′ 𝑎𝑛𝑑 𝑌𝑀 = (𝑦1𝑚 , 𝑦2𝑚 , . , 𝑦𝑡𝑚 )′ .

129 | P a g e
© Yimam JA, UNISA 2019
Hence, the joint probability mass function 𝑃𝑟(𝑌1 = 𝑦1 , 𝑌2 = 𝑦2 , … , 𝑌𝑚 = 𝑦𝑚 ) is decomposed as
follows.

𝑷𝒓(𝒀𝟏 = 𝒚𝟏 , 𝒀𝟐 = 𝒚𝟐 , … , 𝒀𝑴 = 𝒚𝒎 ) = 𝑷𝒓(𝒀𝟏 = 𝒚𝟏 | 𝒀𝟐 = 𝒚𝟐 , … , 𝒀𝑴 = 𝒚𝒎 ) × 𝑷𝒓(𝒀𝟐 =

𝒚𝟐 | 𝒀𝑴 = 𝒚𝒎 ) × … × 𝑷𝒓(𝒀𝑴 = 𝒚𝒎 ).

Taking the above expression into account one can apply the usual pair copula construction. This
extension can consider three different levels of analysis. At first, a pair copula describes the
relations of the responses observed at a specific time. Second each longitudinal series,
corresponding to a given response over time, is modeled separately using a pair copula
decomposition to relate the distributions of the variables describing the observation given in
different times. Finally, the marginal distribution relates the associated factors for each responses
and longitudinal time components.

The second open question is extending our model in in a Bayesian framework. Since the
Bayesian approaches have many advantages in modeling multivariate as well as multivariate
longitudinal outcomes, we believe that there may be significant advantages to estimating our
model in a Bayesian framework. Furthermore, the modular nature of the MCMC in the Gibbs
sampler may facilitate the development of more advanced multivariate longitudinal models.
However, the Bayesian approach requires a good proposal of prior information or distribution for
the marginal distribution during the construction of PCC. This implies that the selection of prior
distribution in the ordinal outcomes setting requires intensive work. As a result, this thesis is
concerned itself on the implementation of the PCC through the Frequentist paradigm and the
likelihood for our PCC is fast to compute. However, in the presence of proper prior distribution,
we believe that joint estimation of marginal and copula parameters for multivariate longitudinal
ordinal model could be easier to develop in a Bayesian context.

The third open question of this thesis is considering non-ignorable missing values in the analysis
of multivariate longitudinal outcomes via PCC. In the univariate longitudinal context, Cui et al.,
(2016) implemented the Peter and Clark (PC) algorithm for both the discrete and continuous data
assumed to be drawn from a Gaussian copula models. Furthermore, Cui et al., (2019) extended
the Gaussian copula models to Copula PC algorithm for incomplete data for mixed data with
missing values. Likewise, Gomes et al. (2019) further extended Gaussian copula for non-

130 | P a g e
© Yimam JA, UNISA 2019
Gaussian responses that are missing not at random using copula selection models. This implies
that in the univariate case, intensive works have been conducted in the copula context for
discrete data. However, in the multivariate longitudinal discrete context, since PCC models in
multivariate longitudinal data analysis are in the growing stage, it needs further research to
handle missing-ness. In our PCC model we did not consider handling missing data since we were
considering compensation of none-response rate and the missing-ness in our data were less than
the none-response rate considered during sample size determination. However, we believe that
considering PCC models in the multivariate longitudinal ordinal outcomes that can treat non-
ignorable missing data could provide valid estimates both in the marginal and copula estimates
of the final model extending one of the above methods that was implemented in the univariate
case.

131 | P a g e
© Yimam JA, UNISA 2019
 References

AAS, K., CZADO, C., FRIGESSI, A. & BAKKEN, H. (2009) Pair-copula constructions of
multiple dependence. Insurance: Mathematics and economics, 44, 182-198.
ABAFITA, J. & KIM, K.-R. (2014) Determinants of household food security in rural Ethiopia:
An empirical analysis. Journal of Rural Development/Nongchon-Gyeongje, 37, 129.
ABDU, J., KAHSSAY, M. & GEBREMEDHIN, M. (2018) Household Food Insecurity,
Underweight Status, and Associated Characteristics among Women of Reproductive Age
Group in Assayita District, Afar Regional State, Ethiopia. Journal of environmental and
public health, 2018.
ABEGAZ, F., NOORAEE, N., VAN DEN HEUVEL, E. & WIT, E. (2015) Logistic Regression
for Multivariate Longitudinal Ordinal Data. University of Groningen.
ABEGAZ, K. H. (2017) Determinants of food security: evidence from Ethiopian Rural
Household Survey (ERHS) using pooled cross-sectional study. Agriculture & Food
Security, 6, 70.
AGIDEW, A.-M. A. & SINGH, K. (2018) Determinants of food insecurity in the rural farm
households in South Wollo Zone of Ethiopia: the case of the Teleyayen sub-watershed.
Agricultural and Food Economics, 6, 10.
AGRESTI, A. (2010) Analysis of ordinal categorical data, John Wiley & Sons.
AHMED, U. I., YING, L., BASHIR, M. K., ABID, M. & ZULFIQAR, F. (2017) Status and
determinants of small farming households' food security and role of market access in
enhancing food security in rural Pakistan. PloS one, 12, e0185466.
ALI, M. M., MIKHAIL, N. & HAQ, M. S. (1978) A class of bivariate distributions including the
bivariate logistic. Journal of multivariate analysis, 8, 405-412.
ASAR, Ö. & İLK, Ö. (2014) Flexible multivariate marginal models for analyzing multivariate
longitudinal data, with applications in R. Computer methods and programs in
biomedicine, 115, 135-146.
ASMAMAW, T., BUDUSA, M. & TESHAGER, M. (2015) Analysis of Vulnerability to Food
Insecurity in the Case of Sayint District, Ethiopia. Asian Journal of rural development, 5,
1-11.

132 | P a g e
© Yimam JA, UNISA 2019
ASPELUND, T. (2002) Non-linear association-marginal models for multivariate categorical
data with application to ordinal receiver operating characteristic analysis, UMI
Dissertation Services.
ASSEFA, B. T. (2015) Food price changes and food security in developing countries: a
multidimensional analysis, Sl: sn.
ASTEMIR, A. (2015) Determinants of food security in rural farm households in Ethiopia. The
Hague, Netherlands.
BALLARD, T., COATES, J., SWINDALE, A. & DEITCHLER, M. (2011) Household hunger
scale: indicator definition and measurement guide. Washington, DC: Food and Nutrition
Technical Assistance II Project, AED.
BANDYOPADHYAY, S., GANGULI, B. & CHATTERJEE, A. (2011) A review of multivariate
longitudinal data analysis. Statistical methods in medical research, 20, 299-330.
BASHIR, M. K. & SCHILIZZI, S. (2012) Measuring food security: Definitional sensitivity and
implications. Contributed Paper Prepared for Presentation at the 56th AARES Annual
Conference, Fremantle, Western Australia, Paper.
BEDFORD, T. & COOKE, R. M. (2001) Probability density decomposition for conditionally
dependent random variables modeled by vines. Annals of Mathematics and Artificial
intelligence, 32, 245-268.
BEDFORD, T. & COOKE, R. M. (2002) Vines: A new graphical model for dependent random
variables. Annals of Statistics, 1031-1068.
BHATNAGAR, S., ATHERTON, J. & BENEDETTI, A. (2015) Comparing alternating logistic
regressions to other approaches to modelling correlated binary data. Journal of Statistical
Computation and Simulation, 85, 2059-2071.
BIESALSKI, H. K., DREWNOWSKI, A., DWYER, J. T., STRAIN, J., WEBER, P. &
EGGERSDORFER, M. (2017) Sustainable nutrition in a changing world, Springer.
BILINSKY, P. & SWINDALE, A. (2007) Months of adequate household food provisioning
(MAHFP) for measurement of household food access: indicator guide, Food and
Nutritional Technical Assistance Project, Academy for Educational ….
BIRHANE, T., SHIFERAW, S., HAGOS, S. & MOHINDRA, K. S. (2014) Urban food
insecurity in the context of high food prices: a community based cross sectional study in
Addis Ababa, Ethiopia. BMC public health, 14, 680.

133 | P a g e
© Yimam JA, UNISA 2019
CAGNONE, S., MOUSTAKI, I. & VASDEKIS, V. (2009) Latent variable models for
multivariate longitudinal ordinal responses. British journal of mathematical and
statistical psychology, 62, 401-415.
CAPALDO, J., KARFAKIS, P., KNOWLES, M. & SMULDERS, M. (2010) A Model of
Vulnerability to Food Insecurity.
CARLETTO, C., ZEZZA, A. & BANERJEE, R. (2013) Towards better measurement of
household food security: Harmonizing indicators and the role of household surveys.
Global food security, 2, 30-40.
CASTRO, A. P. (2000) Food Security and Resource Access: A Final Report on the Community
Assessments in South Wello and Oromiya Zones of Amhara Region, Ethiopia.
CHAGANTY, N. R. & NAIK, D. N. (2002) Analysis of multivariate longitudinal data using
quasi-least squares. Journal of Statistical Planning and Inference, 103, 421-436.
CHOI, J. (2012) Prediction in the joint modeling of mixed types of multivariate longitudinal
outcomes and a time-to-event outcome. University of Pittsburgh.
CLAYTON, D. G. (1978) A model for association in bivariate life tables and its application in
epidemiological studies of familial tendency in chronic disease incidence. Biometrika, 65,
141-151.
COATES, J. (2013) Build it back better: Deconstructing food security for improved
measurement and action. Global Food Security, 2, 188-194.
COCHRAN, W. G. (2007) Sampling techniques, John Wiley & Sons.
COHEN, M. J. & GARRETT, J. L. (2010) The food price crisis and urban food (in) security.
Environment and Urbanization, 22, 467-482.
CUI, R., GROOT, P. & HESKES, T. (2016) Copula PC algorithm for causal discovery from
mixed data. Joint European Conference on Machine Learning and Knowledge Discovery
in Databases. Springer.
CUI, R., GROOT, P. & HESKES, T. (2019) Learning causal structure from mixed data with
missing values using Gaussian copula models. Statistics and Computing, 29, 311-333.
CZADO, C. (2010) Pair-copula constructions of multivariate copulas. Copula theory and its
applications. Springer.
CZADO, C., JESKE, S. & HOFMANN, M. (2013) Selection strategies for regular vine copulae.
Journal de la Société Française de Statistique, 154, 174-191.

134 | P a g e
© Yimam JA, UNISA 2019
DARDANONI, V. & FORCINA, A. (2008) Multivariate ordered regression. mimeo.
DE BRUIN, E. & GRESSE, A. (2018) Dietary diversity amongst adults who buy at shopping
malls in the Nelson Mandela Bay area. Journal of Consumer Sciences.
DISSMANN, J. (2010) Statistical inference for regular vines and application. Diploma the-sis,
Technische Universitat Miinchen.
DISSMANN, J., BRECHMANN, E. C., CZADO, C. & KUROWICKA, D. (2013) Selecting and
estimating regular vine copulae and application to financial returns. Computational
Statistics & Data Analysis, 59, 52-69.
ENDALE, W., MENGESHA, Z. B., ATINAFU, A. & ADANE, A. A. (2014) Food Insecurity in
Farta District, Northwest Ethiopia: a community based cross–sectional study. BMC
research notes, 7, 130.
ENDALEW, B., MUCHE, M. & TADESSE, S. (2015) Assessment of food security situation in
Ethiopia. World Journal of Dairy & Food Sciences, 10, 37-43.
ETANA, D. & TOLOSSA, D. (2017) Unemployment and food insecurity in urban Ethiopia.
African Development Review, 29, 56-68.
FABER, M., SCHWABE, C. & DRIMIE, S. (2009) Dietary diversity in relation to other
household food security indicators. International Journal of Food Safety, Nutrition and
Public Health, 2, 1-15.
FAN, C.-L. (2009) Contributions to the theory of copula, WASHINGTON UNIVERSITY IN
ST. LOUIS.
FANG, H.-B., FANG, K.-T. & KOTZ, S. (2002) The meta-elliptical distributions with given
marginals. Journal of Multivariate Analysis, 82, 1-16.
FAO, A. (2008) Introduction to the Basic Concepts of Food Security. Food Security Information
for Action, Rome.
FAO. (2014) State of Food Insecurity in the World 2013: The Multiple Dimensions of Food
Security, FAO.
FEYISA, M. N. (2018) Determinants of food insecurity among rural households of South
Western Ethiopia. Journal of Development and Agricultural Economics, 10, 404-412.
FISHER, N. I. (1997) Copulas. Encyclopedia of statistical sciences.
FITZMAURICE, G., DAVIDIAN, M., VERBEKE, G. & MOLENBERGHS, G. (2009)
Longitudinal Data Analysis. International Statistical Review, 77, 147-165.

135 | P a g e
© Yimam JA, UNISA 2019
FITZMAURICE, G. M. & LAIRD, N. M. (1993) A likelihood-based method for analysing
longitudinal binary responses. Biometrika, 80, 141-151.
FRAHM, G., JUNKER, M. & SZIMAYER, A. (2003) Elliptical copulas: applicability and
limitations. Statistics & Probability Letters, 63, 275-286.
GANGE, S. J. (1994) Multivariate ordinal responses in clinical trials, University of Wisconsin--
Madison.
GENEST, C. & NEŠLEHOVÁ, J. (2007) A primer on copulas for count data. Astin Bulletin, 37,
475-515.
GODFRAY, H. C. J., BEDDINGTON, J. R., CRUTE, I. R., HADDAD, L., LAWRENCE, D.,
MUIR, J. F., PRETTY, J., ROBINSON, S., THOMAS, S. M. & TOULMIN, C. (2010)
Food security: the challenge of feeding 9 billion people. science, 327, 812-818.
GOMES, M., RADICE, R., CAMARENA BRENES, J. & MARRA, G. (2019) Copula selection
models for non― Gaussian outcomes that are missing not at random. Statistics in
medicine, 38, 480-496.
GRAY, S. M. & BROOKMEYER, R. (2000) Multidimensional longitudinal data: estimating a
treatment effect from continuous, discrete, or time-to-event response variables. Journal of
the American Statistical Association, 95, 396-406.
GUEORGUIEVA, R. (2001) A multivariate generalized linear mixed model for joint modelling
of clustered outcomes in the exponential family. Statistical Modelling, 1, 177-193.
GUMBEL, E. J. (1960) Bivariate exponential distributions. Journal of the American Statistical
Association, 55, 698-707.
HABYARIMANA, J. B. (2015) Determinants of household food insecurity in developing
countries evidences from a probit model for the case of rural households in Rwanda.
Sustainable Agriculture Research, 4.
HAFF, I. H. (2012) Comparison of estimators for pair-copula constructions. Journal of
Multivariate Analysis, 110, 91-105.
HARVILLE, D. A. & MEE, R. W. (1984) A mixed-model procedure for analyzing ordered
categorical data. Biometrics, 393-408.
HEAGERTY, P. J. & ZEGER, S. L. (1996) Marginal regression models for clustered ordinal
measurements. Journal of the American Statistical Association, 91, 1024-1036.

