Food Policy 35 (2010) 211–220

Contents lists available at ScienceDirect

Food Policy
journal homepage: [Link]/locate/foodpol

Traffic lights and food choice: A choice experiment examining

the relationship between nutritional food labels and price
Kelvin Balcombe a, Iain Fraser b,*, Salvatore Di Falco b
a
Dept of Ag and Food Economics, University of Reading, Earley Gate, PO Box 237, Reading RG6 6AR, UK
b
University of Kent, School of Economics, Canterbury Kent, CT2 7NP, UK

a r t i c l e i n f o a b s t r a c t

Article history: In this paper we investigate how consumers respond to the UK nutritional food label Traffic Light System
Received 11 May 2009 (TLS). Employing a choice experiment (CE) we find that consumers appear to behave in a manner consis-
Received in revised form 5 November 2009 tent with our expectations regarding the impact of the TLS. We identify a strong preference on the part of
Accepted 21 December 2009
respondents to avoid a basket of goods containing a mix of foods with any ‘‘Red” lights. In addition, we
find that consumers have a hierarchy of importance in terms of perception of the various nutrients exam-
ined and there are clear behavioural differences associated with particular socio-economic characteristics
Keywords:
confirming early research on the use of nutrition labels. Overall our results indicate significant heteroge-
Nutrients
Traffic Light System
neity in the attitudes and responses of consumers to the TLS nutritional food labels within and across
Choice experiment socio-economic strata.
Bayesian mixed logit Ó 2010 Elsevier Ltd. All rights reserved.

Introduction that can then be converted into a per portion quantity. Thus, for
any food the resulting quantities of these nutrients are measured
There is an ever-growing awareness and understanding of the then compared against the TLS which in turn provides the colour
relationship between food and the role it plays in health and well coding on the food packaging. In practice, the TLS is meant to aid
being. The importance of this relationship is increasing as the consumers in getting the balance of products right in terms of their
health implications of a poor diet in the UK, and elsewhere, have overall diet. It can be used a means to monitor the amount of food
become ever more apparent (HMSO, 2007). In response, there are being consumed that is high in one or more of the nutrients
now a whole raft of policy approaches that attempt to address this identified.
issue including various health campaigns such as the Food Stan- There already exists a large literature that has examined various
dards Agency (FSA) ‘‘The eatwell plate” (FSA, 2007), the FSA ‘‘6 g aspects of the TLS and nutritional labels in general. In comprehen-
per day of salt intake” and the Department of Health ‘‘5-a-day cam- sive reviews of the literature on consumer use and response to
paign” for fruit and vegetable consumption. All of these campaigns nutrition information on food labels, Cowburn and Stockley
have been accompanied by a drive to have food labelled in a man- (2005) and Grunert and Wills (2007) observe that it has been
ner that provides important dietary information to consumers. established that the use of label information can alter overall food
In terms of nutritional food labelling, the UK has voluntarily purchase behaviour. Furthermore, research indicates that most
adopted the Traffic Light System (TLS)1 which indicates the levels consumers are interested in nutrition information and that they
of four key nutrients i.e., fat, sugar, saturates and salt, which are use nutrition labels. However, as many researchers note the use
found in processed food. The TLS system is relatively simple. A of nutrition labels in actual food choice is almost certainly lower
Red2 light indicates a very high level of a specific nutrient, Amber than consumers claim in surveys. In addition, it has been observed
a medium amount and Green low. The choice of colour is based that when consumers are confronted by complex food choices in
on the content of each of the nutrients per 100 g of any food type terms of food selection they are less able to make informed choices.
For example, Black and Rayner (1992) noted that consumers strug-
* Corresponding author. Address: University of Kent, School of Economics, gle to understand how to process information when they are
Canterbury Kent, CT2 7NP, UK. Tel.: +44 (0)1227 823513; fax: +44 (0)1227 827850. shown several nutrients simultaneously. Indeed, there is evidence
E-mail address: [Link]@[Link] (I. Fraser). that consumers will employ a heuristic that sees food choices
1
If a specific nutritional claim is made it is necessary for producers to adhere to the made in terms of a specific nutrient. Evidence from New Zealand
appropriate EC Directive in terms of label format. Within the UK food labelling is
reported by Mhurchu and Gorton (2007) indicates that consumers
regulated by EC Directive 90/496/EEC.
2
For interpretation of colour in Figs. 1 and 2, the reader is referred to the web do not understand how to balance the consumption of nutrients in
version of this article. their diet and often make choices based on the fat content of food

0306-9192/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/[Link].2009.12.005
212 K. Balcombe et al. / Food Policy 35 (2010) 211–220

