Effects of carbon reduction labels: evidence from scanner data.
Kortelainen, Mika ; Raychaudhuri, Jibonayan ; Roussillon, Beatrice 等
Effects of carbon reduction labels: evidence from scanner data.
I. INTRODUCTION
Households in the EU are responsible for 25% of total EU greenhouse
gas emissions. (1) In an effort to reduce household greenhouse gas
emissions, the Carbon Trust Fund in the United Kingdom has introduced a
new product label called the carbon reduction label for many common
household goods. This carbon label shows the approximate number of grams
of carbon dioxide that a product generates during its life cycle, that
is, as the product is grown or manufactured, transported, stored, and
used (Figure 1). More than 27,000 goods (or services) in the United
Kingdom now carry this label and it is estimated that the label appears
on goods worth 3.3 billion pounds in annual sales. (2) The objective of
these carbon labels is to move households' behavior toward lower
amounts of carbon consumption. (3) To examine if the carbon label is
effective one could test if households are willing to pay more for goods
that have a carbon label or a lower carbon footprint (less carbon
dioxide emissions over the lifetime of the good). If consumers are
willing to pay more for carbon-labeled (or low carbon footprint) goods,
there is an incentive for firms to lower the carbon footprint of their
goods, label them accordingly, and charge a higher price. So an indirect
test of the effectiveness of the carbon reduction label is the emergence
of a higher price (or a price premium) for goods that have the carbon
label vis-a-vis other similar goods that do not have the carbon label.
In this study, we investigate the effectiveness of the carbon reduction
label using real market data from a major supermarket chain in the
United Kingdom. In particular, we examine whether a specific category of
carbon-labeled goods--carbon-labeled detergents--obtain a price premium
compared to detergents without the label.
This study contributes to the literature on carbon labeling in two
important ways. First, in contrast with previous empirical studies on
carbon labels, our data are not limited to a specific location or to a
specific store but are based on observed consumer behavior in the whole
of the United Kingdom. Second, in our data we observe transaction prices
for labeled and unlabeled detergents before and after the labeling
started. This feature of our data allows us to utilize standard
microeconometric techniques (elaborated below) to tease out the average
treatment effect. Our empirical analysis concentrates on the impact of
the carbon reduction labels on detergent prices as we do not have either
the aggregate sales data for individual products or the data on
customers' purchases in stores of other supermarket chains.
However, as a robustness check we also estimate simple demand models. In
contrast to previous experimental studies on carbon labels (discussed
below), in the United Kingdom (and therefore also in our data) the
carbon labels used are complicated and include detailed information on
the carbon dioxide emissions of the products (i.e., the number of grams
of C[O.sub.2] emissions). Given this detailed information and the fact
that people's buying behavior may be different in a market setting
than in the laboratory, it is interesting to investigate the impact of
the carbon label with real market data. Real market data also allow us
to account for the effect of search costs, which are typically (or
implicitly) assumed to be zero in the laboratory and in choice
experiments. Recent work by Seiler (2013) shows the presence of high
search costs in the detergent market and hence we would expect search
costs to have an effect on the willingness to pay for carbon-labeled
detergents in our case as well.
We make use of two methods to test for the emergence of a higher
price for carbon-labeled detergents. The first method is a standard
difference-in-differences (DID) regression that takes advantage of the
fact that some of the detergents were labeled sometime after our data
starts. This method allows us to estimate the average impact of the
carbon reduction label on the detergent prices. However, as the impact
of the carbon reduction label can be different for products with
different carbon footprints (i.e., products with different carbon
labels), it is also important to investigate whether treatment effects
vary across labeled products (i.e., if we have heterogeneous treatment
effects). To address this question, we use a relatively new technique
called the synthetic control method. (4) We use this method to estimate
the counterfactual price trajectories for each labeled product
individually. We then compare the price trajectories of the
counterfactual detergents with real carbon-labeled detergents. We also
estimate simple demand models (in a DID setup) to examine the impact of
the carbon reduction label on the sales of carbon-labeled detergents.
The results we get from the DID regressions show that on average the
carbon reduction label has no effect on price, that is, there is no
price premium for detergents that have a carbon label. We do not find
any demand impacts for the carbon reduction label either, although we
note that the results of the demand models might be sensitive to the
sample that we use in estimation. Finally, the results obtained using
the synthetic control method indicate that there is no evidence that
prices would had been higher/lower for products with low/high level of
carbon emissions compared to the corresponding counterfactual products
without the label.