136 | P a g e
© Yimam JA, UNISA 2019
HIRK, R., HORNIK, K. & VANA, L. (2018) Multivariate ordinal regression models: an analysis
of corporate credit ratings. Statistical Methods & Applications, 28, 507-539.
HUANG, G. H., BANDEENۥ ROCHE, K. & RUBIN, G. S. (2002) Building marginal models
for multiple ordinal measurements. Journal of the Royal Statistical Society: Series C
(Applied Statistics), 51, 37-57.
HUNNES, D. E. (2013) Nutrition and Food Security in a Changing Climate: Methods for
predicting household coping strategy use and Food Security in the Ethiopian Context.
UCLA.
HUYNH, V.-N., KREINOVICH, V. & SRIBOONCHITTA, S. (2014) Modeling dependence in
econometrics, Springer.
JANSEN, J. (1990) On the statistical analysis of ordinal data when extravariation is present.
Applied Statistics, 75-84.
JIANG, Z. (2012) Joint Modeling of Multivariate Ordinal Longitudinal Outcome. University of
Pittsburgh.
JOE, H. (1996) Families of m-variate distributions with given margins and m (m-1)/2 bivariate
dependence parameters. Lecture Notes-Monograph Series, 120-141.
JOE, H. (1997) Multivariate models and multivariate dependence concepts, CRC Press.
KELLY, J. L. & PEMBERTON, C. (2016) An assessment of the household food security status
and local foods grown in rural Bahamas. Farm and Business-The Journal of the
Caribbean Agro-Economic Society, 8, 82.
KENNE PAGUI, E. C. & CANALE, A. (2016) Pairwise likelihood inference for multivariate
ordinal responses with applications to customer satisfaction. Applied Stochastic Models
in Business and Industry, 32, 273-282.
KISI, M. A., TAMIRU, D., TESHOME, M. S., TAMIRU, M. & FEYISSA, G. T. (2018)
Household food insecurity and coping strategies among pensioners in Jimma Town,
South West Ethiopia. BMC public health, 18, 1373.
KOPER, N. & MANSEAU, M. (2009) Generalized estimating equations and generalized linear
mixed-effects models for modelling resource selection. Journal of Applied Ecology, 590-
599.
KURADA, R. R. (2011) Modeling and analysis of repeated ordinal data using copula based
likelihoods and estimating equation methods, Old Dominion University.

137 | P a g e
© Yimam JA, UNISA 2019
LAFFONT, C. M., VANDEMEULEBROECKE, M. & CONCORDET, D. (2014) Multivariate
analysis of longitudinal ordinal data with mixed effects models, with application to
clinical outcomes in osteoarthritis. Journal of the American Statistical Association, 109,
955-966.
LEE, K., DANIELS, M. J. & JOO, Y. (2013) Flexible marginalized models for bivariate
longitudinal ordinal data. Biostatistics, 14, 462-476.
LENNON, H. (2016) Gaussian copula modelling for integer-valued time series. University of
Manchester.
LIANG, K.-Y. & ZEGER, S. L. (1986) Longitudinal data analysis using generalized linear
models. Biometrika, 13-22.
LIPSITZ, S. R., KIM, K. & ZHAO, L. (1994) Analysis of repeated categorical data using
generalized estimating equations. Statistics in medicine, 13, 1149-1163.
LIU, J. (2007) Multivariate ordinal data analysis with pairwise likelihood and its extension to
SEM. University of California Los Angeles.
LIU, L. C. (2008) A model for incomplete longitudinal multivariate ordinal data. Statistics in
medicine, 27, 6299-6309.
LIU, L. C. & HEDEKER, D. (2006) A mixed‐effects regression model for longitudinal
multivariate ordinal data. Biometrics, 62, 261-268.
MAGAÑA-LEMUS, D., ISHDORJ, A., ROSSON, C. P. & LARA-ÁLVAREZ, J. (2016)
Determinants of household food insecurity in Mexico. Agricultural and Food Economics,
4, 10.
MBOLANYI, B., EGERU, A. & MFITUMUKIZA, D. (2017) Determinants of household food
security in a rangeland area of Uganda. African Journal of Rural Development, 2, 213-
223.
MENSAH, O., JAMES, A. & TUFFOUR, T. (2013) Determinants of household food security in
the Sekyere-Afram plains district of Ghana. Global Advanced Research Journal of
Agricultural Science, 2, 34-40.
MÉTHOT, J. & BENNETT, E. M. (2018) Reconsidering non-traditional export agriculture and
household food security: A case study in rural Guatemala. PloS one, 13, e0198113.
MOLENBERGHS, G. & VERBEKE, G. (2006) Models for discrete longitudinal data, Springer
Science & Business Media.

138 | P a g e
© Yimam JA, UNISA 2019
MORODA, G. T., TOLOSSA, D. & SEMIE, N. (2018) Food insecurity of rural households in
Boset district of Ethiopia: a suite of indicators analysis. Agriculture & Food Security, 7,
65.
MOTBAINOR, A., WORKU, A. & KUMIE, A. (2016) Level and determinants of food
insecurity in East and West Gojjam zones of Amhara Region, Ethiopia: a community
based comparative cross-sectional study. BMC public health, 16, 503.
NAPIER, C., OLDEWAGE-THERON, W. & MAKHAYE, B. (2018) Predictors of food
insecurity and coping strategies of women asylum seekers and refugees in Durban, South
Africa. Agriculture & Food Security, 7, 67.
NAPOLI, M., DE MURO, P. & MAZZIOTTA, M. (2011) Towards a food insecurity
Multidimensional Index (FIMI). Master in Human Development and Food Security.
NEGATU, W. (2004) Reasons for food insecurity of farm households in South Wollo, Ethiopia:
explanations at grassroots.
NELSEN, R. B. (2007) An introduction to copulas, Springer Science & Business Media.
NGEMA, P., SIBANDA, M. & MUSEMWA, L. (2018) Household Food Security Status and Its
Determinants in Maphumulo Local Municipality, South Africa. Sustainability, 10, 3307.
NICKLAS, S. (2013) Pair Constructions for High-Dimensional Dependence Models in Discrete
and Continuous Time. Universität zu Köln.
NIGUSSIE, Z. & ALEMAYEHU, G. (2013) Levels of household food insecurity in rural areas
of Guraghe zone, Southern Ethiopia. Wudpecker J. Agric. Res, 2, 8-14.
NIKOLOULOPOULOS, A. K. (2017) Weighted scores method for longitudinal ordinal data.
arXiv preprint arXiv:1510.07376.
NOORAEE, N., ABEGAZ, F., ORMEL, J., WIT, E. & VAN DEN HEUVEL, E. R. (2016) An
approximate marginal logistic distribution for the analysis of longitudinal ordinal data.
Biometrics, 72, 253-261.
NOORAEEA, N. (2015) Statistical methods for marginal inference from multivariate ordinal
data, Doctoral Thesis. 205p, University of Groningen.
OLAOMI, J. & YIMAM, J. A. (2019) Modeling the Stability and Determinant Factors of
Household Food Insecurity: A Pair Copula Construction Approach. Joint Conference of
the Sub-Saharan African Network (SUSAN) of the International Biometrics Society (IBS)

139 | P a g e
© Yimam JA, UNISA 2019
 and DELTAS Africa Sub-Saharan Africa Consortium for Advanced
Biostatistics(SSACAB).
PANAGIOTELIS, A., CZADO, C. & JOE, H. (2012) Pair copula constructions for multivariate
discrete data. Journal of the American Statistical Association, 107, 1063-1072.
PANAGIOTELIS, A., CZADO, C., JOE, H. & STÖBER, J. (2015) Model selection for discrete
regular vine copulas.
PERIN, J. (2009) Improved generalized estimating equations for incomplete longitudinal binary
data, covariance estimation in small samples, and ordinal data.
PINSTRUP-ANDERSEN, P. (2009) Food security: definition and measurement. Food security,
1, 5-7.
PIRKTL, V. (2007) Copula models and dependence concepts. UNIVERSITY OF SOUTHERN
CALIFORNIA.
PITT, M., CHAN, D. & KOHN, R. (2006) Efficient Bayesian inference for Gaussian copula
regression models. Biometrika, 93, 537-554.
ROCHON, J. (1996) Analyzing bivariate repeated measures for discrete and continuous outcome
variables. Biometrics, 740-750.
RUSCONE, M. N. & OSMETTI, S. A. (2017) Modelling the dependence in multivariate
longitudinal data by pair copula decomposition. Soft Methods for Data Science. Springer.
RYU, J.-H. & BARTFELD, J. S. (2012) Household food insecurity during childhood and
subsequent health status: the early childhood longitudinal study—kindergarten cohort.
American Journal of Public Health, 102, e50-e55.
SAVU, C. & TREDE, M. (2006) Hierarchical Archimedean Copulas: International Conference
on High Frequency Finance. Konstanz, Germany.
SCHEPSMEIER, U., STOEBER, J., BRECHMANN, E. C., GRAELER, B., NAGLER, T.,
ERHARDT, T., ALMEIDA, C., MIN, A., CZADO, C. & HOFMANN, M. (2015)
Package ‘VineCopula’. R package version, 2.
SCHIRMACHER, D. & SCHIRMACHER, E. (2008) Multivariate dependence modeling using
pair-copulas. Technical report.
SHI, P. & YANG, L. (2016) Pair copula constructions for insurance experience rating. Journal
of the American Statistical Association, 113, 122-133.

140 | P a g e
© Yimam JA, UNISA 2019
SHI, P. & ZHAO, Z. (2018) Predictive Modeling of Multivariate Longitudinal Insurance Claims
Using Pair Copula Construction. arXiv preprint arXiv:1805.07301.
SHONE, M., DEMISSIE, T., YOHANNES, B. & YOHANNIS, M. (2017) Household food
insecurity and associated factors in West Abaya district, Southern Ethiopia, 2015.
Agriculture & Food Security, 6, 2.
SIRISRISAKULCHAI, J. & SRIBOONCHITTA, S. (2014) Modeling dependence of accident-
related outcomes using pair copula constructions for discrete data. Modeling Dependence
in Econometrics. Springer International Publishing, 215-228.
SKLAR, M. (1959) Fonctions de répartition à n dimensions et leurs marges, Université Paris 8.
SMITH, M., MIN, A., ALMEIDA, C. & CZADO, C. (2010) Modeling longitudinal data using a
pair-copula decomposition of serial dependence. Journal of the American Statistical
Association, 105, 1467-1479.
SMITH, M. S. (2015) Copula modelling of dependence in multivariate time series. International
Journal of Forecasting, 31, 815-833.
STÖBER, J., HONG, H. G., CZADO, C. & GHOSH, P. (2015) Comorbidity of chronic diseases
in the elderly: Patterns identified by a copula design for mixed responses. Computational
Statistics & Data Analysis, 88, 28-39.
SUTKOFF, A. (2014) A regular vine-copula approach to endogenous regressors in brand value
estimation. THE UNIVERSITY OF CHICAGO.
SWINDALE, A. & BILINSKY, P. (2007) Household food insecurity access scale (HFIAS) for
measurement of household food access: indicator guide (v. 3). Washington, DC: Food
and Nutrition Technical Assistance Project, Academy for Educational Development.
SYRING, N. A. (2013) Multivariate binary regression models with applications in health care
utilisation. NORTHERN ILLINOIS UNIVERSITY.
TANTU, A. T., GAMEBO, T. D., SHENO, B. K. & KABALO, M. Y. (2017) Household food
insecurity and associated factors among households in Wolaita Sodo town, 2015.
Agriculture & Food Security, 6, 19.
TOULOUMIS, A., AGRESTI, A. & KATERI, M. (2013) GEE for multinomial responses using
a local odds ratios parameterization. Biometrics, 69, 633-640.
TUTZ, G. & HENNEVOGL, W. (1996) Random effects in ordinal regression models.
Computational Statistics & Data Analysis, 22, 537-557.

141 | P a g e
© Yimam JA, UNISA 2019
VANDENHENDE, F. O. & LAMBERT, P. (2000) Modeling repeated ordered categorical data
using copulas. Dis cussion Paper No 00-25: www. stat. ucl. ac.
be/pub/papers/dp/dp00/dpO025. ps Institut de statistique, Uni versit.
VARADHAN, R. & GROTHENDIECK, G. (2011) alabama: Constrained nonlinear
optimization. R package version, 1, 2012.
VERBEKE, G., FIEUWS, S., MOLENBERGHS, G. & DAVIDIAN, M. (2014) The analysis of
multivariate longitudinal data: A review. Statistical methods in medical research, 23, 42-
59.
VERBEKE, G., FIEUWS, S., MOLENBERGHS, G. & DAVIDIAN, M. (2017) The analysis of
multivariate longitudinal data: A review. Statistical methods in medical research, 26,
112-112.
WINEMAN, A. (2016) Multidimensional household food security measurement in rural Zambia.
Agrekon, 55, 278-301.
YANG, L., FREES, E. W. & ZHANG, Z. (2020) Nonparametric estimation of copula regression
models with discrete outcomes. Journal of the American Statistical Association, 115,
707-720.
ZERAY, D. D. N. (2017) Determinants of Rural Household Food Security in Wolaita Zone: The
Case of Humbo Woreda. Journal of Poverty, Investment and Development 32, ISSN
2422-846X.
ZHANG, W., ZHANG, M. & CHEN, Y. (2019) A Copula-Based GLMM Model for Multivariate
Longitudinal Data with Mixed-Types of Responses. Sankhya B, 1-27.