regardless of other nutrient levels. Grunert and Wills (2007) note the nutritional label represents the amount of each nutrient within
that in terms of nutrition information interest, calories and/or fat a basket of goods using colour.
are frequently cited as being of most interest to consumers fol- From a methodological perspective our paper adds to a small
lowed by salt and sugar. literature that having employed a CE to generate stated preference
Given these important observations about how consumers cope data then estimates the resulting Mixed Logit (MXL) model using
when making complex food choices as well as preferences for spe- Bayesian methods. In addition, the model is estimated in WTP
cific nutrients we aim to reveal the relative value that consumers space as opposed to preference space. As Balcombe et al. (2009) re-
attached to specific nutrients. To do this we examine the TLS from port, there are very few examples in the literature adopting this
a different perspective to that previously undertaken in the litera- approach which is surprising given the number of important
ture. We conduct a choice experiment (CE) to examine consumers’ econometric benefits that emerge as a result. Finally, we also em-
willingness-to-pay (WTP) for reductions in the various nutrients as ploy a model specification that allows respondents to be indiffer-
indicated by the TLS, i.e., fat, saturates, sugar and salt, in terms of a ent to the choices presented. We take this approach because of
basket of shopping. This analysis allows us to reveal the relative the observation that many consumers employ a simple heuristic
values placed on the reduction of each specific nutrient in terms when making food choices with respect to nutrients.
of going from Red to Amber and from Amber to Green. The structure of this paper is as follows. We begin by reviewing
The reason why we employ a basket of goods as opposed to spe- the literature on the TLS as well as economic issues related to
cific products is that the TLS is designed to help food purchase nutritional labels of specific interest to our study. In Section choice
choice as part of a healthy diet. Consumers need to consider the experiment design and data we describe the design and implemen-
mix of all food being purchased and consumed, and assess con- tation of the survey instrument used in our CE. Then in Section
sumption against ideal dietary requirements. Thus, there is nothing econometric methods we describe the method of analysis we em-
preventing a consumer from eating a bag of crisps, or a piece of ploy in this paper. In Section results we present our survey results
cheese, as long as they compensate for these food types with mod- and in Section summary and conclusions we provide a summary
eration elsewhere in their diet. Importantly previous research on and conclusions.
nutrition labels has identified that consumers find it difficult to
employ nutritional label information to place a specific food item
within an overall dietary plan (Cowburn and Stockley, 2005). Fur- Literature review
thermore, Wansink and Chandon (2006) note that consumers who
select a healthy food option frequently over compensate with some The TLS
sort of indulgence, yielding a negative impact in terms of their die-
tary intake. For these reasons simply focussing on a single food The TLS has been the subject of ongoing research and develop-
item within a CE could lead to behavioural outcomes that will ment by the FSA since 2004 (see [Link] for more
not capture how the TLS should be used in helping to achieve a details). As noted by Drichoutis et al. (2006) the development of
healthy diet. the TLS can partly be explained as a response to difficulties which
The emphasis here on a basket of goods as opposed to specific consumers had with earlier nutrition label systems. However, the
product is not without precedent. For example, a basket of goods emergence of the TLS has not been a simple or cooperative process
as a means to assess dietary goals has previously been employed on the part of the public and private sectors. Lang (2006) describes
by Jetter and Cassady (2006) who examined the US Department in detail the struggle that the FSA have had in bringing the TLS into
of Agriculture Thrifty Food Plan diet guide that employs a specific operation. He notes that research has found that consumers find
basket of food items. Furthermore, most consumers will engage the TLS easy to use, especially when compared to alternative nutri-
with the TLS as part of the weekly shopping experience and it is ent food label formats. Yet despite the clear public health motiva-
within this context that we should examine the impact of the tion for implementing the scheme the food industry has at times
nutrition label. been less than cooperative. As a result there are a number of alter-
Given the design of our CE and the statistical methods we native nutritional label systems employed in the UK. For example,
employ to analyse the data we make several contributions to the some in the food industry have advocated the use of a Guideline
literature on TLS and nutritional labels in general. First, and our Daily Amounts (GDA) system that relates food intake to a total dai-
main contribution, is that our results provide information about ly target.
changes in consumer choice in relation to the colours. Specifically, There have also been ongoing debates about the use of front or
we find that the relative WTP of moving from Red to Amber is back of packaging labels. In comprehensive reviews of the litera-
much greater than that from moving from Amber to Green. This ture, Cowburn and Stockley (2005) and Grunert and Wills (2007)
is an important insight into how consumers respond to the TLS. report that research suggests that front-of-label information
In particular, it may help to explain and inform the food supply should be simple with the more complex detailed nutritional infor-
chain industry is responding to the dietary demands of consumers mation presented on the back. By presenting information in this
by modifying its processed foods products. Our analysis also re- way consumers are able to make a quick decision at the point of
veals the relative importance of each of the nutrients as revealed purchase as well as being able to examine in more detail at their
by our WTP estimates. leisure specific nutrient details.
Second, the use of CEs to undertake research on nutrition label- Overall there appears to be general support for the use of the
ling is very limited in the literature to date. To date the only other TLS amongst many health professionals in the UK. For example,
paper that has employed a CE to examine issues relating to con- the Children’s Food Campaign (2007) supports the use of the TLS.
sumer use of nutritional labels is Berning et al. (2008). This study They note that the TLS can be employed by consumers very
examines preferences for detailed versus summary nutritional quickly. Other more complex labelling systems increase the risk
information. It reveals that the alternative nutritional label formats of widening existing inequalities in food choice. This is because a
appeal to different groups of respondents. Most other CE applica- large proportion of the UK public finds it difficult to understand
tions in the literature (e.g., Teratanavat and Hooker, 2006; Bond what many of the numerical values employed on food labels imply
et al., 2008) examine health claims on food packaging as opposed about healthy eating.
to nutritional information. They examine presence or absence of However, some researchers are slightly less positive about the
claim or the strength of a claim for a specific product. In our CE TLS. For example, Feunekes et al. (2008) examined front-of-pack
 K. Balcombe et al. / Food Policy 35 (2010) 211–220 213

