Statistical Approaches to Causal Analysis 1st Edition

Visit the link below to download the full version of this book:

https://medipdf.com/product/statistical-approaches-to-causal-analysis-1st-editio
n/

Click Download Now

SAGE Publications Ltd  Matthew McBee 2021
1 Oliver’s Yard
55 City Road This volume published as part of The SAGE Quantitative
London EC1Y 1SP Research Kit (2021), edited by Malcolm Williams,
Richard D. Wiggins and D. Betsy McCoach.
SAGE Publications Inc.
2455 Teller Road Apart from any fair dealing for the purposes of research,
Thousand Oaks, California 91320 private study, or criticism or review, as permitted under the
Copyright, Designs and Patents Act, 1988, this publication
SAGE Publications India Pvt Ltd may not be reproduced, stored or transmitted in any form,
B 1/I 1 Mohan Cooperative Industrial Area or by any means, without the prior permission in writing of
Mathura Road the publisher, or in the case of reprographic reproduction,
New Delhi 110 044 in accordance with the terms of licences issued by
the Copyright Licensing Agency. Enquiries concerning
SAGE Publications Asia-Pacific Pte Ltd reproduction outside those terms should be sent to
3 Church Street the publisher.
#10-04 Samsung Hub
Singapore 049483

Library of Congress Control Number: 2020950514

Editor: Jai Seaman
Assistant editor: Charlotte Bush
British Library Cataloguing in Publication data
Production editor: Manmeet Kaur Tura
Copyeditor: QuADS Prepress Pvt Ltd
A catalogue record for this book is available from the
Proofreader: Elaine Leek
British Library
Indexer: Cathryn Pritchard
Marketing manager: Susheel Gokarakonda
Cover design: Shaun Mercier
Typeset by: C&M Digitals (P) Ltd, Chennai, India
Printed in the UK

ISBN 978-1-5264-2473-0

At SAGE, we take sustainability seriously. Most of our products are printed in the UK using responsibly sourced
papers and boards. When we print overseas, we ensure sustainable papers are used as measured by the PREPS
grading system. We undertake an annual audit to monitor our sustainability.
Dedication

This book is dedicated to my mother, Mary Kay McBee, who passed away before she
could see this book in print.
Contents

List of Figures and Tables xiii

About the Author xxiii
Acknowledgement xxv
Preface xxvii

1 Introduction 1

Internal Validity 2
External Validity 2
Threats to Validity 3
Randomisation6
Non-Experimental Research 7
A Pragmatic Definition of Causation 9
Prediction Versus Explanation 10
Causal Inference Requires External Information 11
Estimation Versus Hypothesis Testing 13
Prerequisites13
Notation14
The R statistical programming environment 14
Installing and Using R and RStudio 16
R Packages 16
Structure of This Book 17

2 Conditioning 19

Simulated Data Set 21

Bias and Inconsistency 21
Obtaining a Biased Estimate of the Causal Effect 25
Covariate Adjustment 26
Visualising Covariate Adjustment 27
Covariate Adjustment Depends on Strong Assumptions 30
viii STATISTICAL APPROACHES TO CAUSAL ANALYSIS

Sample Selection 30
The Bias–Variance Trade-Off 32
Subclassification33
Matching35
Weighting39
Computing the Weights 39
The Problem of Measurement 42
Classical Test Theory Model for Measurement Error 43
Reliability 44
Discussion46
The ‘Curse of Dimensionality’ 47

3 Directed Acyclic Graphs 51

DAGs Are Not Path Models 53

DAG Terminology and Variable Roles 53
Exposure 54
Outcome 55
Mediator 55
Confounder 55
Proxy Confounder 56
Instrument 56
Competing Exposure 57
Collider 58
d-Separation, d-Connectedness and Statistical Independence 58
Conditioning 59
Conditioning on Colliders 61
Colliders and the Real World 62
Spurious Paths 64
Unobservables69
Conditioning on Mediators 69
Criteria for Valid Causal Inference 72
Back-Door Criterion 72
Front-Door Criterion 73
Minimal and Sufficient Adjustment Sets 74
Simultaneous Estimation of Causal Effects 75
Measurement Error and DAGs 76
Using DAGitty77
Practical Recommendations 81
 contents ix

4 Rubin’s Causal Model and the Propensity Score 85

The Counterfactual Framework 86

Defining Causal Effects Under Rubin’s Causal Model 87
The Fundamental Problem of Causal Inference 88
Ignorability90
Bias When Ignorability Does Not Exist 92
Baseline Bias 92
Differential Treatment Effect Bias 93
Conditional Ignorability 93
Conditional Treatment Effects 94
Example: Estimating ATT, ATU and ATE via Linear Regression 95
The Propensity Score 97
Approximating an Experiment 98
Simulated Data Set 101
Propensity Scores 101
Estimating Propensity Scores via Logistic Regression 102
Solving the Curse of Dimensionality 108
Propensity Score Estimation via Boosted Classification Trees 108
Comparing the Two Sets of Propensity Scores 115
Assumptions of Propensity Score Methods 116
Ignorability 116
Stable Unit Treatment Value Assumption 116
Positivity 117