142 | P a g e
© Yimam JA, UNISA 2019
 Appendices

Appendix A: Questionnaires

UNIVERSITY OF SOUTH AFRICA

COLLEGE OF SCIENCE, ENGINEERING AND TECHNOLOGY

Department of Statistics

Household Questionnaire for Modelling the Stability and Determinants of Household Food
Insecurity, February, 2014-2015

Introduction and Consent

My name is___________ and I am attending post graduate class at University of South Africa
(UNISA). We are conducting an assessment on modelling the determinant factors of household
food insecurity using longitudinal multivariate ordinal logistic regression model. I would like to
ask you some questions about you, your household, risks you face relating to food you are
engaged in. The questionnaire usually takes between 20 -25 minutes to complete.

Whatever information you provide will be kept strictly confidential and will not be shown to
other persons. Participation in this assessment is voluntary and you can choose not to answer
any individual questions or all of the questions. However, we hope that you will participate fully
in this assessment since your views are important.

Do you have any questions about the survey? May I begin the interview now?

VERBAL CONSENT GIVEN TO INTERVIEW, CHECK BOX

Interview Information
Date of interview: (dd/mm/yyyy) ____/_____/______

Interviewer's name__________________________Signature__________________

Name of supervisor__________________________Signature__________________
Questionnaire ID: ____________________________

143 | P a g e
© Yimam JA, UNISA 2019
1. Area Identification
No Question Response
1.1 Woreda ____________________
1.2 Kebele ____________________
1.3 Got __________________
1.4 Household ID _____________________
2. Demographic and Socioeconomic Characteristics
No Question Response Skip
2.1 Respondent 1. Household head
2. Housewife
3. Son/daughter
4. Other
2.2 Age of the respondent in years ___________________
2.3 Age of the household head in years ___________________
2.4 Sex of respondent 1. Male
2. Female
2.5 Sex of the household head 1. Male
2. Female
2.6 Who is the household head? 1. Husband
2. Wife
3. Son/daughter
4. Other (specify)---------
2.7 Family size __________________
2.8 Number of under 5 Children __________________
2.9 Whom do the household head live with? 1. Alone
2. Spouse/partner
3. Parents
4. Relatives
5. Others (specify)--------------
2.10 What is the highest level of education of the 1. Unable to read and write

144 | P a g e
© Yimam JA, UNISA 2019
 household head attained? 2. Can read and write
3. Regular Primary education (1-8)
4. General secondary education (9-10)
5. Preparatory education (11-12)
6. TVET
7. College /university education
2.11 What is current marital status of the 1. Never married
household’s head? 2. Cohabiting
3. Married
4. Divorced
5. Widowed
2.12 What is the household head current occupation? 1. Student
(Select that all apply) 2. Unemployed
3. Professional employment
4. Self employed
5. Domestic worker
6. Casual worker
7. Housewife
8. Other (Specify)------------------

3. Economic and Income Related Questions

3.1 Main source of household income 1. Farming

(Select that all apply) 2. Herding 3.9
3. Merchant 3.10
4. Daily labourer 3.12
5. Other (specify)--------------
3.2 How much is your total farmland size? Land size in
1. Hectares ___________
2. Timad _____________
3. Gasha ______________
4. Other ___________

145 | P a g e
© Yimam JA, UNISA 2019
3.3 Slope of your land 1. Plain
2. Hilly
3. Steep
3.4 How do you perceive the quality or fertility of your 1. Fertile
land? 2. Medium fertile
3. Less fertile
4. Poor
3.5 How many times do you cultivate within a year? 1. Yearly
2. Biannual
3. Three-times
3.6 What type of the following cereals did you harvest 1. Barely
during the last 12 months? 2. Millet
(Select that all apply) 3. Wheat
4. Sorghum
5. Teff
6. Bean
7. Pea
8. Others specify ---------
3.7 What are the main problems for farmers’ incomes 1. Pests
in your village? 2. Rainfall shortage
(Select that all apply) 3. Disease
4. Lack of improved agricultural
product input
5. Households head death
6. Excessive temperature
7. Excess rainfall
8. Increase in market price
9. Fall in market price
10. Property loss
11. Other specify ---------------------
3.8 Have you used any of the following agricultural 1. Chemical fertilizer

146 | P a g e
© Yimam JA, UNISA 2019
 technologies during the last 12 months production 2. Pesticides
season? 3. Improved seeds
(Select that all apply) 4. Farm credit
5. Access to irrigation water
6. Nothing
7. Compost
8. Others Specify ------------------
3.9 What type of livestock do you have? 1. Ox 6. Donkey
(Select that all apply) 2. Cow 7. Mule
3. Sheep 8. Camel
4. Goat 9. Chicken
5. Horse 10. Other
(Specify)------
3.10 The weathering condition of your village? 1. Hot (Kolla)
2. Medium (Wenadega)
3. Cold (Dega)
3.11 How was the availability of rain on your village 1. Very high
since last year? 2. High
3. Enough
4. Little
5. Very little
6. Too much
7. Too little
8. Other specify _____________
3.12 Which season is the main production season in your 1. Winter (Dec-Feb)
village? 2. Summer (Jun-Aug)
(Select that all apply) 3. Autumn (March-May)
4. Spring (Sep-Nov)
3.13 Member in the household contributing financially
to incomes ______________
3.14 Number of persons contributing financially to

147 | P a g e
© Yimam JA, UNISA 2019
 incomes ________________
3.15 Average monthly income of your family 1. Less than 500
2. 500-1500
3. 1501-2500
4. 2501-3000
5. 3001-4800
6. 4801-5000
7. Greater than 5001(Specify) ------
4. Household Food Insecurity Availability, Access and Utilisation Scale tool
A. Household Food Insecurity Utilisation Scale Tool
4.1 In the past [24 hours], did you or any household member ate CEREAL CROPS (bread, 1. Yes
noodles, biscuits, cookies or any other foods made from millet, sorghum, maize, rice, 2. No
wheat other locally available grains)
4.2 In the past [24 hours], did you or any household member ate VITAMIN A RICH 1. Yes
VEGETABLES AND FRUITS (carrots, squash, sweet potatoes, ripe mangoes, 2. No
papayas or other locally available vitamin A-rich fruits or vegetables)
4.3 In the past [24 hours], did you or any household member ate MEAT (beef, pork, lamb, 1. Yes
goat, rabbit, wild game, chicken, duck, or other birds, liver, kidney, heart or other 2. No
organ meats or blood-based foods)
4.4 In the past [24 hours], did you or any household member ate EGGS 1. Yes
2. No
4.5 In the past [24 hours], did you or any household member ate FISH (fresh or dried fish 1. Yes
or shellfish) 2. No
4.6 In the past [24 hours], did you or any household member ate LEGUMES, NUTS AND 1. Yes
SEEDS (beans, peas, lentils, nuts, seeds or foods made from these) 2. No
4.7 In the past [24 hours], did you or any household member ate MILK AND MILK 1. Yes
PRODUCTS (milk, cheese, yogurt or other milk products) 2. No
4.8 In the past [24 hours], did you or any household member ate OILS AND FATS (oil, 1. Yes
fats or butter added to food or used for cooking) 2. No
4.9 In the past [24 hours], did you or any household member ate SWEETS (sugar, honey, 1. Yes
sweetened soda or sugary foods such as chocolates, sweets or candies) 2. No

148 | P a g e
© Yimam JA, UNISA 2019
4.10 Did your households get enough and safe drinking water? 1. Yes
2. No
4.11 Did bone problem happen among your under five families? 1. Yes
2. No
4.12 Did diarrhoea disease mostly happen among your under five families? 1. Yes
2. No
4.13 Did anaemia disease happen among your under five families? 1. Yes
2. No
4.14 Did pregnant woman take balanced diet food than the other family members? 1. Yes
2. No
4.15 Did breast feeding woman take balanced diet food than the other family members? 1. Yes
2. No
4.16 Did you prepare appropriate place for dusts? 1. Yes
2. No
4.17 Does someone among your family members who didn’t eat food that others ate it? 1. Ye
2. No
4.18 Has someone who can eat raw food (raw meat, milk and others) among your family 1. Yes
members? 2. No

4.19 Do you toilet? 1. Yes

2. No
B. Household Food Insecurity Availability Scale tool
4.20 Have you plough land for cereal crops? 1. Yes
2. No
4.21 If yes for 4.20, which of these 1. Enough of the kinds of food we want to eat 4.23
statements best describes any cereal 2. Enough but not always the kinds of food we
crops eaten in your household in the last want
12 months through your own 3. Sometimes not enough to eat
production? 4. Often not enough to eat
4.22 Which of these statements best 1. Enough of the kinds of food we want to eat
describes that you usually able to buy all 2. Enough but not always the kinds of food we

149 | P a g e
© Yimam JA, UNISA 2019
 of the cereal crops that you need for you want
and your family from the local market 3. Sometimes not enough to eat
of your village or surrounding? 4. Often not enough to eat
4.23 Have you plough land for fruits? 1. Yes
2. No
4.24 If yes for 4.23, which of these 1. Enough of the kinds of fruit we want to eat 4.26
statements best describes any fresh fruit 2. Enough but not always the kinds of fruit we
eaten in your household in the last 12 want
months through your own production? 3. Sometimes not enough to eat
Interviewer: Do not include juice or fruit 4. Often not enough to eat
that is frozen or canned.
4.25 Which of these statements best 1. Enough of the kinds of fruit we want to eat
describes that you usually able to buy all 2. Enough but not always the kinds of fruit we
of the fruit that you need for you and want
your family from the local market of 3. Sometimes not enough to eat
your village or surrounding? 4. Often not enough to eat
4.26 Have you plough land for vegetables? 1. Yes
2. No
4.27 If yes for 4.26, which of these 1. Enough of the kinds of vegetables we want 4.27
statements best describes any vegetables to eat
eaten in your household in the last 12 2. Enough but not always the kinds of
months through your own production? vegetables we want
3. Sometimes not enough to eat
4. Often not enough to eat
4.28 Which of these statements best 1. Enough of the kinds of vegetables we want
describes that you usually able to buy all to eat
of the vegetables that you need for you 2. Enough but not always the kinds of
and your family from the local market vegetables we want
of your village or surrounding? 3. Sometimes not enough to eat
4. Often not enough to eat
4.29 Have you animals that produces milk? 1. Yes

150 | P a g e
© Yimam JA, UNISA 2019
 2. No
4.30 If yes for 4.29, which of these 1. Enough of the kinds of food we want to eat 4.32
statements best describes any milk 2. Enough but not always the kinds of food we
products eaten in your household in the want
last 12 months through your own 3. Sometimes not enough to eat
production? 4. Often not enough to eat
4.31 Which of these statements best 1. Enough of the kinds of food we want to eat
describes that you usually able to buy all 2. Enough but not always the kinds of food we
of the milk products that you need for want
you and your family from the local 3. Sometimes not enough to eat
market of your village or surrounding? 4. Often not enough to eat
4.32 Have you animals that produces milk? 1. Yes
2. No
4.33 Which of these statements best 1. Enough of the kinds of food we want to eat 4.35
describes any meat products eaten in 2. Enough but not always the kinds of food we
your household in the last 12 months want
through your own production? 3. Sometimes not enough to eat
4. Often not enough to eat
4.34 Which of these statements best 1. Enough of the kinds of food we want to eat
describes that you usually able to buy all 2. Enough but not always the kinds of food we
of the meat products that you need for want
you and your family from the local 3. Sometimes not enough to eat
market of your village or surrounding? 4. Often not enough to eat
4.35 Which of these statements best 1. Enough of the kinds of food we want to eat
describes the kind of foods eaten in your 2. Enough but not always the kinds of food we
household in the last 12 months through want
food aid in your village 3. Sometimes not enough to eat
4. Often not enough to eat
5. Not aided
4.36 Which of these statements best 1. Efficient water we want to use
describes the kind of water used in your 2. Efficient but not always the kinds of water

151 | P a g e
© Yimam JA, UNISA 2019
 household in the last 12 months in your we want
village? 3. Sometimes not efficient to use
4. Often not efficient to use
C. Household Food Insecurity Access Scale tool
4.37 What best describes the food consumed in 1. Always enough of what wanted
the household during the past 12 2. Enough but not always what wanted
months.(due to lack of money to buy food) 3. Sometimes not enough food
4. Often not enough food
4.38 In past 12 months were you and your 1. No
household members worried that your 2. Yes
food would run out before you had money *No follow up question on frequency
to buy more?
4.39 In past 12 months did you have to eat the 1. No
same food daily because you did not have 2. Yes
money to buy other food? *No follow up question on frequency
4.40 In the past 12 months have you or any 1. No
other adult in your household eaten less 2. Yes How often?
food than you wanted to because you did 1. More than half the time
not have enough money to buy food? 2. Less than half the time but more than 30 days
3. Less than 30 days but more than 10 days
4. Less than 10 days
4.41 Did you or another adult in your 1. No
household skip meals during the past 12 2. Yes How often?
months because you did not have enough 1. More than half the time
money to buy food? 2. Less than half the time but more than 30 days
3. Less than 30 days but more than 10 days
4. Less than 10 days
4.42 Did you or another adult in your 1. No
household stop eating for an entire day 2. Yes How often?
(during the past 12 months) because you 1. Less than half the time but more than 30 days
did not have enough money to buy food? 2. Less than 30 days but more than 10 days

152 | P a g e
© Yimam JA, UNISA 2019
 3. Less than 10 days
4.43 In the past 12 months, did you or anyone 1. No
in the household borrow money for food 2. Yes How often?
from friends or relatives? 1. Less than half the time but more than 30 days
2. Less than 30 days but more than 10 days
3. Less than 10 days
4.44 In the past 12 months, did you or anyone 1. No
in the household buy food on a credit 2. Yes How often?
account or credit card? 1. Less than half the time but more than 30 days
2. Less than 30 days but more than 10 days
3. Less than 10 days
5 Household Food Insecurity Coping Mechanisms
5.1 Since crises, how did you 1. Eating less/skipping meals
overcome the food 2. Eating food less preferred
3. Food or cash aid
shortage?
4. Migrating household head to other villages
(Select that all apply) 5. Migrating the younger household members to town
6. Selling assets
7. Eating wild food
8. Selling trees
9. Gardening (to grow food, mainly vegetables and green leaves)
10. Trade (commercial activities)
11. Little crafts
12. Small livestock raising
13. Other/specify --------------------------------
5.2 Did anyone in the family 1. Yes, emergency food rations
benefit from food aid 2. Yes, safety net food rations
rations in the last one year? 3. No
(Select that all apply) 4. Other (specify)…………….
5.3 If you are safety net user 1. Yes
have you graduate now? 2. No
3. No safety net user

153 | P a g e
© Yimam JA, UNISA 2019
Appendix B: Internal consistence analysis of the data collection tools.