nutrition labelling for various label formats including the TLS. They survey was conducted in Spain and the sample size was 400 indi-
conducted survey work across four European countries. One of viduals. The motivation behind this research stems from the fact
their studies looked at three specific products with each survey that the EU have been considering the adoption of mandatory food
participant shown three out of the six nutritional labelling systems labels on food much like the Nutritional Labelling and Education
being examined. The respondents had to rate the each of the label Act (NLEA) introduced in the US in 1994. This study found that con-
systems based on liking, comprehension, credibility and perceived sumers were WTP 11% more for a product with this information,
healthiness. In general the TLS performed very well except in terms although this varied by type of consumer. In related research Gra-
of perceived healthiness per product category. In this part of their cia et al. (2007) examined how consumers value the potential
study Feunekes et al. report that the TLS did not differentiate as introduction of mandatory nutrition labels. Employing a three-
well as some of the other nutritional labels with respect to health- equation multivariate probit model they reveal that individuals
ier and less healthy products. However, they do conclude by noting with food related health problems know more about nutritional la-
that the TLS like several of the other formats examined in the study bels, and being knowledgeable makes an individual more likely to
is understood, liked and considered as credible by consumers. use a label and that label users do consider mandatory labels as
It also needs to be appreciated that although consumers may beneficial.
prefer simple information presented on the front of packages this Berning et al. (2008) employed a CE to examine how consum-
does not mean that they react to it in the manner intended (Verbe- ers respond to detailed versus summary nutritional information.
ke, 2005). For example, Grunert and Wills (2007) observe that even The CE employed a number of nutrition label designs and infor-
though the TLS is a relatively simple nutritional labelling system, mation sets for a single product, a can of tomato soup. Employing
consumers can struggle to understand it when attempting to con- a face-to-face survey method a total of 410 individuals partici-
struct a meal. In addition, when it comes to actual use of the TLS pated. The data was analysed using a MXL model employing Clas-
not everybody responds to Red as intended because taste can over- sical methods assuming normal distributions for all random
ride considerations related to health. However, there is also some parameters in one model and triangular in another. The main
indirect evidence that consumers do use the TLS. Several food finding is that different nutritional label formats appeal to differ-
retailers, including Sainsbury’s and Waitrose, have adopted the ent groups of shoppers.
TLS. Based on this there is some evidence reported by Grunert In addition to the research that has employed specific con-
and Wills (2007) that suggests that by adopting the TLS these sumer surveys to examine issues of nutritional label use, there
retailers have subsequently found it advantageous to reformulate have also been a number of studies that have examined actual
certain products, so as to remove Red lights from various products. food purchase behaviour (e.g., Kim et al., 2000 and Variyam,
This would suggest that consumers are responding to the TLS and 2008). These papers have all made extensive use of the US
retailers have responded as they see a change in the mix of prod- Department of Agriculture’s 1994–1996 Continuing Survey of
ucts being purchased. Food Intakes by Individuals as well as the Diet and Health Knowl-
edge Survey. The reason why this particular part of the literature
Economics and nutritional labels is relevant to our research is that it provides insights into the rel-
ative importance of nutrients being consumed once nutritional
There is a large literature that has examined various economic labelling has been introduced. For example, Kim et al. report sig-
issues associated with the liking, use, understanding and develop- nificant effects as a result of label use on the consumption of fats,
ment of nutritional labels. In addition to the material already cited cholesterol, sodium and fiber. However, Variayam challenges
other examples in the literature include Drichoutis et al. (2005, these findings by employing methods of estimation to deal with
2006); Gracia et al. (2007); Kim et al. (2000); Loureiro et al. issues of self-selection in the data. These findings raise doubts
(2006); Verbeke (2005) and Variyam (2008).3 about the extent to which the introduction of the NLEA resulted
In general it has been established in the literature that consum- in changes in nutrient intake except for fiber and iron. Thus, these
ers are interested in the provision of nutrition information results indicate only a minimal response on the part of the public
although the extent and detail of this information demanded by to nutritional labels.
consumers varies by product as well as context. In a comprehen-
sive review of the literature Drichoutis et al. (2006) state that
nutritional label use does impact purchase behaviour as consumers Choice experiment design and data
want to avoid things that are considered bad for them. However,
they also find that there is little consistency in research attempting Designing and implementing the choice experiment
to explain what determines label use in relation to income, age or
working status. There is, however, evidence to suggest that being Our CE was design and implemented so as to examine consum-
female and being educated are positively related to label use. Fur- ers’ WTP for reductions in the nutrients in the TLS. As we have
thermore, having time to shop, being concerned about diet for var- explained we decided to design our CE around a basket of goods
ious reasons such as a diet-disease relation and having prior as opposed to specific food items. We did this because the TLS is
knowledge about nutrition are all positively related to label use. designed to help consumer food choice as part of achieving a
Finally, as might be expected consumers who are price sensitive healthy diet. As we will explain the design of the CE we employed
are less likely to be interested in and use labels. allowed us to estimate the WTP for each nutrient in terms of going
In terms of actual studies examining nutritional labels Loureiro from Red to Amber and from Amber to Green.
et al. (2006) employed a survey designed to reveal consumer WTP A critical decision was to frame the choices in terms of a basket
for a nutritional label on a box of cookies. The WTP values were of goods rather than a specific item. Our rationale was that a spe-
derived by employing a Contingent Valuation (CV) survey cific item would not reflect the purchasing behaviour of the indi-
instrument that used a dichotomous choice question format. The vidual in a general enough way. For example, the response of
consumers to the nutrients in a meat pie or ready meal would
3
hardly reflect their responses to a range of items. Consumers are
There is related literature that has examined the impact of health claims on food
packaging. Examples include Teratanavat and Hooker (2006) and Bond et al. (2008).
obviously ready to tolerate high levels of nutrients in some items
Both used CE to examine how specific health claims impact consumer choice and the but not others. Our piloting suggested that consumers could read-
resulting WTP for a product. ily conceptualize a representative basket.
214 K. Balcombe et al. / Food Policy 35 (2010) 211–220

We began our CE survey instrument design process by present- choices are hypothetical, we needed them to make their choices as
ing a number of draft choice cards to a small group of participants if they are real. Furthermore, we also needed to establish the legit-
(i.e., students and staff). From this we developed our set of CE attri- imacy of the cost element in the choice set so that the resulting
butes and the associated levels which we then used to conduct a WTP estimates are meaningful. The development of the informa-
focus group exercise. The nutrients and the associated TLS colours tion provided in this part of the survey instrument was the result
constituted our main CE attributes. In keeping with the TLS we of extensive pre-testing with focus group participants as well as
included four nutrients: Salt, Sugar, Fat and Saturates. Next we discussions with experts in food marketing.
devised our basket of goods based on the report produced by Having identified our set of attributes (number of attributes by
Synovate (2005) for the FSA. Based on the mix of goods in our number of levels) we then generated the choice sets employed in
hypothetical basket we then referred to the National Statistics the CE. Employing a full factorial design, ensuring balance across
(2007) publication, Family Food in 2005–06 to establish the all attributes, yielded a total of 24 choice sets. To keep the survey
expected cost of this basket of goods for an average UK household. task manageable and to avoid response fatigue we blocked the
This yielded a value of £20. Having established the status quo price 24 choice sets into four groups of six. Thus, each respondent only
we determined the number of price points used in the survey. We answered six choice sets. Within each choice set the status quo
decided to employ five price points with £20 as the mid point. The option was always Option 1, with Options 2 and 3 alternative
other price points used in the CE are £15, £18, £25 and £30. Impor- options. We generated the alternative options randomly from the
tantly, we allowed the cost of the basket to be lower than the sta- set of 24, always ensuring a balance in terms of attribute levels
tus quo. We considered it important to offer this option because across the design. We also included a ‘‘Don’t Know” as part of
some consumers may well be far more price sensitive than health the choice set. An example of the choice card used in the CE is
sensitive, a behaviour observed in previous CE. presented in Fig. 2.
In terms of the survey instrument we began by briefly
explaining to participants that the government has introduced
the TLS to signal the impact of certain nutrients on health. We Survey returns
then described the system using simple language. Next we pro-
vided an explanation of the task involved in the CE. This is The survey was distributed in the mail to 3000 UK households.
shown in Fig. 1. It was a single shot survey with participants offered a simple finan-
An important facet of the description of the CE task was that we cial incentive to complete the survey. We received 477 useable re-
needed to make it clear to all survey participants that although the turns and these are used in our analysis.