5 Propensity Score Analysis 119

Simulated Data Set 120

Descriptive Statistics and Biased Treatment Effect Estimate 121
Obtaining a Biased Estimate of the Treatment Effect 121
Propensity Score Matching 124
Matching Algorithms 124
Estimating Treatment Effects with Matching 129
Example Analysis 129
Stratifying on the Propensity Score 131
Weighting with the propensity score 136
From Propensity Scores to Weights 139
Stabilised Weights and Truncated Weights 141
Example of an Analysis Using Propensity Score Weights 143
Doubly Robust Estimation 147
x STATISTICAL APPROACHES TO CAUSAL ANALYSIS

6 Instrumental Variable Analysis 151

Endogeneity and Bias 153

Defining Instrumental Variables 156
Finding an Instrument 159
The Two-Stage Least Squares Estimator 162
Step 1: Obtain the Predicted Values of the Exposure Variable 163
Step 2: Estimate the Treatment Effect 164
Simultaneous Two-Stage Least Squares 165
Sample Size Issues 169
Measurement Error 169
Imperfect Measurement of the Exposure 170
Measurement Error in the Instrument 172
Local Average Treatment Effects 174
Assumptions of Instrumental Variable Analysis 175

7 Regression Discontinuity Design 177

The Forcing Variable and Treatment Assignment 179

Sharp RDD 180
Extrapolation via Parametric Regression 182
Example of Sharp RDD Analysis 191
Fuzzy RDD 194
Example of Fuzzy RDD Analysis 197
Local Average Treatment Effects 199
Assumptions of the RDD 202

8 Conclusion 205

Some Practical Advice 207

Disclose Your DAG 207
Test Your DAG 208
Sensitivity Analysis 208
Don’t Ignore Precision 209
Don’t Fool Yourself 209
What to Learn Next 210
Campbell and Stanley’s Versus Rubin’s Perspectives
on Causation 210
More About DAGs 210
 contents xi

Principal Stratification 210

Fixed Effects 211
Closing211

Glossary 213
References 217
Index 229
List of Figures and Tables

List of figures

1.1 RStudio 16

2.1 Example of a DAG for the conditioning example data set 20

2.2 Illustration of an unbiased and consistent estimator. Sample size is
on the x-axis. The estimated regression coefficient of A on C is on the
y-axis. Each dot is a regression slope coefficient from a simulated study.
The true value of the causal effect is depicted by the horizontal line 23
2.3 Illustration of a consistent estimator that is biased at small
sample sizes. The estimator is convergent and asymptotically unbiased 24
2.4 Two examples of inconsistent estimators. Left panel: The estimates
converge to the wrong value as the sample grows. These estimates are
both biased and inconsistent. Right panel: The estimates are unbiased
but inconsistent; they do not converge to the true value 25
2.5 Simple regression of C on A. Values of B are depicted by point shading
and size, with bigger, lighter points higher on B. Because A causes B, cases
with high values of A also tend to have high values of B. And because
B causes A, cases with high values of B tend to have high values of C. This
creates a spurious relationship between A and C that is portrayed by the
regression line 28
2.6 Conditioning via covariate adjustment with linear regression. Upper left:
3D scatter plot of A (x-axis) and B (y-axis) versus C (z-axis) viewed from
above (with respect to C). Upper right: Rotation to show the points
plotted against the regression surface. Lower left: Rotation to view the
regression surface with respect to B. Variation on A is invisible and
irrelevant from this perspective, as it occurs along the line of sight of
the observer. Lower right: Rotation to view the regression surface
with respect to A. Variation on B is invisible and irrelevant from this
perspective29
xiv STATISTICAL APPROACHES TO CAUSAL ANALYSIS

2.7 The bias–variance trade-off. The figure displays point estimates

and 95% confidence intervals by sample selection bandwidth. The dotted
vertical line indicates the true value of the causal effect. The values on
the x-axis represent the sample selection bandwidth, which is the range
of allowable values on variable B that were retained in the analysis. Wider
bandwidths retain a larger sample and therefore exhibit less sampling
variance (narrower confidence intervals), but are more biased 32
2.8 Conditioning via matching: Scatter plot of Btreatment versus Bcontrol after
matching. Each point represents one matched pair. Dotted lines indicate
caliper boundaries 37
2.9 Conditioning via weighting. Left panel: Distribution of academic ability by
treatment group before weighting. The target distribution is outlined. Right
panel: Distribution of academic ability by group after weighting.
The distribution of academic ability in the control group has been altered
by the weighting to resemble its distribution in the treatment group 39
2.10 Biased and inconsistent estimates when conditioning on an
imperfectly measured confounder. Sample size is on the x-axis.
The estimated regression coefficient of A on C is on the y-axis. Each
dot is a regression coefficient from a simulated study. Bias occurs when
the estimates do not converge to their true value (zero) as the sample
size increases. Upper left: Confounder variable omitted from the model.
Upper right: Conditioning on a poorly measured confounder. Lower left:
Conditioning on an ‘adequately’ measured confounder. Lower right:
Conditioning on a superbly measured confounder 45
2.11 Point estimates and 95% confidence intervals by conditioning method.
The dotted vertical line indicates the true value of the causal effect 46
2.12 Conditioning via covariate adjustment can involve extensive
extrapolation across regions of sparse or absent data 47