Items Scale Mean Scale Corrected Cronbach's

if Item Variance if Item-Total Alpha if Item
Deleted Item Deleted Correlation Deleted
Any household member ate CEREAL CROPS 61.00 74.572 .114 .737
Any household member ate VITAMIN A RICH 60.26 72.642 .243 .731
VEGETABLES AND FRUITS
Any household member ate MEAT 60.12 74.049 .115 .735
Any household member ate EGGS 60.13 73.863 .116 .735
Any household member ate FISH 60.05 74.678 .112 .737
Any household member ate LEGUMES, NUTS 60.85 73.268 .173 .733
AND SEEDS
Any household member ate MILK AND MILK 60.31 72.401 .257 .730
PRODUCTS
Any household member ate OILS AND FATS 60.73 72.182 .278 .729
Any household member ate SWEETS 60.66 71.920 .294 .728
Did your households get enough and safe 60.70 73.010 .166 .733
drinking water
Any cereal crops eaten in your household in the 58.90 67.433 .360 .721
last 12 months through your own production
You are usually able to buy all of the cereal crops 60.46 71.192 .183 .732
that you need for you and your family from the
local market
Any fresh fruit eaten in your household in the last 57.53 67.432 .365 .720
12 months through your own production
You are usually able to buy all of the fruit that 59.60 67.093 .298 .726
you need for you and your family from the local
market
Any vegetables eaten in your household in the 57.78 65.426 .417 .715
last 12 months through your own production

154 | P a g e
© Yimam JA, UNISA 2019
You are usually able to buy all of the vegetables 60.11 67.892 .303 .725
that you need for you and your family from the
local market
Any milk products eaten in your household in the 58.24 66.221 .350 .721
last 12 months through your own production
You are usually able to buy all of the milk 59.53 66.916 .297 .726
products that you need for you and your family
from the local market
Any meat products eaten in your household in the 58.02 68.625 .305 .725
last 12 months through your own production
You are usually able to buy all of the meat 60.08 70.398 .147 .738
products that you need for you and your family
from the local market
The kind of water used in your household in the 60.22 71.127 .145 .736
last 12 months in your village
Food consumed in the household during the past 59.11 66.207 .515 .711
12 months. (due to lack of money to buy food)
You and your household members worried that 60.18 71.795 .431 .726
your food would run out before you had money to
buy more
You have to eat the same food daily because you 60.51 72.332 .233 .730
did not have money to buy other food
You or any other adult in your household eaten 59.59 64.010 .372 .720
less food than you wanted to because you did not
have enough money to buy food
You or another adult in your household skip 60.36 68.241 .307 .725
meals during the past 12 months because you did
not have enough money to buy food

155 | P a g e
© Yimam JA, UNISA 2019
You or another adult in your household stop 60.94 73.694 .108 .735
eating for an entire day (during the past 12
months) because you did not have enough money
to buy food
You or anyone in the household borrow money 60.37 67.808 .290 .726
for food from friends or relatives
You or anyone in the household buy food on a 60.49 68.342 .262 .728
credit account or credit card
Cronbach's Alpha .735

156 | P a g e
© Yimam JA, UNISA 2019
Appendix C: The joint probability distribution based on the D-vine pair copula was displayed as follows.

Pr(Y1 = y1 , Y2 = y 2 , Y3 = y 3 )
C12 ( F1 ( y1 − i1 ), F2 ( y 2 )) − C12 ( F1 ( y1 − i1 ), F2 ( y 2 − 1)) C 23 ( F2 ( y 2 ), F3 ( y 3 − i3 )) − C 23 ( F2 ( y 2 − 1), F3 ( y 3 − i3 ))
={   (−1) i1 + i3
C13|2 ( , )}[F2 ( y 2 ) − F2 ( y 2 − 1)].
i1 = 0 , 1 i3 = 0 , 1 F2 ( y 2 ) − F2 ( y 2 − 1) F2 ( y 2 ) − F2 ( y 2 − 1)
Among the six proposed bivariate copula functions, AMH bivariate copula was selected for F1 ( y1 ) and F2 ( y2 ) , Frank for F2 ( y2 ) and
F3 ( y3 ) , and Frank for F1|2 ( y1 | y2 ) and F3|2 ( y3 | y2 ) . Hence the simplified joint probability distribution based on the D pair copula was

displayed as follows.
when i1 = 0 and i3 = 0 ,
𝐴𝑀𝐻 ++
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 )) = 𝐹1 (𝑦1 ) ∗ 𝐹2 (𝑦2 )/(1−𝜃12 (1 − 𝐹1 (𝑦1 )) ∗ (1 − 𝐹2 (𝑦2 )))
𝐴𝑀𝐻 +−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 − 1)) = 𝐹1 (𝑦1 ) ∗ 𝐹2 (𝑦2 − 1)/(1−𝜃12 (1 − 𝐹1 (𝑦1 )) ∗ (1 − 𝐹2 (𝑦2 − 1)))

((exp (−𝜃23 𝐹2 (𝑦2 )) − 1)(exp (−𝜃23 𝐹3 (𝑦3 )) − 1))

= 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 )) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 ++
𝐶23 )
23 exp(−𝜃23 ) − 1

(exp (−𝜃23 𝐹2 (𝑦2 − 1)) − 1)(exp (−𝜃23 𝐹3 (𝑦3 )) − 1)

= 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 )) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 −+
𝐶23 )
23 exp(−𝜃23 ) − 1
++ +− ++ −+
𝐶 𝐴𝑀𝐻 − 𝐶12 𝐴𝑀𝐻 𝐹𝑟
𝐶23 − 𝐶23𝐹𝑟
(exp (−𝜃13|2 ( 12 )) − 1) (exp (−𝜃13|2 ( )) − 1)
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
= −1⁄𝜃
𝐹𝑟 00
𝐶13|2 𝑙𝑜𝑔 1 +
13|2 exp(−𝜃13|2 ) − 1
( )

when i1 = 1 and i3 = 0
𝐴𝑀𝐻 −+
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 )) = 𝐹1 (𝑦1 − 1) ∗ 𝐹2 (𝑦2 )/(1−𝜃12 (1 − 𝐹1 (𝑦1 − 1)) ∗ (1 − 𝐹2 (𝑦2 )))

137 | P a g e
 𝐴𝑀𝐻 −−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 − 1)) = 𝐹1 (𝑦1 − 1) ∗ 𝐹2 (𝑦2 − 1)/(1−𝜃12 (1 − 𝐹1 (𝑦1 − 1)) ∗ (1 − 𝐹2 (𝑦2 − 1)))

((exp (−𝜃23 𝐹2 (𝑦2 )) − 1)(exp (−𝜃23 𝐹3 (𝑦3 )) − 1))

= 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 )) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 ++
𝐶23 )
23 exp(−𝜃23 ) − 1

(exp (−𝜃23 𝐹2 (𝑦2 − 1)) − 1)(exp (−𝜃23 𝐹3 (𝑦3 )) − 1)

= 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 )) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 −+
𝐶23 )
23 exp(−𝜃23 ) − 1
−+ −− ++ −+
𝐶 𝐴𝑀𝐻 − 𝐶12 𝐴𝑀𝐻 𝐹𝑟
𝐶23 − 𝐶23𝐹𝑟
(exp (−𝜃13|2 ( 12 )) − 1) (exp (−𝜃13|2 ( )) − 1)
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
= −1⁄𝜃
𝐹𝑟 10
𝐶13|2 𝑙𝑜𝑔 1 +
13|2 exp(−𝜃13|2 ) − 1
( )
when i1 = 0 and i3 = 1
𝐴𝑀𝐻 ++
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 )) = 𝐹1 (𝑦1 ) ∗ 𝐹2 (𝑦2 )/(1−𝜃12 (1 − 𝐹1 (𝑦1 )) ∗ (1 − 𝐹2 (𝑦2 )))
𝐴𝑀𝐻 +−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 − 1)) = 𝐹1 (𝑦1 ) ∗ 𝐹2 (𝑦2 − 1)/(1−𝜃12 (1 − 𝐹1 (𝑦1 )) ∗ (1 − 𝐹2 (𝑦2 − 1)))

((exp (−𝜃23 𝐹2 (𝑦2 )) − 1)(exp (−𝜃23 𝐹3 (𝑦3 − 1)) − 1))

= 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 − 1)) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 +−
𝐶23 )
23 exp(−𝜃23 ) − 1

(exp (−𝜃23 𝐹2 (𝑦2 − 1)) − 1)(exp (−𝜃23 𝐹3 (𝑦3 − 1)) − 1)

= 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 − 1)) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 −−
𝐶23 )
23 exp(−𝜃23 ) − 1
++ +− +− −−
𝐶 𝐴𝑀𝐻 − 𝐶12 𝐴𝑀𝐻 𝐹𝑟
𝐶23 − 𝐶23𝐹𝑟
(exp (−𝜃13|2 ( 12 )) − 1) (exp (−𝜃13|2 ( )) − 1)
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
= −1⁄𝜃
𝐹𝑟 01
𝐶13|2 𝑙𝑜𝑔 1 +
13|2 exp(−𝜃13|2 ) − 1
( )
when i1 = 1 and i3 = 1 ,

138 | P a g e
 𝐴𝑀𝐻 −+
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 )) = 𝐹1 (𝑦1 − 1) ∗ 𝐹2 (𝑦2 )/(1−𝜃12 (1 − 𝐹1 (𝑦1 − 1)) ∗ (1 − 𝐹2 (𝑦2 )))
𝐴𝑀𝐻 −−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 − 1)) = 𝐹1 (𝑦1 − 1) ∗ 𝐹2 (𝑦2 − 1)/(1−𝜃12 (1 − 𝐹1 (𝑦1 − 1)) ∗ (1 − 𝐹2 (𝑦2 − 1)))

((exp (−𝜃23 𝐹2 (𝑦2 )) − 1)(exp (−𝜃23 𝐹3 (𝑦3 − 1)) − 1))

= 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 − 1)) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 +−
𝐶23 )
23 exp(−𝜃23 ) − 1

(exp (−𝜃23 𝐹2 (𝑦2 − 1)) − 1)(exp (−𝜃23 𝐹3 (𝑦3 − 1)) − 1)

= 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 − 1)) = −1⁄𝜃 𝑙𝑜𝑔 (1 +
𝐹𝑟 −−
𝐶23 )
23 exp(−𝜃23 ) − 1
−+ −− +− −−
𝐶 𝐴𝑀𝐻 − 𝐶12 𝐴𝑀𝐻 𝐹𝑟
𝐶23 − 𝐶23𝐹𝑟
(exp (−𝜃13|2 ( 12 )) − 1) (exp (−𝜃13|2 ( )) − 1)
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
= −1⁄𝜃
𝐹𝑟 11
𝐶13|2 𝑙𝑜𝑔 1 +
13|2 exp(−𝜃13|2 ) − 1
( )
Hence,
𝐹𝑟 𝐹𝑟 00 𝐹𝑟 01
𝐹𝑟 10 11
𝑃(𝑌1 = 𝑦1 , 𝑌2 = 𝑦2 , 𝑌3 = 𝑦3 ) = (𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)]
As a result, the likelihood function is
𝑛 𝐽−1 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗
𝑇
𝐹𝑟 𝐹𝑟 𝐹𝑟 00 𝐹𝑟 01 10 11
L(𝑐, 𝛽, 𝜃|𝑋) = ∏ ∏ ∏[(𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)]]
𝑖=1 𝑡=1 𝑗=1

The log-likelihood is
𝑛 𝑇 𝐽−1
𝐹𝑟 𝐹𝑟 𝐹𝑟 00 𝐹𝑟 01 10 11
l(𝑐, 𝛽, 𝜃|𝑋) = ∑ ∑ ∑ 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗 ∗ 𝑙𝑜𝑔 ((𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)])
𝑖=1 𝑡=1 𝑗=1

𝑛 𝑇 𝐽−1
𝐹𝑟 𝐹𝑟 𝐹𝑟 00 𝐹𝑟 01 10 11
l(𝑐, 𝛽, 𝜃|𝑋) = ∑ ∑ ∑ 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗 ∗ {𝑙𝑜𝑔(𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 ) + 𝑙𝑜𝑔( 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1))}
𝑖=1 𝑡=1 𝑗=1

139 | P a g e
I. The steps possessed to simplify and obtain the log-likelihood maximum likelihood for the D- vine longitudinal discrete random
variables application to household food security.
The joint probability distribution of longitudinal ordinal data based on the D pair copula was displayed as follows.
Pr(Y1 = y1 , Y2 = y 2 , Y3 = y 3 )
C12 ( F1 ( y1 − i1 ), F2 ( y 2 )) − C12 ( F1 ( y1 − i1 ), F2 ( y 2 − 1)) C 23 ( F2 ( y 2 ), F3 ( y 3 − i3 )) − C 23 ( F2 ( y 2 − 1), F3 ( y 3 − i3 ))
={   (−1) i1 + i3
C13|2 ( , )}[F2 ( y 2 ) − F2 ( y 2 − 1)].
i1 = 0 , 1 i3 = 0 , 1 F2 ( y 2 ) − F2 ( y 2 − 1) F2 ( y 2 ) − F2 ( y 2 − 1)
Among the six proposed bivariate copula functions, Gumbel copula was selected for F1 ( y1 ) and F2 ( y2 ) , Gumbel for F2 ( y2 ) and
F3 ( y3 ) , and Gumbel for F1|2 ( y1 | y2 ) and F3|2 ( y3 | y2 ) . Hence the simplified joint probability distribution given by

when i1 = 0 and i3 = 0 ,
1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢++
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 ))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) ) ]

1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢+−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 ))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) ) ]