Section C: Survey Description

W e are g oing to ask yo u 6 que stion s co nce rnin g fo od ch oice.

You simply indica te the ba ske t you would buy if offere d these options in
a shop.

The option s pre sen te d relate to an ind ividu al’s wee kly ba sket o f fo od wh ich
m ig ht in clu de:

• Re ad y m eals
• Ch icken bu rge rs/p izzas
• P asta read y m eals/cu rry read y m eals
• Ca ke /crisps
• Ce re al b ars/bre akfast cereals

We de scribe the b asket of food u sin g the Traffic Light S yste m a nd co st.

The cost of the b asket is fo r a typical UK con su m er fo r one w ee k’s shop ping.

This is, on a verage , ab out £ 20.

In the s urvey , an option would be de scribed as follows :

This colum n rep resen ts

Food Bask et O ption A
th e nu trie nt con te nt in a
ba sket of go ods. We
Salt G reen
sim p ly call this ba sket
Sugar G reen Option A . U sin g the
Traffic Light S ystem we
Fat Am ber co de ea ch n utrien t.
Saturates G reen Finally, £2 0 is th e p rice
fo r b uyin g this ba sket
Price £20 of good s

Fig. 1. The choice experiment survey description page.

 K. Balcombe et al. / Food Policy 35 (2010) 211–220 215

CHOICE CARD 1 Finally, we also asked survey respondents a number of ques-

tions concerning food and health issues. Some 94% of respondents
Food Basket Option 1 Option2 Option3 consider themselves to be health conscious. In terms of food label
Salt Amber Red Green use the majority claim to read food labels some of the time. We
Sugar Amber Green Amber Don’t also asked participants to rank the nutrients that appear on the
Fat Red Amber Red Know FSA TLS. Ranking, salt, sugar, fat and saturated fats on a scale of
1–4 in terms of most to least important in terms of wellbeing we
Saturates Amber Amber Red
found the following average scores: salt = 2.4; sugar = 2.9;
Price of basket £20 £25 £30 fat = 2.6; and saturates = 2.1. These indicate that saturates followed
Tick ONE and by salt are considered the most important in terms of making a
only one box negative impact on the wellbeing of respondents.
Fig. 2. Sample choice card.
Econometric methods

We begin our analysis of the survey returns by examining The main objective of this paper is to estimate consumer WTP
various socio-economic descriptive statistics that are reported in from a CE designed to avoid high levels of nutrients as indicated
Table 1. by the TLS. To analyse our CE data we employ Bayesian methods
Beginning with various socio-economic data the average age of to estimate the MXL model following Balcombe et al. (2009).
respondents is 48 years. This is slightly higher than the average age Formally, let xj,s,n denote a k  1 vector of attributes from the CE
in the UK which is 39 in 2007. Also it is slightly higher than re- presented to the jth individual (j = 1, . . .. . ., J) in the sth option
ported in related research. For example, Loureiro et al. (2006) have (s = 1,. . ., S) of the nth choice set (n = 1,. . ., N). Next assume that
a sample average of 46.8 years and Berning et al. (2008) report Uj,s,n is the utility that the jth individual attains from xj,s,n. In addi-
42.4 years. In terms of income our sample average (excluding zero tion, let yj,s,n be an indicator variable that is equal to one if the jth
responses which accounted for some 20% of the sample) is £24,500 individual chooses the sth option within the nth choice set, and
and was reasonably evenly distributed. In comparison to popula- zero otherwise.
tion data for the UK average income for non-retired households An individual j is assumed to receive linear utility from the sth
was £37,600 in 2006/07 and the average for all households includ- choice in the nth choice set, although the parameters may be trans-
ing retirees £30,000. Thus, the reported levels of income are a little formed. Consequently, the utility function is of the form
below the UK average. U j;s;n ¼ x0j;s;n tðbj Þ þ ej;s;n ð1Þ
Next we consider the gender mix of respondents. We had 81%
females and 19% males. This is slightly higher than in related sur- where bj is a (k  1) vector describing the preferences of the jth indi-
veys that report rates of female participation in the range of 60– vidual and t(.) is some transformation of the parameters. The func-
70%. To illustrate the effect of this characteristic our analysis we tion t(.) can take a number of forms (see Balcombe et al., 2009). The
present results that correct for this imbalance where we assume error term ej,s,n is assumed to be extreme value (Gumbel) distrib-
a 50:50 gender mix which is in keeping with current UK uted, independent of x0j;s;n and uncorrelated across individuals or
demographics. choices.
In terms of children in the household, 67% of respondents did Specifically, we employ the censored normal distribution for all
not have children living at home. Currently within the UK there random parameters, except the price of the basket, such that the
are estimated to be 33% single households and just over 60% of preference distribution is censored from below at zero, with a mass
households have no children. Turning to the number of children point occurring at zero. By censoring the normal distribution from
in the household, our sample has on average 0.6. below at zero yields a mass point at zero so that with b normally

Table 1
Survey descriptive statistics.

Socio-economic data Variable

Age (years) Average age in sample 48.34 (13.3)
Income (£ 000’s) Average income in sample 24.5 (21)
Gender Male 19%
Female 81%
Marital status Married 73%
Single 11%
Other 16%
Number of children in household children 0 67%
1 13%
2 14%
3 5%
4 1%
Educational achievement School to 16 25%
A Level or equivalent 11%
Further education 32%
University Undergraduate 18%
University postgraduate 12%
Higher 2%
Employment status Employed 48%,
Self-employed 12%
Unemployed 15%
Other 25%
216 K. Balcombe et al. / Food Policy 35 (2010) 211–220

distributed with mean b and variance r the transformation is Table 2

t = max(0,b). Values of b and r, and hence t, are estimated giving MXL results for indifference model.