Our study is one of the first empirical papers to systematically
study the effects of a carbon label using real market data in a
quasi-experimental setup. Although there is a relatively large
theoretical and empirical literature on labeling, there have been very
few empirical studies investigating the carbon label and its
effectiveness. This lacuna may result from the fact that carbon labels
were introduced only a few years ago. In recent years, a few
experimental studies have investigated the effectiveness of carbon
label. Using a simple experiment Michaud, Llerena, and Joly (2013) find
a significant price premium for low carbon footprint products. However,
the choice setting that the study exploits in the experimental design is
very different from a real life purchase choice and therefore the
external validity of the results can be weak. (5) Also, using a conjoint
choice experiment (included in a 2008 U.S. survey), Onozaka and Mcfadden
(2011) find some evidence to show that labels which signal carbon
intensity of a product can have a negative impact on the effectiveness
of other environmental labels. Finally, Vanclay et al. (2011) study the
effectiveness of a traffic light style carbon label placed on the
shelves in one grocery store in Australia. Interestingly, they find the
shelf labels to have a small positive impact on the sales of the least
carbon intensive products and a negative impact on the sales of most
carbon intensive products during an 8-week follow-up period. Although
their results may capture real market behavior, the study is limited in
scope and duration. Moreover, as the experimental design they use is not
very rigorous, it is difficult to evaluate the impact of the labels
based on their results. (6)
The rest of the article is structured as follows: Section II
presents a literature review and our setting which help to delineate the
different types of price premium that we could possibly observe in the
data; Section III describes the data for the paper and the methods used
in the empirical analysis; Section IV gives the results of the empirical
analysis, while Section V discusses the results and policy implications;
and Section VI concludes.
II. LITERATURE REVIEW AND SETTING
In this section, we review the main theoretical and empirical
literature on the effects of labeling. In addition, we outline some
factors that may affect the effectiveness of environmental labels and
discuss how these factors pertain to our setting.
A. Theoretical Results
The theoretical literature predicts that the introduction of a
carbon label should lead to a price premium (in our case a "carbon
premium"). The intuition behind this result is the following; a
consumer gets higher utility from consuming a more environmentally
friendly good, which leads to a higher willingness to pay for that good
and in turn leads to a higher price for that good.7 The label allows
firms to either segment the market or differentiate their products. More
specifically, a large number of theoretical models mainly use the Mussa
and Rosen's (1978) utility function or its variant (Bonroy and
Constantatos 2015): (8)
(1) U ([theta]) = [theta]q - p,
where 0 represents the consumer willingness to pay for the product
and it is assumed to be uniformly distributed in [0; 1], while q
represents the product quality or environmental friendliness and p the
product price.
In the above setup, we can expect the introduction of a label to
lead to an increase of the labeled product price independent of market
conditions or on the level of competition. We now go through three
distinct cases, where a price premium can emerge for labeled products.
First, it is well known that the introduction of a label in a
competitive market will segment the market into two: the labeled market
and the unlabeled market. In the labeled market, demand will increase
(consumers have higher willingness to pay for environmentally friendly
product) and supply may decrease (generally only few firms can comply
with the label requirements) resulting in a higher price for the labeled
product (see Bonroy and Constantatos 2015; Mattoo and Singh 1994; Sedjo
and Swallow 2002; Zago and Pick 2004). In the unlabeled market, if the
number of firms remaining in the market is higher than the number of
consumers still willing to buy the unlabeled product, then the price
will decrease (otherwise the price can remain constant or increase).
Second, a monopoly can also choose to segment their market due to the
label. In this case, monopoly could switch to a mass marketing strategy
where it provides one product for the whole market or to a segmentation
strategy where it produces two different variants of a good: a green and
a polluting variant. The segmentation strategy allows the monopoly to
exploit more consumer surplus by imposing a higher price for the green
(labeled) product. If the fixed costs of producing the green product are
not too high then the segmentation is always the most profitable
strategy and results in a price increase for the labeled product. In
this case, the price of the unlabeled product can decrease or remain the
same depending on potential cannibalization (Chen 2001). (9) Third, the
introduction of the label allows for vertical product differentiation
(Bonroy and Constantatos 2015). (10) In a duopoly, facing price
competition and selling homogenous products, the Bertrand paradox will
lead to price equaling marginal cost with no profit for either of the
two firms. After the introduction of the label, price competition among
firms may decrease because of vertical product differentiation with the
green product corresponding to the high quality product. In this case,
the price of the labeled and the unlabeled product will both increase.