1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢++
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 ))) ) ]

1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢−+
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 ))) ) ]

1⁄
𝜃13|2 𝜃13|2 𝜃13|2
𝐺𝑢++ 𝐺𝑢+− 𝐺𝑢++ 𝐺𝑢−+
𝐺𝑢 00 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 = 𝑒𝑥𝑝 − ((−𝑙𝑜𝑔 ( )) + (−𝑙𝑜𝑔 ( )) )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
[ ]

140 | P a g e
when i1 = 1 and i3 = 0
1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢−+
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 − 1))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) ) ]

1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢−−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 − 1))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) ) ]

1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢++
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 ))) ) ]

1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢−+
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 ))) ) ]

1⁄
𝜃13|2 𝜃13|2 𝜃13|2
𝐺𝑢−+ 𝐺𝑢−− 𝐺𝑢++ 𝐺𝑢−+
𝐺𝑢 10 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 = 𝑒𝑥𝑝 − ((−𝑙𝑜𝑔 ( )) + (−𝑙𝑜𝑔 ( )) )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
[ ]
when i1 = 0 and i3 = 1
1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢++
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 ))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) ) ]

1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢+−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 ))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) ) ]

1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢+−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 − 1))) ) ]

141 | P a g e
 1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢−−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 − 1))) ) ]

1⁄
𝜃13|2 𝜃13|2 𝜃13|2
𝐺𝑢++ 𝐺𝑢+− 𝐺𝑢+− 𝐺𝑢−−
𝐺𝑢01 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 = 𝑒𝑥𝑝 − ((−𝑙𝑜𝑔 ( )) + (−𝑙𝑜𝑔 ( )) )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
[ ]
when i1 = 1 and i3 = 1 ,
1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢−+
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 )) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 − 1))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) ) ]

1⁄
𝜃12 𝜃12 𝜃12
𝐺𝑢−−
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹1 (𝑦1 − 1))) + (−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) ) ]

1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢+−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 ))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 − 1))) ) ]

1⁄
𝜃23 𝜃23 𝜃23
𝐺𝑢−−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 − 1)) = 𝑒𝑥𝑝 [− ((−𝑙𝑜𝑔(𝐹2 (𝑦2 − 1))) + (−𝑙𝑜𝑔(𝐹3 (𝑦3 − 1))) ) ]

1⁄
𝜃13|2 𝜃13|2 𝜃13|2
𝐺𝑢 −+ 𝐺𝑢 −− 𝐺𝑢 +−
𝐺𝑢 −−
𝐺𝑢11 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 = 𝑒𝑥𝑝 − ((−𝑙𝑜𝑔 ( )) + (−𝑙𝑜𝑔 ( )) )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
[ ]
Hence,
𝐺𝑢 00 𝐺𝑢 01 𝐺𝑢 10
𝐺𝑢 11
𝑃(𝑌1 = 𝑦1 , 𝑌2 = 𝑦2 , 𝑌3 = 𝑦3 ) = (𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)]
As a result, the likelihood function is

142 | P a g e
 𝑛 𝐽−1 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗
𝑇
𝐺𝑢00 𝐺𝑢01 𝐺𝑢10 𝐺𝑢11
L(𝑐, 𝛽, 𝜃|𝑋) = ∏ ∏ ∏[(𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)]]
𝑖=1 𝑡=1 𝑗=1

The log-likelihood is
𝑛 𝑇 𝐽−1
𝐺𝑢 𝐺𝑢 𝐺𝑢 𝐺𝑢 00 01 10 11
l(𝑐, 𝛽, 𝜃|𝑋) = ∑ ∑ ∑ 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗 ∗ 𝑙𝑜𝑔 ((𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)])
𝑖=1 𝑡=1 𝑗=1

𝑛 𝑇 𝐽−1
𝐺𝑢 𝐺𝑢 𝐺𝑢 𝐺𝑢 00 01 10 11
l(𝑐, 𝛽, 𝜃|𝑋) = ∑ ∑ ∑ 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗 ∗ {𝑙𝑜𝑔(𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 ) + 𝑙𝑜𝑔( 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1))}
𝑖=1 𝑡=1 𝑗=1

II. The steps possessed to simplify and obtain the log-likelihood maximum likelihood for the D- vine multivariate longitudinal
discrete random variables application to household food security.
M-dimensional multivariate discrete random variables were observed repeatedly for 𝑇 time points and re-ordered the observations of
the multivariate series into the univariate outcomes of dimensions 𝑁 = 𝑇 ∗ 𝑛 given by 𝑌 = (𝑌1 , 𝑌2 , … , 𝑌𝑀 ), where 𝑌1 =
(𝑦11 , 𝑦21 , … , 𝑦𝑁1 )′ , 𝑌2 = (𝑦12 , 𝑦22 , … , 𝑦𝑁2 )′ 𝑎𝑛𝑑 𝑌𝑀 = (𝑦1𝑚 , 𝑦2𝑚 , . , 𝑦𝑁𝑚 )′ . The joint probability distribution of multivariate
longitudinal ordinal data based on the D pair copula was displayed as follows.
Pr(Y1 = y1 , Y2 = y 2 , Y3 = y 3 )
C12 ( F1 ( y1 − i1 ), F2 ( y 2 )) − C12 ( F1 ( y1 − i1 ), F2 ( y 2 − 1)) C 23 ( F2 ( y 2 ), F3 ( y 3 − i3 )) − C 23 ( F2 ( y 2 − 1), F3 ( y 3 − i3 ))
={   (−1) i1 + i3
C13|2 ( , )}[F2 ( y 2 ) − F2 ( y 2 − 1)].
i1 = 0 , 1 i3 = 0 , 1 F2 ( y 2 ) − F2 ( y 2 − 1) F2 ( y 2 ) − F2 ( y 2 − 1)
Among the six proposed bivariate copula functions, Clayton (Cl) copula was selected for F1 ( y1 ) and F2 ( y2 ) , AMH for F2 ( y2 ) and
F3 ( y3 ) , and Independent for F1|2 ( y1 | y2 ) and F3|2 ( y3 | y2 ) . Hence the simplified joint probability distribution given by

when i1 = 0 and i3 = 0 ,

++ −𝜃12 −𝜃12 −1⁄𝜃

𝐶𝑙 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 )) = ((𝐹1 (𝑦1 )) + (𝐹2 (𝑦2 )) − 1)

143 | P a g e
 −𝜃12 −𝜃12 −1⁄𝜃
𝐶𝑙+− 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 − 1)) = ((𝐹1 (𝑦1 )) + (𝐹2 (𝑦2 − 1)) − 1)
𝐴𝑀𝐻 ++
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 )) = 𝐹2 (𝑦2 ) ∗ 𝐹2 (𝑦3 )/(1−𝜃23 (1 − 𝐹2 (𝑦2 )) ∗ (1 − 𝐹3 (𝑦3 )))
𝐴𝑀𝐻 −+
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 )) = 𝐹2 (𝑦2 − 1) ∗ 𝐹2 (𝑦3 )/(1−𝜃23 (1 − 𝐹2 (𝑦2 − 1)) ∗ (1 − 𝐹3 (𝑦3 )))
𝐶𝑙 ++ 𝐶𝑙 +− 𝐴𝑀𝐻 ++
𝐴𝑀𝐻 −+
𝐼𝑛𝑑00 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 =( )( )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
when i1 = 1 and i3 = 0

−+ −𝜃12 −𝜃12 −1⁄𝜃

𝐶𝑙 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 )) = ((𝐹1 (𝑦1 − 1)) + (𝐹2 (𝑦2 )) − 1)

−𝜃12 −𝜃12 −1⁄𝜃

𝐶𝑙−− 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 − 1)) = ((𝐹1 (𝑦1 − 1)) + (𝐹2 (𝑦2 − 1)) − 1)
𝐴𝑀𝐻 ++
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 )) = 𝐹2 (𝑦2 ) ∗ 𝐹2 (𝑦3 )/(1−𝜃23 (1 − 𝐹2 (𝑦2 )) ∗ (1 − 𝐹3 (𝑦3 )))
𝐴𝑀𝐻 −+
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 )) = 𝐹2 (𝑦2 − 1) ∗ 𝐹2 (𝑦3 )/(1−𝜃23 (1 − 𝐹2 (𝑦2 − 1)) ∗ (1 − 𝐹3 (𝑦3 )))
𝐶𝑙 −+ 𝐶𝑙 −− 𝐴𝑀𝐻 ++
𝐴𝑀𝐻 −+
𝐼𝑛𝑑10 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 =( )( )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
when i1 = 0 and i3 = 1
++ −𝜃12 −𝜃12 −1⁄𝜃
𝐶𝑙 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 )) = ((𝐹1 (𝑦1 )) + (𝐹2 (𝑦2 )) − 1)
−𝜃12 −𝜃12 −1⁄𝜃
𝐶𝑙+− 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 ), 𝐹2 (𝑦2 − 1)) = ((𝐹1 (𝑦1 )) + (𝐹2 (𝑦2 − 1)) − 1)
𝐴𝑀𝐻 +−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 − 1)) = 𝐹2 (𝑦2 ) ∗ 𝐹2 (𝑦3 − 1)/(1−𝜃23 (1 − 𝐹2 (𝑦2 )) ∗ (1 − 𝐹3 (𝑦3 − 1)))
𝐴𝑀𝐻 −−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 − 1)) = 𝐹2 (𝑦2 − 1) ∗ 𝐹2 (𝑦3 − 1)/(1−𝜃23 (1 − 𝐹2 (𝑦2 − 1)) ∗ (1 − 𝐹3 (𝑦3 − 1)))
𝐶𝑙 ++ 𝐶𝑙 +− 𝐴𝑀𝐻 +−
𝐴𝑀𝐻 −−
𝐼𝑛𝑑01 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 =( )( )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)

144 | P a g e
when i1 = 1 and i3 = 1 ,
−𝜃12 −𝜃12 −1⁄𝜃
𝐶𝑙−+ 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 )) = ((𝐹1 (𝑦1 − 1)) + (𝐹2 (𝑦2 )) − 1)
−− −𝜃12 −𝜃12 −1⁄𝜃
𝐶𝑙 12
𝐶12 = 𝐶12 (𝐹1 (𝑦1 − 1), 𝐹2 (𝑦2 − 1)) = ((𝐹1 (𝑦1 − 1)) + (𝐹2 (𝑦2 − 1)) − 1)
𝐴𝑀𝐻 +−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 ), 𝐹3 (𝑦3 − 1)) = 𝐹2 (𝑦2 ) ∗ 𝐹2 (𝑦3 − 1)/(1−𝜃23 (1 − 𝐹2 (𝑦2 )) ∗ (1 − 𝐹3 (𝑦3 − 1)))
𝐴𝑀𝐻 −−
𝐶23 = 𝐶23 (𝐹2 (𝑦2 − 1), 𝐹3 (𝑦3 − 1)) = 𝐹2 (𝑦2 − 1) ∗ 𝐹2 (𝑦3 − 1)/(1−𝜃23 (1 − 𝐹2 (𝑦2 − 1)) ∗ (1 − 𝐹3 (𝑦3 − 1)))
𝐶𝑙 −+ 𝐶𝑙 −−𝐴𝑀𝐻 𝐴𝑀𝐻 +− −−
𝐼𝑛𝑑11 𝐶12 − 𝐶12 𝐶23 − 𝐶23
𝐶13|2 =( )( )
𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1) 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)
Hence,
𝐼𝑛𝑑 𝐼𝑛𝑑 𝐼𝑛𝑑 00 𝐼𝑛𝑑 01 10 11
𝑃(𝑌1 = 𝑦1 , 𝑌2 = 𝑦2 , 𝑌3 = 𝑦3 ) = (𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)]
As a result, the likelihood function is
𝑛 𝐽−1 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗
𝑇
𝐼𝑛𝑑00 𝐼𝑛𝑑01 𝐼𝑛𝑑10 𝐼𝑛𝑑11
L(𝑐, 𝛽, 𝜃|𝑋) = ∏ ∏ ∏[(𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)]]
𝑖=1 𝑡=1 𝑗=1
The log-likelihood is
𝑛 𝑇 𝐽−1
𝐼𝑛𝑑 𝐼𝑛𝑑 𝐼𝑛𝑑 𝐼𝑛𝑑 00 01 10 11
l(𝑐, 𝛽, 𝜃|𝑋) = ∑ ∑ ∑ 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗 ∗ 𝑙𝑜𝑔 ((𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 )[ 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1)])
𝑖=1 𝑡=1 𝑗=1
𝑛 𝑇 𝐽−1
𝐼𝑛𝑑 𝐼𝑛𝑑 𝐼𝑛𝑑 𝐼𝑛𝑑 00 01 10 11
l(𝑐, 𝛽, 𝜃|𝑋) = ∑ ∑ ∑ 𝑦1𝑖𝑗 𝑦2𝑖𝑗 𝑦3𝑖𝑗 ∗ {𝑙𝑜𝑔(𝐶13|2 − 𝐶13|2 − 𝐶13|2 + 𝐶13|2 ) + 𝑙𝑜𝑔( 𝐹2 (𝑦2 ) − 𝐹2 (𝑦2 − 1))}
𝑖=1 𝑡=1 𝑗=1

145 | P a g e
Appendix D: R codes for the log-likelihood D-vines

I. The R code to estimate the parameters for the simplified log-likelihood function in
Appendix B I.

# Import the Food security data and Pre-processing ##

### X are Covariates for All Dimensions ##
### Dep is Dependent variables ##

library(Alabama)

X<-[Link]([Link](), header=TRUE, sep=",")

Dep<-[Link]([Link](), header=TRUE, sep=",")
Y<-Dep$Y1
Z<-Dep$Y2
W<-Dep$Y3
y1<-ifelse(Y==1,1,0)
y2<-ifelse(Y==2,1,0)
y3<-ifelse(Y==3,1,0)
y4<-ifelse(Y==4,1,0)
z1<-ifelse(Z==1,1,0)
z2<-ifelse(Z==2,1,0)
z3<-ifelse(Z==3,1,0)
z4<-ifelse(Z==4,1,0)
w1<-ifelse(W==1,1,0)
w2<-ifelse(W==2,1,0)
w3<-ifelse(W==3,1,0)
w4<-ifelse(W==4,1,0)
X=[Link](X)
G=function(z)
{
G=exp(z)/(1+exp(z))