the population proportions massed at zero and above zero. Attributes Posterior parameter estimates
With respect to price we employ a log-normal transformation Mean Standard deviation Median
t(b) = exp(b) with the distribution bounded below at zero and with
Price 1.720 1.504 1.297
zero probability mass at zero. Saturates-A/G 0.437 0.424 0.355
The reason why we employ a bounded distribution is because Saturates-R/G 1.727 0.591 1.741
we consider it necessary to accommodate indifference on the part Fat-A/G 0.120 0.197 0.000
of respondents in our CE. If we allow a respondent to be indiffer- Fat-R/G 1.208 0.519 1.207
Sugar-A/G 0.381 0.372 0.311
ent with respect to some attributes this implies that we are
Sugar-R/G 1.422 0.437 1.438
allowing marginal utility to be equal to zero for these attributes. Salt-A/G 0.263 0.310 0.151
The reason for wishing to allow indifference is that we do not Salt-R/G 1.932 0.540 1.959
believe that respondents are negatively disposed to any particular SQ 0.366 0.316 0.323
attributes. But, in the case of food choice based on nutrient con-
tent it is highly likely that some individual’s will be indifferent hypothesis is no-difference between the first and second half of the
to some of the attributes because of health concerns or dietary sampled values (with a sub-set of values removed from the
requirements and/or restrictions. Their choices will be driven middle).4
by a sub-set of the attributes and there is evidence from the
literature to support this view (e.g., Black and Rayner, 1992,
and Grunert and Wills, 2007). Results
Another important feature of the MXL model we employ is that
we estimate our model in WTP space. In order to estimate the MXL Our preferred specification allows for potential heterogeneity in
in WTP space we employed a re-parameterisation of the form: all the parameters characterising preferences. The results we pres-
 0 ent are for a MXL estimated in WTP space. As previously indicated
tðbj Þ ¼ t1 ðb1j Þ 1; t 2 ðb2j Þ; . . . . . . . . . t k ðbkj Þ ð2Þ
we allow for the possibility of indifference with respect to the var-
which means that the quantities t2 ðb2j Þ; . . . . . . . . . tk ðbkj Þ are the ious attributes by employing a censored normal distribution for all
Marginal Rates of Substitution (MRS) with the numeraire being parameters except the price of the basket of goods, which we have
the first attribute, which will always be the price or cost attribute modelled as a log-normal distribution. The model specification
within the CE. The benefits to be gained by estimating in WTP estimated takes the following form
space are that the MRS are estimated directly and this can signifi-
cantly reduce the instability associated with WTP estimates de- U j;s;n ¼ b1;j ðpricej;s;n þ b2;j SatAGj;s;n þ b3;j SatRGj;s;n
rived in preference space. The instability is avoided as the need þ b4;j FatAGj;s;n þ b5;j FatRGj;s;n þ b6;j SugAGj;s;n
to derive estimates based on the ratio of random variables, which
are by construction volatile, is no longer necessary. The instability
þ b7;j SugRGj;s;n þ b8;j SaltAGj;s;n þ b9;j SaltRGj;s;n
of WTP estimates derived in preference space has been found to be þ b10;j StatusQuoj;s;n Þ þ ej;s;n ð3Þ
particularly problematic when the payment vehicle (price or cost
attribute) is variable and not bounded above zero. As a result this where price is the cost of the basket, Sat is saturates, Fat is fat, Sug is
has lead researchers to fix the payment coefficient (e.g., Bond et al., sugar and Salt is salt. For each of the nutrients we have estimated
2008) which is an ad hoc approach to resolving instability. Indeed, the change from Amber to Green (AG) and Red to Green (RG). These
fixing the price coefficient may violate other modelling require- parameters provide us with a measure of how much our survey
ment as well as being behaviourally inappropriate if we assume respondents are WTP to reduce their exposure to higher levels of
that individuals’ responses vary independently of socio-economic the various nutrients. The StatusQuo captures Option 1 in all choice
characteristics. sets which we devised based on an examination of current con-
To implement the Bayesian approach to estimation of the MXL sumption activity and the levels of the various nutrients being con-
we simulate the posterior distribution of the mean and variance/ sumed. Finally, each of the parameters is then conditioned on a set
covariance of the preference parameters {bj}. In the Bayesian liter- of socio-economic characteristics as discussed in Section choice
ature the algorithm used to undertake the simulation is referred to experiment design and data. In this case each parameter is ex-
as the ‘‘the sampler”. A detailed description of the algorithm, in this pressed as follows
case Gibbs with a Metropolis–Hastings (M–H) Step which is based
on Balcombe et al. (2009) is provided in the Appendix A. bij ¼ a0;i;j þ a1;i;j Genderj þ a2;i;j Agej þ a3;i;j Childrenj
For the analysis we present we have generated all posterior dis- þ a4;i;j Educationj þ ui;j ð4Þ
tributions by mapping 10,000 draws from the posterior sampler.
We have paid particular attention to the performance of the sam-
pler to ensure convergence has been achieved. To test for conver- These estimates can then be used to construct the WTP esti-
gence we initially observe the values of the parameters mates for groups by Gender, Age, Children and Education. In each
sequentially generated by the sampler. If our model is correctly case respondents are divided into two groups for each of the char-
specified and performing appropriately our parameters should acteristics: Male/Female, Young/Old, With/Without Children and
move away from their initial starting points and become stable Less/More Educated. Importantly, the error terms (ui,j in Eq. (4))
about a mean. To ensure that there is minimal dependence in the can be correlated. This would be expected because we assume that
sampled values we estimate the autocorrelation coefficients for people that responded more with respect to a particular nutrient
the sequential values generated by the sampler. To minimise (e.g., Salt) would also be more responsive with respect to the other
problems of dependence we draw every kth value (in this model nutrients also (e.g., Fat).
we took very 1 in 500 values) in a sequence generated by the Our results and the resulting WTPs are reported in Table 2.
sampler where the number of draws is set so as to minimise the
degree of dependence. Following Koop (2003) we formally test 4
Details of the likelihood and priors employed in the estimation are provided in an
for convergence by employing a modified t-test for which the null appendix to the paper.
 K. Balcombe et al. / Food Policy 35 (2010) 211–220 217

Table 3
Mixed Logit Correlation Matrix.