However, the price increase of the labeled product will be much higher
(more than twice in the most simple case with zero marginal cost) than
the price increase of the unlabeled product (for a literature review
presenting these different models, see Bonroy and Constantatos 2015 and
Mason 2013). It is also worth noting that an environmentally friendly
product is typically considered costlier to produce than its polluting
equivalent. The labeling cost can be divided in two parts: the first
part is a compliance cost as the product should achieve a certain
environmental quality in order to get the label; and the second part
corresponds to the certification cost coming from the monitoring and
testing of the product. A possible increase in cost has to be
compensated by an increase in price.
There are also theoretical reasons to expect small or even zero
price premiums. If we relax the assumption that the consumers fully
(blindly) trust and understand the label then the emergence of a price
premium is less likely as we explain in the following three scenarios
below. First, consumers may believe that firms can fraudulently label
polluting or environmentally unfriendly products, in which case price
premiums may fail to emerge. In competitive markets, firms have a high
incentive to fraudulently affix a label on their polluting products (see
Hamilton and Zilberman 2006). Second, the type of certification matters
for the emergence of a premium. Self-labeling may produce smaller price
premiums than third party certification as the former is less credible.
Finally, consumers may be unsure about the value of the label. They may
know it means that the firms have passed the certification standard
requirements, but they probably will not know the difficulty of these
requirements. In such a case, consumers may use the number of firms
using labeling to estimate the difficulty of having the label and thus
its value. The price premiums can therefore be decreasing in the number
of products which have the label (Harbaugh, Maxwell, and Roussillon
2011).
To conclude, the emergence of price premiums for a labeled product
is to be expected in most situations. (11) The situations where the
emergence of a premium is unlikely include the case of a competitive
market, where one can have fraudulent labels, or when there are a large
number of firms using the label. In any case, it should be noted that
the results elaborated above depend on a number of specific assumptions
that may not be true in reality.
B. Price Premiums in Practice
Empirically, the emergence of a price premium for an environmental
label and the magnitude of the premium conditional on its emergence
depends on the following three factors:
1. Consumers' valuation of the environmental characteristic.
2. Consumers' awareness of the label. A consumer needs to look
for the label resulting in a search cost.
3. Consumers' understanding of the label. This depends on
consumers' cognitive ability to process the information on the
label.
Regarding consumers' valuation of the environmental
characteristic, many studies using the hedonic approach have found
considerable price premium for organic products (see, e.g., Griffith and
Nesheim 2013 or Nimon and Beghin 1999). However, organic products are
often considered as a tastier and healthier alternative to their
nonorganic counterparts and therefore they incorporate some private
benefit as well as attributes of a public good type. Note that in
general taste and nutritional aspects of the good are much more
important for consumers than the environmental characteristics of the
good (see Bougherara and Combris 2009; Fletcher and Downing 2011; Gadema
and Oglethorpe 2011; Griffith and Nesheim 2013). (12)
For the specific case of the carbon label, Michaud et al. (2013)
and Vanclay et al. (2011) have found a price premium using an
experimental approach suggesting that consumers value products with a
low carbon footprint. (13) In a recent survey, Hartikainen et al. (2014)
find positive attitudes toward carbon labels in an online-survey of
1,010 Finnish consumers and report that "90% stated that a carbon
footprint would have at least a little impact on their buying decision,
but the information became meaningful only when many other purchasing
criteria (such as price and taste) were satisfied."
Consumers' difficulty in noticing the label, which is
typically more likely to be the case in real markets than in
experimental and stated preference settings, appears to be an important
factor in the emergence of a price premium (Rubik and Frankl 2005;
Thogersen 2000). Related to this, Noussair, Robin, and Ruffieux (2004)
show in an experimental framework that consumers may not read the label
and thus buy GMO products despite their claimed animosity toward these
products. In our case, the carbon reduction label is placed at the back
of the detergent, which may affect the salience of the label.