147 | P a g e
return(G)
}

g=function(z)
{
g=exp(z)/(1+exp(z))^2
return(g)
}
### Cumulative logit multivariate Ordinal longitudinal Model ###
alpha1<-vector(length=3,mode="numeric")
alpha1[1]<-0
alpha2<-vector(length=3,mode="numeric")
alpha2[1]<-0
alpha3<-vector(length=3,mode="numeric")
alpha3[1]<-0
beta<-vector(length=16,mode="numeric") ### coefficients for Availability covariates ###
gamma<-vector(length=16,mode="numeric") ### coefficients for Accessibility covariates ###
zeta<-vector(length=6,mode="numeric") ### coefficients for Utilisation covariates ###
r1<-vector(length=1,mode="numeric") ### PCC parameter for Availability and Accessibility
###
r2<-vector(length=1,mode="numeric") ### PCC parameter for Accessibility and Utilisation
###
r3<-vector(length=1,mode="numeric") ### PCC parameter for Availability|Accessibility and
Utilisation|Accessibility ###
par<-vector(length=60,mode="numeric")

[Link] <-function(par)
{
comp1<-comp2<-comp3<-comp4<-vector(length=dim(X)[1],mode="numeric")
z10<-z11<-z12<-z13<-z14<-z20<-z21<-z22<-z23<-z24<-z30<-z31<-z32<-z33<-z34<-
vector(length=dim(X)[1],mode="numeric")

148 | P a g e
AMH_11<-AMH_21<-AMH_31<-AMH_41<-AMH_12<-AMH_22<-AMH_32<-AMH_42<-
AMH_13<-AMH_23<-AMH_33<-AMH_43<-AMH_14<-AMH_24<-AMH_34<-AMH_44<-
vector(length=dim(X)[1],mode="numeric")
Fr_001<-Fr_101<-Fr_011<-Fr_111<- Fr_002<-Fr_102<-Fr_012<-Fr_112<-Fr_003<-Fr_103<-
Fr_013<-Fr_113<-Fr_004<-Fr_104<-Fr_014<-Fr_114<-
vector(length=dim(X)[1],mode="numeric")
Fr_11<-Fr_21<-Fr_31<-Fr_41<-Fr_12<-Fr_22<-Fr_32<-Fr_42<-Fr_13<-Fr_23<-Fr_33<-
Fr_43<-Fr_14<-Fr_24<-Fr_34<-Fr_44<-vector(length=dim(X)[1],mode="numeric")
f1<-f2<-f3<-f4<-vector(length=dim(X)[1],mode="numeric")
alpha1<-par[1:3]; alpha2<-par[4:6]; alpha3<-par[7:9]; beta<-par[10:25]; gamma<-par[26:41];
zeta<-par[42:57]; r1<-par[58]; r2<-par[59]; r3<-par[60]
for (ii in 1: dim(X)[1])
{
z10[ii]<--Inf+sum(beta*X[ii,])
z11[ii]<-alpha1[1]+sum(beta*X[ii,])
z12[ii]<-alpha1[2]+sum(beta*X[ii,])
z13[ii]<-alpha1[3]+sum(beta*X[ii,])
z14[ii]<-100+sum(beta*X[ii,])
z20[ii]<--Inf+sum(gamma*X[ii,])
z21[ii]<-alpha2[1]+sum(gamma*X[ii,])
z22[ii]<-alpha2[2]+sum(gamma*X[ii,])
z23[ii]<-alpha2[3]+sum(gamma*X[ii,])
z24[ii]<-100+sum(gamma*X[ii,])
z30[ii]<--Inf+sum(zeta*X[ii,])
z31[ii]<-alpha3[1]+sum(zeta*X[ii,])
z32[ii]<-alpha3[2]+sum(zeta*X[ii,])
z33[ii]<-alpha3[3]+sum(zeta*X[ii,])
z34[ii]<-100+sum(zeta*X[ii,])
}

d<-[Link](z10, z11, z12, z13, z14)

149 | P a g e
z<-[Link](z20, z21, z22, z23, z24)
w<-[Link](z30, z31, z32, z33, z34)

AHM_11<-(G(d[,2])*G(z[,2]))/(1-r1*(1-G(d[,2]))*(1-G(z[,2])))
AHM_21<-(G(d[,2])*G(z[,1]))/(1-r1*(1-G(d[,2]))*(1-G(z[,1])))
AHM_31<-(G(d[,1])*G(z[,2]))/(1-r1*(1-G(d[,1]))*(1-G(z[,2])))
AHM_41<-(G(d[,1])*G(z[,1]))/(1-r1*(1-G(d[,1]))*(1-G(z[,1])))

Fr_11<-(-1/r2)*log(1+((exp(-r2*G(z[,2]))-1)*(exp(-r2*G(w[,2]))-1)/(exp(-r2)-1)))
Fr_21<-(-1/r2)*log(1+((exp(-r2*G(z[,1]))-1)*(exp(-r2*G(w[,2]))-1)/(exp(-r2)-1)))
Fr_31<-(-1/r2)*log(1+((exp(-r2*G(z[,2]))-1)*(exp(-r2*G(w[,1]))-1)/(exp(-r2)-1)))
Fr_41<-(-1/r2)*log(1+((exp(-r2*G(z[,1]))-1)*(exp(-r2*G(w[,1]))-1)/(exp(-r2)-1)))

f2<-G(z[,2])-G(z[,1])

Fr_001<-(-1/r3)*log(1+((exp(-r3*(AHM_11 - AHM_21)/f2)-1)*(exp(-r3*(Fr_11 - Fr_21)/f2)-

1)/(exp(-r3)-1)))
Fr_101<-(-1/r3)*log(1+((exp(-r3*(AHM_31 - AHM_41)/f2)-1)*(exp(-r3*(Fr_11 - Fr_21)/f2)-
1)/(exp(-r3)-1)))
Fr_011<-(-1/r3)*log(1+((exp(-r3*(AHM_11 - AHM_21)/f2)-1)*(exp(-r3*(Fr_31 - Fr_41)/f2)-
1)/(exp(-r3)-1)))
Fr_111<-(-1/r3)*log(1+((exp(-r3*(AHM_31 - AHM_41)/f2)-1)*(exp(-r3*(Fr_31 - Fr_41)/f2)-
1)/(exp(-r3)-1)))

comp1<-y1*z1*w1*log((Fr_001 - Fr_101 - Fr_011 + Fr_111)*f2)

AHM_12<-(G(d[,3])*G(z[,3]))/(1-r1*(1-G(d[,3]))*(1-G(z[,3])))
AHM_22<-(G(d[,3])*G(z[,2]))/(1-r1*(1-G(d[,3]))*(1-G(z[,2])))
AHM_32<-(G(d[,2])*G(z[,3]))/(1-r1*(1-G(d[,2]))*(1-G(z[,3])))
AHM_42<-(G(d[,2])*G(z[,2]))/(1-r1*(1-G(d[,2]))*(1-G(z[,2])))

150 | P a g e
Fr_12<-(-1/r2)*log(1+((exp(-r2*G(z[,3]))-1)*(exp(-r2*G(w[,3]))-1)/(exp(-r2)-1)))
Fr_22<-(-1/r2)*log(1+((exp(-r2*G(z[,2]))-1)*(exp(-r2*G(w[,3]))-1)/(exp(-r2)-1)))
Fr_32<-(-1/r2)*log(1+((exp(-r2*G(z[,3]))-1)*(exp(-r2*G(w[,2]))-1)/(exp(-r2)-1)))
Fr_42<-(-1/r2)*log(1+((exp(-r2*G(z[,2]))-1)*(exp(-r2*G(w[,2]))-1)/(exp(-r2)-1)))

f3<-G(z[,3])-G(z[,2])

Fr_002<-(-1/r3)*log(1+((exp(-r3*(AHM_12 - AHM_22)/f3)-1)*(exp(-r3*(Fr_12 - Fr_22)/f3)-

1)/(exp(-r3)-1)))
Fr_102<-(-1/r3)*log(1+((exp(-r3*(AHM_32 - AHM_42)/f3)-1)*(exp(-r3*(Fr_12 - Fr_22)/f3)-
1)/(exp(-r3)-1)))
Fr_012<-(-1/r3)*log(1+((exp(-r3*(AHM_12 - AHM_22)/f3)-1)*(exp(-r3*(Fr_32 - Fr_42)/f3)-
1)/(exp(-r3)-1)))
Fr_112<-(-1/r3)*log(1+((exp(-r3*(AHM_32 - AHM_42)/f3)-1)*(exp(-r3*(Fr_32 - Fr_42)/f3)-
1)/(exp(-r3)-1)))

comp2<-y2*z2*w2*log((Fr_002 - Fr_102 - Fr_012 + Fr_112)*f3)

AHM_13<-(G(d[,4])*G(z[,4]))/(1-r1*(1-G(d[,4]))*(1-G(z[,4])))
AHM_23<-(G(d[,4])*G(z[,3]))/(1-r1*(1-G(d[,4]))*(1-G(z[,3])))
AHM_33<-(G(d[,3])*G(z[,4]))/(1-r1*(1-G(d[,3]))*(1-G(z[,4])))
AHM_43<-(G(d[,3])*G(z[,3]))/(1-r1*(1-G(d[,3]))*(1-G(z[,3])))

Fr_13<-(-1/r2)*log(1+((exp(-r2*G(z[,4]))-1)*(exp(-r2*G(w[,4]))-1)/(exp(-r2)-1)))
Fr_23<-(-1/r2)*log(1+((exp(-r2*G(z[,3]))-1)*(exp(-r2*G(w[,4]))-1)/(exp(-r2)-1)))
Fr_33<-(-1/r2)*log(1+((exp(-r2*G(z[,4]))-1)*(exp(-r2*G(w[,3]))-1)/(exp(-r2)-1)))
Fr_43<-(-1/r2)*log(1+((exp(-r2*G(z[,3]))-1)*(exp(-r2*G(w[,3]))-1)/(exp(-r2)-1)))

f4<-G(z[,4])-G(z[,3])

151 | P a g e
Fr_003<-(-1/r3)*log(1+((exp(-r3*(AHM_13 - AHM_23)/f4)-1)*(exp(-r3*(Fr_13 - Fr_23)/f4)-
1)/(exp(-r3)-1)))
Fr_103<-(-1/r3)*log(1+((exp(-r3*(AHM_33 - AHM_43)/f4)-1)*(exp(-r3*(Fr_13 - Fr_23)/f4)-
1)/(exp(-r3)-1)))
Fr_013<-(-1/r3)*log(1+((exp(-r3*(AHM_13 - AHM_23)/f4)-1)*(exp(-r3*(Fr_33 - Fr_43)/f4)-
1)/(exp(-r3)-1)))
Fr_113<-(-1/r3)*log(1+((exp(-r3*(AHM_33 - AHM_43)/f4)-1)*(exp(-r3*(Fr_33 - Fr_43)/f4)-
1)/(exp(-r3)-1)))

comp3<-y3*z3*w3*log((Fr_003 - Fr_103 - Fr_013 + Fr_113)*f4)

AHM_14<-(G(d[,5])*G(z[,5]))/(1-r1*(1-G(d[,5]))*(1-G(z[,5])))
AHM_24<-(G(d[,5])*G(z[,4]))/(1-r1*(1-G(d[,5]))*(1-G(z[,4])))
AHM_34<-(G(d[,4])*G(z[,5]))/(1-r1*(1-G(d[,4]))*(1-G(z[,5])))
AHM_44<-(G(d[,4])*G(z[,4]))/(1-r1*(1-G(d[,4]))*(1-G(z[,4])))

Fr_14<-(-1/r2)*log(1+((exp(-r2*G(z[,5]))-1)*(exp(-r2*G(w[,5]))-1)/(exp(-r2)-1)))
Fr_24<-(-1/r2)*log(1+((exp(-r2*G(z[,4]))-1)*(exp(-r2*G(w[,5]))-1)/(exp(-r2)-1)))
Fr_34<-(-1/r2)*log(1+((exp(-r2*G(z[,5]))-1)*(exp(-r2*G(w[,4]))-1)/(exp(-r2)-1)))
Fr_44<-(-1/r2)*log(1+((exp(-r2*G(z[,4]))-1)*(exp(-r2*G(w[,4]))-1)/(exp(-r2)-1)))

f5<-G(z[,5])-G(z[,4])

Fr_004<-(-1/r3)*log(1+((exp(-r3*(AHM_14 - AHM_24)/f5)-1)*(exp(-r3*(Fr_14 - Fr_24)/f5)-

1)/(exp(-r3)-1)))
Fr_104<-(-1/r3)*log(1+((exp(-r3*(AHM_34 - AHM_44)/f5)-1)*(exp(-r3*(Fr_14 - Fr_24)/f5)-
1)/(exp(-r3)-1)))
Fr_014<-(-1/r3)*log(1+((exp(-r3*(AHM_14 - AHM_24)/f5)-1)*(exp(-r3*(Fr_34 - Fr_44)/f5)-
1)/(exp(-r3)-1)))
Fr_114<-(-1/r3)*log(1+((exp(-r3*(AHM_34 - AHM_44)/f5)-1)*(exp(-r3*(Fr_34 - Fr_44)/f5)-
1)/(exp(-r3)-1)))

152 | P a g e
comp4<-y4*z4*w4*log((Fr_004 - Fr_104 - Fr_014 + Fr_114)*f5)

-sum(comp1+comp2+comp3+comp4)
}
hin<-function(par)
{
alpha<-par[1:9]
h<-rep(NA,1)
h[1]<-alpha1[2]-alpha[1]
h[2]<-alpha1[3]-alpha1[2]
h[3]<-alpha2[2]-alpha2[1]
h[4]<-alpha2[3]-alpha2[2]
h[5]<-alpha3[2]-alpha3[1]
h[6]<-alpha3[3]-alpha3[2]
h
}
[Link]<-function(par){
alpha<-par[1:9]
j<-matrix(NA,6, length(par))
j[1,]<-
c(1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,0,0,0,0,0)
j[2,]<-
c(0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,0,0,0,0,0)
j[3,]<-
c(0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,0,0,0,0,0)