Price Saturates Saturates Fat Fat Sugar Sugar Salt Salt Status
coefficient (Amber) (Red) (Amber) (Red) (Amber) (Red) (Amber) (Red) quo
Price coefficient 1.000
Saturates 0.362 1.000
(Amber)
Saturates (Red) 0.216 0.288 1.000
Fat (Amber) 0.368 0.534 0.193 1.000
Fat (Red) 0.414 0.558 0.289 0.725 1.000
Sugar (Amber) 0.371 0.524 0.211 0.742 0.768 1.000
Sugar (Red) 0.379 0.564 0.228 0.699 0.754 0.764 1.000
Salt (Amber) 0.334 0.472 0.143 0.682 0.653 0.658 0.623 1.000
Salt (Red) 0.395 0.483 0.327 0.659 0.724 0.687 0.677 0.656 1.000
Status quo 0.206 0.276 0.154 0.300 0.329 0.316 0.283 0.341 0.344 1.000

In Table 2 we report the mean, standard deviation and median ing pattern for Saturates, especially RG, in that the correlation
of the resulting posterior densities. We report both the mean and estimates are much smaller than for all the other nutrients.
median, because for the indifference model these estimates Thus, reductions in nutrient consumption of Salt, Sugar and Fat
represent quite different aspects of the underlying distribution. If are not matched by anywhere near the same reduction in Satu-
over 50% of consumers are indifferent to an attribute the median rates. Overall, our correlation estimates provide support for our
might be zero, but the mean may still be (potentially) quite large. CE in terms of respondents interacting appropriately with the
As a result Balcombe et al. (2009) advocate that researchers report task required.
both the mean as well median. In the resulting analysis we focus We next report results for the same model specification includ-
on the median, although for this data set the differences are minor. ing a number of key socio-economic variables in our model which
From the results presented in Table 2 we can see that the are interacted with each of the parameters in the model as de-
change from Red to Green (RG) yields a much larger estimate for scribed by Eq. (4). The socio-economic variables we have employed
all nutrients compared to the change from Amber to Green (AG). are gender (male or female), age (young (less than 46) or old (more
This implies that there is a strong aversion to Red for all nutrients, than 46)), children in then household (yes or no), and level of edu-
which is in keeping with results reported by Drichoutis et al. cation (school or higher). By including these socio-economic vari-
(2006). We can also see that the largest mean and median esti- ables we are able to provide estimates of how these various
mates are for Salt and the smallest for Fat. This finding is interest- characteristics impact the WTP estimates.
ing given the high profile campaign employed by the FSA in the UK The first thing to note in Table 4 relates to the differences
to draw attention to salt in the diet and the associated health ef- between Men and Women. It is very clear that Men have lower
fects of excessive consumption.5 WTP estimates for all nutrients compared to Women. This im-
In addition, for Fat (AG) the median is equal to zero indicating plies that Women are WTP more to avoid nutrient quantities la-
that consumers are indifferent to the change in this level of this belled Red and Amber compared to Green. These differences
nutrient. This implies that consumers are concerned about reduc- equate to Men being WTP almost a third less than Women.
ing high levels of Fat in their diet but far less concerned about The second set of results in Table 4, refer to differences associ-
reducing Fat significantly. We can also see that the status quo ated with Age. On the whole there appears to be only a minimal
parameter estimate is positive indicating a positive preference difference between our Age categories which suggests that this
for this option. However, the relative magnitude of the estimate socio-economic characteristic is not important in terms of
is much smaller than any of the Red to Green changes for all the explaining WTP.
nutrients. The lower part of Table 4 reports results for differences in WTP
Overall we note that the resulting WTP estimates are reason- depending on if a household includes children or not. Although the
ably large. For example, the WTP estimate for moving from Red differences are not that large there is evidence to suggest that
to Green for Salt is £19. Although this is higher than we anticipated households with children are WTP more to consume food with
the magnitude of the estimate can be partly linked to the price at- lower levels of all nutrients compared to those without children.
tached to the basket of goods. It will be important in subsequent Finally, we examine how the level of education affects WTP. As
research to examine this aspect of the CE design in more detail we might expect from existing results in the literature, those
to see if modifying the price has a significant impact on the choices respondents with a higher level of education are WTP more to have
made and resulting WTP estimates. lower levels of nutrients.
Next we report the correlation structure between the regression The final part of our analysis presents results when we take ac-
coefficients. Our estimates are shown in Table 3. count of the bias in the sample of the high proportion of females
The results in Table 3 show that the regression coefficients (81%). By imposing an equal sample weight of 50% we have recal-
for the various food ingredient attributes are all negatively re- culated our WTP estimates. These results are reported in Table 5.
lated to price. Furthermore, and as we might expect, all the The revised results in Table 5 are compared against the original
nutrients are positively correlated. However, there is an interest- results reproduced from Table 2. As can be seen, by re-weighting
the sample to increase/decrease the proportion of males/females
5
The FSA have a stated objective of reducing average salt consumption of adults in we have reduced all the resulting WTP estimates. This result is
the UK to 6g a day by 2010. This objective comes from the Scientific Advisory not that surprising given the results we have previously considered
Committee on Nutrition who where concerned about excessive salt consumption and regarding Gender and presented in Table 4. However, these results
related health problems such as high blood pressure and the associated increased risk do indicate that if we wish to extrapolate from our results to those
of strokes and cardiovascular disease. To achieve this objective the FSA launched its
first Salt Campaign in September 2004, it has engaged with various food organisations
of the general public it will be necessary to ensure that we
asking them to reduce the salt content of food, and in March 2006 it published targets correctly reflect current socio-economic and demographic
for reductions in salt consumed by consumers backed up by high-profile TV ads. characteristics.
218 K. Balcombe et al. / Food Policy 35 (2010) 211–220

Table 4
MXL results – socio-economic variables.