3.1 An example of a DAG 52

3.2 An improper DAG containing a feedback loop 52
3.3 Example of a DAG illustrating ancestors and descendants 54
3.4 Example of a DAG with mediation 55
3.5 Example of a DAG with proxy confounders. Z and B are
confounders. A and B are proxy confounders of Z57
3.6 Example of a DAG with an instrument 57
3.7 Example of a DAG with a collider 58
3.8 Another example of a DAG containing a collider 58
3.9 DAG for selection on college admission example 62
 list of figures and tables xv

3.10 Bias from conditioning on a collider. The regression line fitted to only
the college attendees (solid line) has a negative slope. The regression line
fitted to all of the data (dashed line) has the correct zero slope 65
3.11 DAG containing a spurious path 65
3.12 Effect of conditioning. Left panel: Conditioning on A. Centre panel:
Conditioning on B. Right panel: Conditioning on Z. Any of these
conditioning decisions block the spurious connection between X and Y
and would permit the unbiased estimation of their non-existent causal
relationship66
3.13 A DAG containing an unobservable (U) and an instrument (A).
The causal effect of X on Y cannot be identified via conditioning 69
3.14 DAG containing direct and indirect effects 70
3.15 Effect of conditioning on mediators. Left panel: Conditioning on A.
Centre panel: Conditioning on B. Right panel: Conditioning on A
and B71
3.16 An illustration of the front-door criterion. U is an unobservable
confounder73
3.17 Simultaneous estimation of causal effects and potential DAGs.
Panel 1: A and X as competing exposures. This is the only condition
in which simultaneous estimation should be used. Panel 2: A as
confounder of X → Y, X as mediator of A → Y. Panel 3: A as mediator of
X → Y, X as confounder of A → Y. Panel 4: A as mediator of X → Y. X and
U as confounders of A → Y. The causal effect A → Y is not estimable after
conditioning on X due to U75
3.18 A DAG with measurement error. X is the exposure variable and Y is
the response. Ztrue is the unobservable confounder true score. Zobs is
the observed version of the confounder, which is contaminated with
measurement error. Conditioning on Zobs cannot remove all the Ztrue’s
confounding77
3.19 dagitty.net is a tool for drawing and analysing DAGs 78
3.20 dagitty.net analysis of the DAG after conditioning on the collision
node Z. Two sufficient adjustment sets are identified 78
3.21 Equivalent DAGs. These graphs generate the same implied conditional
independencies. They cannot be distinguished on the basis of data 80

4.1 Combinations of the values of X1 and X2 that produce a propensity
score of p = .8 100
4.2 Equivalence in distribution on X variables after conditioning on the
propensity score 100
xvi STATISTICAL APPROACHES TO CAUSAL ANALYSIS

4.3 Example of a DAG for the propensity score analysis examples 101
4.4 Logit link mapping values of the linear predictor to p103
4.5 Logistic regression–derived propensity score density by treatment
status from the analysis data 105
4.6 Example of propensity score distribution with no common support.
Causal effects cannot be estimated with propensity score methods
under this condition 106
4.7 Example of propensity score distribution with moderate overlap.
Left panel: Distribution of propensity scores by group. Right panel:
Distribution of propensity scores after discarding cases 107
4.8 Propensity score dimension reduction. Black circles are treatment cases.
White triangles are control cases. Panel 1: Three-dimensional scatter plot
of confounders A (x-axis) and B (y-axis) versus treatment status (D; z-axis +
shape and colour) viewed from above. Panel 2: Rotation to show the points
plotted against the predicted probability (propensity score) surface. Panel 3:
Further rotation to display the sigmoid shape of the surface with vertical
reference lines connecting each point to its estimated propensity score.
Panel 4: Location of each point with respect to the propensity score (z-axis).
The formerly two-dimensional location of each point has been reduced
to its location along a one-dimensional number line 109
4.9 The two-dimensional space of variables A and B is reduced to a
one-dimensional location along a number line of propensity scores 109
4.10 Example of classification tree. The numbers below each leaf are the
predicted probabilities of treatment 110
4.11 A classification tree partitions the data into areas of maximum
class similarity 111
4.12 Boosted model performance in-sample (bottom line) and out-of-sample
(top line) by ensemble size. The dashed vertical reference line indicates
maximum out-of-sample classification performance 113
4.13 Boosted model–estimated relationships between model variables
and the (logit) probability of receiving treatment 114
4.14 Boosted classification tree–derived propensity score density by
treatment status from the analysis data 115
4.15 Scatter plot of logistic propensity scores versus boosted
propensity scores 116

5.1 Example of a DAG for the propensity score analysis examples 120
5.2 Frequency distributions of variables A and B by treatment
condition (D)121