153 | P a g e
j[4,]<-
c(0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,0,0,0,0,0)
j[5,]<-
c(0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,0,0,0,0,0)
j[6,]<-
c(0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
,0,0,0,0,0,0,0,0,0)
j
}
par<-c(-4.331,-1.526,0.834,-4.685,-2.885,-.806,-1.526,2.243,4.140,-0.514,-0.698,-0.192,-1.09,-
0.231,-1.258,-0.445,0.432,-0.492,-0.503,0.85,0.586,1.65,-0.578,-1.622,-1.191,-.865,-.751,-
.242,.647,.185,-.223,.058,-.050,-.086,-.267,-1.586,.388,.159,.132,-.920,-.589,-1.065,-.908,-.111,-
.544,-.184,1.117,.910,.349,-.582,.066,1.542,.492,.290,.010,.012,.169, 0.5, 1.5, 1.2)
[Link]<- auglag(par, [Link] , hin = hin, [Link] = [Link])
s.e<- sqrt(diag(solve([Link]$hessian)))

II. The R code to estimate the parameters for the simplified log-likelihood function in
Appendix B II.

library(alabama)

X<-[Link]([Link](), header=TRUE, sep=",")

Dep<-[Link]([Link](), header=TRUE, sep=",")
Y<-Dep$Y1
Z<-Dep$Y2
W<-Dep$Y3

y1<-ifelse(Y==1,1,0)
y2<-ifelse(Y==2,1,0)
y3<-ifelse(Y==3,1,0)
y4<-ifelse(Y==4,1,0)

154 | P a g e
z1<-ifelse(Z==1,1,0)
z2<-ifelse(Z==2,1,0)
z3<-ifelse(Z==3,1,0)
z4<-ifelse(Z==4,1,0)
w1<-ifelse(W==1,1,0)
w2<-ifelse(W==2,1,0)
w3<-ifelse(W==3,1,0)
w4<-ifelse(W==4,1,0)

X=[Link](X)

G=function(z)
{
G=exp(z)/(1+exp(z))
return(G)
}

g=function(z)
{
g=exp(z)/(1+exp(z))^2
return(g)
}

### Cumulative logit Ordinal Model ###

alpha1<-vector(length=3,mode="numeric")
alpha1[1]<-0
alpha2<-vector(length=3,mode="numeric")
alpha2[1]<-0
alpha3<-vector(length=3,mode="numeric")
alpha3[1]<-0

155 | P a g e
beta<-vector(length=7,mode="numeric")
gamma<-vector(length=7,mode="numeric")
zeta<-vector(length=7,mode="numeric")
r1<-vector(length=1,mode="numeric")
r2<-vector(length=1,mode="numeric")
r3<-vector(length=1,mode="numeric")

par<-vector(length=33,mode="numeric")

[Link] <-function(par)
{
comp1<-comp2<-comp3<-comp4<-vector(length=dim(X)[1],mode="numeric")
z10<-z11<-z12<-z13<-z14<-z20<-z21<-z22<-z23<-z24<-z30<-z31<-z32<-z33<-z34<-
vector(length=dim(X)[1],mode="numeric")

Gu_11<-Ga_21<-Ga_31<-Ga_41<-Ga_12<-Ga_22<-Ga_32<-Ga_42<-Ga_13<-Ga_23<-
Ga_33<-Ga_43<-Ga_14<-Ga_24<-Ga_34<-Ga_44<-vector(length=dim(X)[1],mode="numeric")
Gu_11<-Gu_21<-Gu_31<-Gu_41<-Gu_12<-Gu_22<-Gu_32<-Gu_42<-Gu_13<-Gu_23<-
Gu_33<-Gu_43<-Gu_14<-Gu_24<-Gu_34<-Gu_44<-
vector(length=dim(X)[1],mode="numeric")
Ga_001<-Ga_011<-Ga_101<-Ga_111<-Ga_002<-Ga_012<-Ga_102<-Ga_112<-Ga_003<-
Ga_013<-Ga_103<-Ga_113<-Ga_004<-Ga_014<-Ga_104<-Ga_114<-
vector(length=dim(X)[1],mode="numeric")
Gu_001<-Gu_011<-Gu_101<-Gu_111<-Gu_002<-Gu_012<-Gu_102<-Gu_112<-Gu_003<-
Gu_013<-Gu_103<-Gu_113<-Gu_004<-Gu_014<-Gu_104<-Gu_114<-
vector(length=dim(X)[1],mode="numeric")

f1<-f2<-f3<-f4<-vector(length=dim(X)[1],mode="numeric")
alpha1<-par[1:3]; alpha2<-par[4:6]; alpha3<-par[7:9]; beta<-par[10:16]; gamma<-par[17:23];
zeta<-par[24:30]; r1<-par[31]; r2<-par[32]; r3<-par[33]

156 | P a g e
for (ii in 1: dim(X)[1])
{
z10[ii]<--Inf+sum(beta*X[ii,])
z11[ii]<-alpha1[1]+sum(beta*X[ii,])
z12[ii]<-alpha1[2]+sum(beta*X[ii,])
z13[ii]<-alpha1[3]+sum(beta*X[ii,])
z14[ii]<-100+sum(beta*X[ii,])

z20[ii]<--Inf+sum(gamma*X[ii,])
z21[ii]<-alpha2[1]+sum(gamma*X[ii,])
z22[ii]<-alpha2[2]+sum(gamma*X[ii,])
z23[ii]<-alpha2[3]+sum(gamma*X[ii,])
z24[ii]<-100+sum(gamma*X[ii,])

z30[ii]<--Inf+sum(zeta*X[ii,])
z31[ii]<-alpha3[1]+sum(zeta*X[ii,])
z32[ii]<-alpha3[2]+sum(zeta*X[ii,])
z33[ii]<-alpha3[3]+sum(zeta*X[ii,])
z34[ii]<-100+sum(zeta*X[ii,])

}
d<-[Link](z10, z11, z12, z13, z14)
z<-[Link](z20, z21, z22, z23, z24)
w<-[Link](z30, z31, z32, z33, z34)

Ga_11<-exp(-((-log(G(d[,2])))^r1 + (-log(G(z[,2])))^r1)^1/r1)
Ga_21<-exp(-((-log(G(d[,2])))^r1 + (-log(G(z[,1])))^r1)^1/r1)
Ga_31<-exp(-((-log(G(d[,1])))^r1 + (-log(G(z[,2])))^r1)^1/r1)
Ga_41<-exp(-((-log(G(d[,1])))^r1 + (-log(G(z[,1])))^r1)^1/r1)

Gu_11<-exp(-((-log(G(z[,2])))^r2 + (-log(G(w[,2])))^r2)^1/r2)

157 | P a g e
Gu_21<-exp(-((-log(G(z[,1])))^r2 + (-log(G(w[,2])))^r2)^1/r2)
Gu_31<-exp(-((-log(G(z[,2])))^r2 + (-log(G(w[,1])))^r2)^1/r2)
Gu_41<-exp(-((-log(G(z[,1])))^r2 + (-log(G(w[,1])))^r2)^1/r2)
f2<-G(z[,2])-G(z[,1])

Ga_001<-exp(-((-log((Ga_11 - Ga_21)/f2))^r3 + (-log((Gu_11 - Gu_21)/f2))^r3)^1/r3)

Ga_101<-exp(-((-log((Ga_11 - Ga_21)/f2))^r3 + (-log((Gu_31 - Gu_41)/f2))^r3)^1/r3)
Ga_011<-exp(-((-log((Ga_31 - Ga_41)/f2))^r3 + (-log((Gu_11 - Gu_21)/f2))^r3)^1/r3)
Ga_111<-exp(-((-log((Ga_31 - Ga_41)/f2))^r3 + (-log((Gu_31 - Gu_41)/f2))^r3)^1/r3)

comp1<-y1*z1*w1*log((Ga_001 - Ga_101 - Ga_011 + Ga_111)*f2)

Ga_12<-exp(-((-log(G(d[,3])))^r1 + (-log(G(z[,3])))^r1)^1/r1)
Ga_22<-exp(-((-log(G(d[,3])))^r1 + (-log(G(z[,2])))^r1)^1/r1)
Ga_32<-exp(-((-log(G(d[,2])))^r1 + (-log(G(z[,3])))^r1)^1/r1)
Ga_42<-exp(-((-log(G(d[,2])))^r1 + (-log(G(z[,2])))^r1)^1/r1)

Gu_12<-exp(-((-log(G(z[,3])))^r2 + (-log(G(w[,3])))^r2)^1/r2)
Gu_22<-exp(-((-log(G(z[,2])))^r2 + (-log(G(w[,3])))^r2)^1/r2)
Gu_32<-exp(-((-log(G(z[,3])))^r2 + (-log(G(w[,2])))^r2)^1/r2)
Gu_42<-exp(-((-log(G(z[,2])))^r2 + (-log(G(w[,2])))^r2)^1/r2)

f3<-G(z[,3])-G(z[,2])

Ga_002<-exp(-((-log((Ga_12 - Ga_22)/f3))^r3 + (-log((Gu_12 - Gu_22)/f3))^r3)^1/r3)

Ga_102<-exp(-((-log((Ga_12 - Ga_22)/f3))^r3 + (-log((Gu_32 - Gu_42)/f3))^r3)^1/r3)
Ga_012<-exp(-((-log((Ga_32 - Ga_42)/f3))^r3 + (-log((Gu_12 - Gu_22)/f3))^r3)^1/r3)
Ga_112<-exp(-((-log((Ga_32 - Ga_42)/f3))^r3 + (-log((Gu_32 - Gu_42)/f3))^r3)^1/r3)

comp2<-y2*z2*w2*log((Ga_002 - Ga_102 - Ga_012 + Ga_112)*f3)

158 | P a g e
Ga_13<-exp(-((-log(G(d[,4])))^r1 + (-log(G(z[,4])))^r1)^1/r1)
Ga_23<-exp(-((-log(G(d[,4])))^r1 + (-log(G(z[,3])))^r1)^1/r1)
Ga_33<-exp(-((-log(G(d[,3])))^r1 + (-log(G(z[,4])))^r1)^1/r1)
Ga_43<-exp(-((-log(G(d[,3])))^r1 + (-log(G(z[,3])))^r1)^1/r1)

Gu_13<-exp(-((-log(G(z[,4])))^r2 + (-log(G(w[,4])))^r2)^1/r2)
Gu_23<-exp(-((-log(G(z[,3])))^r2 + (-log(G(w[,4])))^r2)^1/r2)
Gu_33<-exp(-((-log(G(z[,4])))^r2 + (-log(G(w[,3])))^r2)^1/r2)
Gu_43<-exp(-((-log(G(z[,3])))^r2 + (-log(G(w[,3])))^r2)^1/r2)

f4<-G(z[,4])-G(z[,3])

Ga_003<-exp(-((-log((Ga_13 - Ga_23)/f4))^r3 + (-log((Gu_13 - Gu_23)/f4))^r3)^1/r3)

Ga_103<-exp(-((-log((Ga_13 - Ga_23)/f4))^r3 + (-log((Gu_33 - Gu_43)/f4))^r3)^1/r3)
Ga_013<-exp(-((-log((Ga_33 - Ga_43)/f4))^r3 + (-log((Gu_13 - Gu_23)/f4))^r3)^1/r3)
Ga_113<-exp(-((-log((Ga_33 - Ga_43)/f4))^r3 + (-log((Gu_33 - Gu_43)/f4))^r3)^1/r3)

comp3<-y3*z3*w3*log((Ga_003 - Ga_103 - Ga_013 + Ga_113)*f4)

Ga_14<-exp(-((-log(G(d[,5])))^r1 + (-log(G(z[,5])))^r1)^1/r1)
Ga_24<-exp(-((-log(G(d[,5])))^r1 + (-log(G(z[,4])))^r1)^1/r1)
Ga_34<-exp(-((-log(G(d[,4])))^r1 + (-log(G(z[,5])))^r1)^1/r1)
Ga_44<-exp(-((-log(G(d[,4])))^r1 + (-log(G(z[,4])))^r1)^1/r1)

Gu_14<-exp(-((-log(G(z[,5])))^r2 + (-log(G(w[,5])))^r2)^1/r2)
Gu_24<-exp(-((-log(G(z[,4])))^r2 + (-log(G(w[,5])))^r2)^1/r2)
Gu_34<-exp(-((-log(G(z[,5])))^r2 + (-log(G(w[,4])))^r2)^1/r2)
Gu_44<-exp(-((-log(G(z[,4])))^r2 + (-log(G(w[,4])))^r2)^1/r2)

f5<-G(z[,5])-G(z[,4])

159 | P a g e
Ga_004<-exp(-((-log((Ga_14 - Ga_24)/f5))^r3 + (-log((Gu_14 - Gu_24)/f5))^r3)^1/r3)
Ga_104<-exp(-((-log((Ga_14 - Ga_24)/f5))^r3 + (-log((Gu_34 - Gu_44)/f5))^r3)^1/r3)
Ga_014<-exp(-((-log((Ga_34 - Ga_44)/f5))^r3 + (-log((Gu_14 - Gu_24)/f5))^r3)^1/r3)
Ga_114<-exp(-((-log((Ga_34 - Ga_44)/f5))^r3 + (-log((Gu_34 - Gu_44)/f5))^r3)^1/r3)

comp4<-y4*z4*w4*log((Ga_004 - Ga_104 - Ga_014 + Ga_114)*f5)

-sum(comp1+comp2+comp3+comp4)
}
hin<-function(par)
{
alpha<-par[1:9]
h<-rep(NA,1)
h[1]<-alpha1[2]-alpha[1]
h[2]<-alpha1[3]-alpha1[2]
h[3]<-alpha2[2]-alpha2[1]
h[4]<-alpha2[3]-alpha2[2]
h[5]<-alpha3[2]-alpha3[1]
h[6]<-alpha3[3]-alpha3[2]
h}
[Link]<-function(par){
alpha<-par[1:9]
j<-matrix(NA,6, length(par))
j[1,]<-c(1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[2,]<-c(0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[3,]<-c(0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[4,]<-c(0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[5,]<-c(0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[6,]<-c(0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j
}

160 | P a g e
par<-c(-2.515, .104, 2.926, -2.157, .444, 3.298, -3.366, -.244, 2.113, .259, .990, -.222, .299, .905,
-.360, -.363, .258, 1.103, -.215, .405, .927, .003, -.186, .022, .896, -.577, -.456, -.717, .412, .984,
1.1, 1.1, 1.1)
[Link]<- auglag(par, [Link] , hin = hin, [Link] = [Link])
s.e<- sqrt(diag(solve([Link]$hessian)))