Attributes Women Attributes Men

Mean Standard deviation Median Mean Standard deviation Median
Price 1.751 1.516 1.330 Price 1.560 1.365 1.181
Sat-A/G 0.465 0.431 0.394 Sat-A/G 0.319 0.373 0.187
Sat-R/G 1.847 0.537 1.842 Sat-R/G 1.203 0.548 1.200
Fat-A/G 0.117 0.195 0 Fat-A/G 0.121 0.197 0
Fat-R/G 1.278 0.497 1.276 Fat-R/G 0.883 0.489 0.874
Sug-A/G 0.394 0.374 0.327 Sug-A/G 0.344 0.356 0.257
Sug-R/G 1.518 0.387 1.519 Sug-R/G 0.995 0.404 0.993
Salt-A/G 0.274 0.317 0.162 Salt-A/G 0.210 0.281 0.056
Salt-R/G 2.051 0.470 2.050 Salt-R/G 1.387 0.508 1.385
SQ 0.383 0.316 0.352 SQ 0.292 0.291 0.229
Older Younger
Mean Standard deviation Median Mean Standard deviation Median
Price 1.692 1.475 1.277 Price 1.708 1.473 1.291
Sat-A/G 0.463 0.432 0.392 Sat-A/G 0.406 0.415 0.308
Sat-R/G 1.722 0.565 1.729 Sat-R/G 1.722 0.613 1.742
Fat-A/G 0.123 0.199 0 Fat-A/G 0.116 0.193 0
Fat-R/G 1.283 0.509 1.284 Fat-R/G 1.124 0.517 1.125
Sug-A/G 0.401 0.376 0.341 Sug-A/G 0.354 0.360 0.273
Sug-R/G 1.498 0.416 1.509 Sug-R/G 1.346 0.444 1.368
Salt-A/G 0.278 0.317 0.172 Salt-A/G 0.246 0.304 0.117
Salt-R/G 2.038 0.501 2.041 Salt-R/G 1.826 0.550 1.854
SQ 0.415 0.323 0.387 SQ 0.323 0.298 0.275

Children No children
Mean Standard deviation Median Mean Standard deviation Median
Price 1.853 1.584 1.402 Price 1.652 1.414 1.255
Sat-A/G 0.419 0.421 0.329 Sat-A/G 0.444 0.429 0.363
Sat-R/G 1.926 0.566 1.933 Sat-R/G 1.625 0.587 1.641
Fat-A/G 0.112 0.190 0 Fat-A/G 0.123 0.207 0
Fat-R/G 1.267 0.508 1.270 Fat-R/G 1.172 0.521 1.176
Sug-A/G 0.350 0.358 0.267 Sug-A/G 0.396 0.372 0.334
Sug-R/G 1.495 0.414 1.504 Sug-R/G 1.385 0.446 1.405
Salt-A/G 0.270 0.317 0.155 Salt-A/G 0.259 0.309 0.143
Salt-R/G 2.037 0.510 2.055 Salt-R/G 1.876 0.550 1.908
SQ 0.333 0.304 0.284 SQ 0.381 0.318 0.346
Higher education School education
Mean Standard deviation Median Mean Standard deviation Median
Price 1.693 1.458 1.276 Price 1.697 1.461 1.280
Sat-A/G 0.479 0.438 0.409 Sat-A/G 0.411 0.416 0.316
Sat-R/G 1.901 0.591 1.910 Sat-R/G 1.633 0.574 1.656
Fat-A/G 0.103 0.183 0 Fat-A/G 0.125 0.200 0
Fat-R/G 1.212 0.515 1.211 Fat-R/G 1.192 0.517 1.195
Sug-A/G 0.382 0.369 0.309 Sug-A/G 0.377 0.369 0.304
Sug-R/G 1.493 0.433 1.509 Sug-R/G 1.388 0.435 1.406
Salt-A/G 0.285 0.321 0.182 Salt-A/G 0.254 0.308 0.139
Salt-R/G 2.062 0.530 2.082 Salt-R/G 1.861 0.533 1.893
SQ 0.367 0.315 0.329 SQ 0.369 0.316 0.332

Table 5
MXL results indifference model – re-weighted for gender.

Attributes Re-weighted sample Base sample

Posterior parameter estimates Posterior parameter estimates
Mean Standard deviation Median Mean Standard deviation Median
Price 1.656 1.441 1.256 1.720 1.504 1.297
Sat-A/G 0.392 0.402 0.290 0.437 0.424 0.355
Sat-R/G 1.525 0.542 1.521 1.727 0.591 1.741
Fat-A/G 0.119 0.196 0 0.120 0.197 0.000
Fat-R/G 1.081 0.493 1.075 1.208 0.519 1.207
Sug-A/G 0.369 0.365 0.292 0.381 0.372 0.311
Sug-R/G 1.256 0.395 1.256 1.422 0.437 1.438
Salt-A/G 0.242 0.299 0.109 0.263 0.310 0.151
Salt-R/G 1.719 0.489 1.717 1.932 0.540 1.959
SQ 0.337 0.303 0.291 0.366 0.316 0.323

Summary and conclusions our results indicate a very strong preference on the part of UK con-
sumers to reduce the quantity of any nutrient associated with a
In this paper we have developed and analysed a CE to examine Red Light. From this response we can conclude that the role of
how consumers respond to the TLS introduced by the FSA. Overall the TLS to inform consumers appears to be understood. We have
 K. Balcombe et al. / Food Policy 35 (2010) 211–220 219

also found that consumers are most concerned by Salt and Satu- Given this notation we can we can define the probability that an
rated Fats when it comes to judging nutrient content and much individual j will make a given set of choices. Formally,
less so by Fat and Sugars. In addition, we have identified that par- 0x
!yj;s;n
ticular parts of the population respond differently to the TLS in a
N Y
Y Sn
etðbj Þj;s;n
pj ¼ PS n tðbn Þ0x
manner that we would expect from previous research on the use
s¼1 e
j;s;n
n¼1 s¼1
of nutrition labels.
Overall we consider these findings provide support for the use Therefore, we can express the likelihood function for the
of the TLS in terms of consumer understanding and planned shop- choices made (given a selection of a model t) as:
ping behaviour. If the policy agenda is to encourage consumers to Z !
Y
J
move away from purchasing items with very high levels of nutri- LðHjD; tÞ ¼ pj dFðBjH; ZÞ
ents, then this is a significant finding. However, it is difficult to B j¼1
know if the hypothetical behaviours identified will be replicated
Thus, the posterior distribution of the parameters pðHjD; tÞ (in a
in actual shopping behaviour. This is, of course a problem with
Bayesian framework) is:
all stated preference methods and studies. Currently, anecdotal
evidence suggests that this is the case but further research combin- pðHjD; tÞ / LðHjD; tÞpðHÞ
ing revealed and stated preference data (data from actual purchase
behaviour and data generated by a hypothetical CE) could signifi- Where pðHÞ is the prior distribution for the parameters that are
cantly enhance our understanding of this issue. In addition, it is un- independently normal for the parameters ac and ab and Inverse
clear if consumers are really responding to the information content Wishart (IW) for X. More specifically for the prior for a we assume
associated with the TLS or whether we are simply observing a deci-
ða0c ; a0b Þ0 ¼ a  fN ðajl; A0 Þ
sion based on the colour scheme used. Further research to examine
the extent to which information as opposed to colour of the TLS where l is the mean and A0 is a diagonal matrix. With respect to the
guides decisions is worthy of further consideration. Indeed, it prior for X
may well be the case that a key difference between the TLS and
other schemes are that the colours not only provide information X  fIW ðXjT 0 ; m0 Þ
but convey a strong normative message. This would go some
For the priors we employ, the hyper parameters, which are set a pri-
way to explaining the very high WTPs we observe.
ori, are l, A0, T0 and m0. Our choice of priors is determined by refer-
ence to the literature as well as our own experiments with the data
Acknowledgements
to check for robustness of results generated.
We acknowledge the financial support provided by the Food
Standards Agency in helping us to undertake this research. We References
are appreciate the extensive comments made on earlier versions
of this paper by three anonymous referees as well as seminar par- Balcombe, K., Chalak, A., Fraser, I.M., 2009. Model selection in the Bayesian
ticipants at the University of Kent, University of Portsmouth, Uni- mixed logit. Journal of Environmental Economics and Management 57 (2),
226–237.
versity of Bradford La Trobe University, University of Sydney, and Berning, J.P., Chouinard, H., McCluskey, J.J., 2008. Consumer Preferences for Detailed
participants at the 12th EAAE Congress in Ghent. Versus Summary Formats of Nutrition Information on Grocery Store Shelf
Labels, Journal of Agricultural and Food Industrial Organization, Volume 6,
(Article 6), <[Link]
Appendix A. Likelihood and priors Black, A., Rayner, M., 1992. Just Read the Label, The Stationary Office, London.
Bond, C.A., Thilmany, D.D., Keeling Bond, J., 2008. What to choose? the value of label
claims to fresh produce consumers. Journal of Agricultural and Resource
Following Balcombe et al. (2009) we can assume that the
Economics 33 (3), 402–427.
parameters bj are ordered. Thus, they may contain fixed parame- Children’s Food Campaign 2007. Missing the Target, October (www.
0
ters c0j as well as random parameters bj B0j ¼ ðc0j bj Þ. [Link]).
Both sets of parameters can be conditioned on appropriate so- Cowburn, G., Stockley, L., 2005. Consumer understanding and use of nutrition
labelling, a systematic review. Public Health Nutrition 8 (1), 21–28.
cio-economic variables for any individual where preferences are Drichoutis, A.C., Lazaridis, P., Nayga, R.M., 2005. Nutrition knowledge and consumer
determined by a vector zj, which is a (h  1) vector of variables. use of nutritional food labels. European Review of Agricultural Economics 32
Thus, defining Z j ¼ Ik  zj ; the components of b0j are defined as: (1), 93–118.
Drichoutis, A.C., Lazaridis, P., Nayga, R.M., 2006. Consumers’ Use of Nutritional
cj ¼ Z 0j ac Labels: A Review of Research Studies and Issues, Academy of Marketing Science
Review, 9.
Feunekes, G.I.J., Gortemaker, I.A., Willems, A.A., Lion, R., van den Kommer, M., 2008.
bj ¼ Z 0j abþuj Front-of-Pack Nutrition Labelling: Testing Effectiveness of Different Nutrition
Labelling Formats Front-of-Pack in Four European Countries, Appetite, 50, pp.
57–70.
FSA, 2007. Front of Pack Nutritional Signpost Labelling Technical Guidance, Issue 1,
where uj is a independently and identically normally distributed January 2007. <[Link]
vector with variance covariance matrix X. [Link]>.
Gracia, A., Loureiro, M., Nayga, R., 2007. Do consumers perceive benefits from the
Next define the set of all stated choices by an individual as implementation of a EU mandatory nutritional labelling program? Food Policy
Y = {yj,s,n}j,s,n, the set of characteristics describing all respondents 32 (2), 160–174.
by Z = {zj}j, the set of options given to the jth individual is Grunert, K.G., Wills, J.M., 2007. A review of European research on consumer
response to nutrition information on food labels. Journal of Public Health 15,
Xj = {xj,s,n}s,n, the set of all option sets given to all respondents is
385–399.
X = {Xj}j and the set of all data is D = {Y,Z,X}. Finally, we describe HMSO (2007). Health and Wellbeing – Healthy Eating ([Link]).
the collection of all parameters describing the model as Jetter, K.M., Cassady, D.L., 2006. The availability and cost of healthier food
H = (a,X) the set {bj}j denoted as B refers to B as latent data. Note, alternatives. American Journal of Preventative Medicine 30 (1), 38–44.
Kim, S.Y., Nayga Jr., R.M., Capps Jr., O., 2000. The effect of food label use on nutrient
for notational convenience the multiple integral intakes: an endogenous switching regression analysis. Journal of Agricultural
Z Z Z and Resource Economics 25 (1), 215–231.
...... db1 . . . . . . . . . dbn is expressed as dB: Koop, G., 2003. Bayesian Econometrics, Wiley Publishers, West Sussex, England.
bn b1 B Lang, T., 2006. Food, the law and public health: three models of the relationship.
Public Health 120, 41-40.
220 K. Balcombe et al. / Food Policy 35 (2010) 211–220

Loureiro, M., Gracia, A., Nayga, R., 2006. Do consumers value nutritional labels. Teratanavat, R., Hooker, N., 2006. Consumer valuations and preference hetero-
European review of Agricultural Economics 33 (2), 249–268. geneity for a novel functional food. Journal of Food Science 71 (7), S533–
Mhurchu, C.N., Gorton, D., 2007. Nutrition labels and claims in New Zealand and 541.
Australia: a review of use and understanding. Australian and New Zealand Variyam, J.N., 2008. Do nutrition labels improve dietary outcomes? Health
Journal of Public Health 31 (2), 105–112. Economics 17, 695–708.
National Statistics 2007. Family Food in 2005–06, A National Statistics Publication Verbeke, W., 2005. Agriculture and the food industry in the information age.
by DEFRA, London. European Review of Agricultural Economics 32 (3), 347–368.
Synovate, 2005. Quantitative Evaluation of Alternative Food Signposting Concepts, Wansink, B., Chandon, P., 2006. Can ‘‘low-fat” nutrition labels lead to obesity?
Reported prepared for the FSA, November. Journal of Marketing Research XLIII 605, 617.