III. Code to estimate the parameters for the simplified log-likelihood function in Appendix B
III.

# Import the Food security data and Preprocessing ##

### X1 are Covariates for Availability Dimension ##
### X2 are Covariates for Accessibility Dimension ##
### X3 are Covariates for Utilisation Dimension ##
### Dep are Dependent variables ##

library(alabama)

X1<-[Link]([Link](), header=TRUE, sep=",")

X2<-[Link]([Link](), header=TRUE, sep=",")
X3<-[Link]([Link](), header=TRUE, sep=",")
Dep<-[Link]([Link](), header=TRUE, sep=",")
Y<-Dep$Y1
Z<-Dep$Y2
W<-Dep$Y3
y1<-ifelse(Y==1,1,0)
y2<-ifelse(Y==2,1,0)
y3<-ifelse(Y==3,1,0)
y4<-ifelse(Y==4,1,0)
z1<-ifelse(Z==1,1,0)
z2<-ifelse(Z==2,1,0)
z3<-ifelse(Z==3,1,0)

161 | P a g e
z4<-ifelse(Z==4,1,0)
w1<-ifelse(W==1,1,0)
w2<-ifelse(W==2,1,0)
w3<-ifelse(W==3,1,0)
w4<-ifelse(W==4,1,0)
X1=[Link](X1)
X2=[Link](X2)
X3=[Link](X3)
G=function(z)
{
G=exp(z)/(1+exp(z))
return(G)
}

g=function(z)
{
g=exp(z)/(1+exp(z))^2
return(g)
}
### Cumulative logit multivariate Ordinal longitudinal Model ###
alpha1<-vector(length=3,mode="numeric")
alpha1[1]<-0
alpha2<-vector(length=3,mode="numeric")
alpha2[1]<-0
alpha3<-vector(length=3,mode="numeric")
alpha3[1]<-0
beta<-vector(length=13,mode="numeric") ### coefficients for Availability covariates ###
gamma<-vector(length=11,mode="numeric") ### coefficients for Accessibility covariates ###
zeta<-vector(length=9,mode="numeric") ### coefficients for Utilisation covariates ###
r1<-vector(length=1,mode="numeric") ### PCC parameter for Availability and Accessibility
###

162 | P a g e
r2<-vector(length=1,mode="numeric") ### PCC parameter for Accessibility and Utilisation
###
par<-vector(length=44,mode="numeric")

[Link] <-function(par)
{
comp1<-comp2<-comp3<-comp4<-vector(length=dim(X1)[1],mode="numeric")
z10<-z11<-z12<-z13<-z14<-z20<-z21<-z22<-z23<-z24<-z30<-z31<-z32<-z33<-z34<-
vector(length=dim(X1)[1],mode="numeric")
Cl_11<- Cl_21<- Cl_31<- Cl_41<-Cl_11<- Cl_21<- Cl_31<- Cl_41<-Cl_11<- Cl_21<- Cl_31<-
Cl_41<-Cl_11<- Cl_21<- Cl_31<- Cl_41<-vector(length=dim(X1)[1],mode="numeric")
AMH_11<-AMH_21<-AMH_31<-AMH_41<-AMH_12<-AMH_22<-AMH_32<-AMH_42<-
AMH_13<-AMH_23<-AMH_33<-AMH_43<-AMH_14<-AMH_24<-AMH_34<-AMH_44<-
vector(length=dim(X1)[1],mode="numeric")
IND_001<-IND_101<-IND_011<-IND_111<-IND_002<-IND_102<-IND_012<-IND_112<-
IND_003<-IND_103<-IND_013<-IND_113<-IND_004<-IND_104<-IND_014<-IND_114<-
vector(length=dim(X1)[1],mode="numeric")
f1<-f2<-f3<-f4<-vector(length=dim(X1)[1],mode="numeric")
alpha1<-par[1:3]; alpha2<-par[4:6]; alpha3<-par[7:9]; beta<-par[10:22]; gamma<-par[23:33];
zeta<-par[34:42]; r1<-par[44]; r2<-par[44]
for (ii in 1: dim(X1)[1])
{
z10[ii]<--Inf+sum(beta*X1[ii,])
z11[ii]<-alpha1[1]+sum(beta*X1[ii,])
z12[ii]<-alpha1[2]+sum(beta*X1[ii,])
z13[ii]<-alpha1[3]+sum(beta*X1[ii,])
z14[ii]<-100+sum(beta*X1[ii,])
z20[ii]<--Inf+sum(gamma*X2[ii,])
z21[ii]<-alpha2[1]+sum(gamma*X2[ii,])
z22[ii]<-alpha2[2]+sum(gamma*X2[ii,])
z23[ii]<-alpha2[3]+sum(gamma*X2[ii,])

163 | P a g e
z24[ii]<-100+sum(gamma*X2[ii,])
z30[ii]<--Inf+sum(zeta*X3[ii,])
z31[ii]<-alpha3[1]+sum(zeta*X3[ii,])
z32[ii]<-alpha3[2]+sum(zeta*X3[ii,])
z33[ii]<-alpha3[3]+sum(zeta*X3[ii,])
z34[ii]<-100+sum(zeta*X3[ii,])
}

d<-[Link](z10, z11, z12, z13, z14)

z<-[Link](z20, z21, z22, z23, z24)
w<-[Link](z30, z31, z32, z33, z34)

Cl_11<-((G(d[,2]))^(-r1) + (G(z[,2]))^(-r1) - 1)^(-1/r1)

Cl_21<-((G(d[,2]))^(-r1) + (G(z[,1]))^(-r1) - 1)^(-1/r1)
Cl_31<-((G(d[,1]))^(-r1) + (G(z[,2]))^(-r1) - 1)^(-1/r1)
Cl_41<-((G(d[,1]))^(-r1) + (G(z[,1]))^(-r1) - 1)^(-1/r1)

AMH_11<-(G(z[,2])*G(w[,2]))/((1-r2*(1-G(z[,2]))*(1-G(w[,2]))))
AMH_21<-(G(z[,1])*G(w[,2]))/((1-r2*(1-G(z[,1]))*(1-G(w[,2]))))
AMH_31<-(G(z[,2])*G(w[,1]))/((1-r2*(1-G(z[,2]))*(1-G(w[,1]))))
AMH_41<-(G(z[,1])*G(w[,1]))/((1-r2*(1-G(z[,1]))*(1-G(w[,1]))))

f2<-G(z[,2])-G(z[,1])
IND_001<-((AMH_11-AMH_21)/f2)*((Cl_11-Cl_21)/f2)
IND_101<-((AMH_31-AMH_41)/f2)*((Cl_11-Cl_21)/f2)
IND_011<-((AMH_11-AMH_21)/f2)*((Cl_31-Cl_41)/f2)
IND_111<-((AMH_31-AMH_41)/f2)*((Cl_31-Cl_41)/f2)

comp1<-y1*z1*w1*log((IND_001 - IND_101 - IND_011 + IND_111)*f2)

Cl_12<-((G(d[,3]))^(-r1) + (G(z[,3]))^(-r1) - 1)^(-1/r1)

164 | P a g e
Cl_22<-((G(d[,3]))^(-r1) + (G(z[,2]))^(-r1) - 1)^(-1/r1)
Cl_32<-((G(d[,2]))^(-r1) + (G(z[,3]))^(-r1) - 1)^(-1/r1)
Cl_42<-((G(d[,2]))^(-r1) + (G(z[,2]))^(-r1) - 1)^(-1/r1)

AMH_12<-(G(z[,3])*G(w[,3]))/((1-r2*(1-G(z[,3]))*(1-G(w[,3]))))
AMH_22<-(G(z[,2])*G(w[,3]))/((1-r2*(1-G(z[,2]))*(1-G(w[,3]))))
AMH_32<-(G(z[,3])*G(w[,2]))/((1-r2*(1-G(z[,3]))*(1-G(w[,2]))))
AMH_42<-(G(z[,2])*G(w[,2]))/((1-r2*(1-G(z[,2]))*(1-G(w[,2]))))

f3<-G(z[,3])-G(z[,2])
IND_002<-((AMH_12-AMH_22)/f3)*((Cl_12-Cl_22)/f3)
IND_102<-((AMH_32-AMH_42)/f3)*((Cl_12-Cl_22)/f3)
IND_012<-((AMH_12-AMH_22)/f3)*((Cl_32-Cl_42)/f3)
IND_112<-((AMH_32-AMH_42)/f3)*((Cl_32-Cl_42)/f3)

comp2<-y2*z2*w2*log((IND_002 - IND_102 - IND_012 + IND_112)*f3)

Cl_13<-((G(d[,4]))^(-r1) + (G(z[,4]))^(-r1) - 1)^(-1/r1)

Cl_23<-((G(d[,4]))^(-r1) + (G(z[,3]))^(-r1) - 1)^(-1/r1)
Cl_33<-((G(d[,3]))^(-r1) + (G(z[,4]))^(-r1) - 1)^(-1/r1)
Cl_43<-((G(d[,3]))^(-r1) + (G(z[,3]))^(-r1) - 1)^(-1/r1)

AMH_13<-(G(z[,4])*G(w[,4]))/((1-r2*(1-G(z[,4]))*(1-G(w[,4]))))
AMH_23<-(G(z[,3])*G(w[,4]))/((1-r2*(1-G(z[,3]))*(1-G(w[,4]))))
AMH_33<-(G(z[,4])*G(w[,3]))/((1-r2*(1-G(z[,4]))*(1-G(w[,3]))))
AMH_43<-(G(z[,3])*G(w[,3]))/((1-r2*(1-G(z[,3]))*(1-G(w[,3]))))

f4<-G(z[,4])-G(z[,3])
IND_003<-((AMH_13-AMH_23)/f4)*((Cl_13-Cl_23)/f4)
IND_103<-((AMH_33-AMH_43)/f4)*((Cl_13-Cl_23)/f4)
IND_013<-((AMH_13-AMH_23)/f4)*((Cl_33-Cl_43)/f4)

165 | P a g e
IND_113<-((AMH_33-AMH_43)/f4)*((Cl_33-Cl_43)/f4)

comp3<-y3*z3*w3*log((IND_003 - IND_103 - IND_013 + IND_113)*f4)

Cl_14<-((G(d[,5]))^(-r1) + (G(z[,5]))^(-r1) - 1)^(-1/r1)

Cl_24<-((G(d[,5]))^(-r1) + (G(z[,4]))^(-r1) - 1)^(-1/r1)
Cl_34<-((G(d[,4]))^(-r1) + (G(z[,5]))^(-r1) - 1)^(-1/r1)
Cl_44<-((G(d[,4]))^(-r1) + (G(z[,4]))^(-r1) - 1)^(-1/r1)

AMH_14<-(G(z[,5])*G(w[,5]))/((1-r2*(1-G(z[,5]))*(1-G(w[,5]))))
AMH_24<-(G(z[,4])*G(w[,5]))/((1-r2*(1-G(z[,4]))*(1-G(w[,5]))))
AMH_34<-(G(z[,5])*G(w[,4]))/((1-r2*(1-G(z[,5]))*(1-G(w[,4]))))
AMH_44<-(G(z[,4])*G(w[,4]))/((1-r2*(1-G(z[,4]))*(1-G(w[,4]))))

f5<-G(z[,5])-G(z[,4])
IND_004<-((AMH_14-AMH_24)/f5)*((Cl_14-Cl_24)/f5)
IND_104<-((AMH_34-AMH_44)/f5)*((Cl_14-Cl_24)/f5)
IND_014<-((AMH_14-AMH_24)/f5)*((Cl_34-Cl_44)/f5)
IND_114<-((AMH_34-AMH_44)/f5)*((Cl_34-Cl_44)/f5)

comp4<-y4*z4*w4*log((IND_004 - IND_104 - IND_014 + IND_114)*f5)

-sum(comp1+comp2+comp3+comp4)
}

hin<-function(par)
{
alpha<-par[1:9]
h<-rep(NA,1)
h[1]<-alpha1[2]-alpha[1]

166 | P a g e
h[2]<-alpha1[3]-alpha1[2]
h[3]<-alpha2[2]-alpha2[1]
h[4]<-alpha2[3]-alpha2[2]
h[5]<-alpha3[2]-alpha3[1]
h[6]<-alpha3[3]-alpha3[2]
h
}
[Link]<-function(par)
{
alpha<-par[1:9]
j<-matrix(NA,9, length(par))
j[1,]<-c(1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[2,]<-c(0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[3,]<-c(0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[4,]<-c(0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[5,]<-c(0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j[6,]<-c(0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
j
}
par<-c(-4.066, -1.203, 1.072, -2.577, -.679, 1.256, -5.060, -1.239, 1.185, -.060, -.409, -.385, -
1.137, -1.443, .380, -.730, .279, .458, .265, 1.064, -.680, .834, .073, -1.095, -1.207, .268, .442,
.250, .274, .294, .091, -.766, -.634, .039, -.539, -.056, -1.278, -1.429, .332, -.341, .362, -.363, 1.5,
0.5)
[Link]<- auglag(par, [Link] , hin = hin, [Link] = [Link])
s.e<- sqrt(diag(solve([Link]$hessian)))

167 | P a g e
Appendix E: Plagiarism Report

168 | P a g e
169 | P a g e
Appendix F: Ethical Clearance Approval

170 | P a g e
171 | P a g e
Appendix F: Language editing certificate
7542 Galangal Street
Lotus Gardens
Pretoria
0008
07 January 2021
TO WHOM IT MAY CONCERN
This certificate serves to confirm that I have edited JA Yimam’s thesis entitled, Modelling the
Stability and Determinants of Household Food Insecurity: A Multivariate Longitudinal
Ordinal Logistic Regression Approach.

I found the work easy and intriguing to read. Much of my editing basically dealt with
obstructionist technical aspects of language, which could have otherwise compromised smooth
reading as well as the sense of the information being conveyed. I hope that the work will be
found to be of an acceptable standard. I am a member of Professional Editors’ Guild.
Hereunder are my particulars:

Jack Chokwe (Mr)

Contact numbers: 072 214 5489
jackchokwe@[Link]

172 | P a g e