Finally, the manner in which information about environmental
quality is communicated to the consumer also appears to matter for the
emergence of a price premium. Information about the environmental
quality of a good can be of many types. The two most common types of
indicators of environmental quality are (1) simple labels of approval
(e.g., an eco-label such as the EU flower or the Nordic Swan), and (2)
labels showing detailed information on the product in the same way as
nutritional information (e.g., energy cards or the information showing
the percentage of material made from recycled materials). An eco-label
informs the consumer that the product is complying with a certain
standard of environmental quality. For labels which involve detailed
environmental information, consumers can observe the exact
"amount" of an environmental attribute usually expressed in
numbers and possibly a scale to determine whether the product is
environmentally friendly or not (rather like the Guideline Daily Amount
for nutritional information). Often, these labels are mandatory, which
means all the products in the same product category have to be labeled.
Several studies using experimental field data have documented that more
information is not always better and that consumers prefer simpler
information to more detailed information (see BIO Intelligence Service
2012; Kiesel and Villas-Boas 2013; Teisl, Rubin, and Noblet 2008;
Wansink and Chandon 2006; Wansink, Sonka, and Hasler 2004).
In our case, the carbon label is voluntary, which is why only some
of the products in a product category have the label. The specific form
of the carbon label used on detergents is called the carbon reduction
label and it indicates the approximate amount of C[O.sub.2] emissions
generated by the labeled product or detergent with the sentence:
"We have committed to reduce this carbon footprint.'" In
addition, the label indicates the carbon footprint of a labeled product
in the same product category (see Figure 1). (14) Previous research
suggests that consumers appear to have difficulties in understanding
this label. Based on the survey of 428 UK shoppers, Gadema and
Oglethorpe (2011) found that "81% either strongly agreed or agreed
that understanding carbon footprint information and comparing carbon
footprints was difficult and confusing." Indeed, by reading the
carbon reduction label on a single product, the consumer knows the
C[O.sub.2] emissions generated by the labeled product, but does not know
whether this amount of C[O.sub.2] emissions is environmentally friendly.
In order to understand the label and to find the number of labeled
products and their associated C[O.sub.2] emissions, the consumer needs
to review all the products within the product category. Even if all of
this information can be collected by the consumer, he/she does not
necessarily have a scale or a reference point to understand this
information. Thus, given the particular form of the carbon reduction
label, it is difficult to predict its impact on the market and on
prices. However, based on the theoretical and empirical studies on
environmental labels, we can outline the following scenarios or
predictions:
1. If consumers value the carbon reduction label and interpret it
perfectly, we would expect to find a price premium that varies among
different labeled products.
2. If consumers have limited ability and use the label as a proxy
for environmental quality, we would expect all labeled products
(detergents) to obtain the same price premium. (15)
3. If consumers find it too complicated to assess the labels, we
would expect to find no premium at all for any labeled product.
III. DATA AND EMPIRICAL APPROACHES
A. Data
For our empirical analysis, we utilize a unique data set based on a
noted supermarket chain's scanner data. The data consist of
detailed purchase information on Clubcard account holders of the
supermarket chain, 60,000 customers in total. This sample is a
representative (random) sample for all the Clubcard account holders of
this supermarket chain in the United Kingdom. For these customers, we
have detailed information on product sales and daily transaction prices
of 339 distinct products. Among these products, there are 43 detergents,
the names of which are given in Table l. (16) Of these 43 detergents,
only five detergents (shown in bold in Table 1) are carbon labeled. All
of these labeled detergents belong to the supermarket brand that has
many detergents in the unlabeled category as well. (17) These
carbon-labeled products have the following carbon footprints: (4) 700 g
of C[O.sub.2] per wash, (7) 750g of C[O.sub.2] per wash, (17) 850 g of
C[O.sub.2] per wash, (32) 700 g of C[O.sub.2] per wash, and (41) 600 g
of C[O.sub.2] per wash. The label given on the back of the product
package informs customers of the amount of C[O.sub.2] emissions produced
during the product's life cycle on average and demonstrates a
commitment to reduce the detergent's carbon footprint (Figure 1).
In addition, the label gives information on the carbon footprint of a
benchmark product and advice on how customers could reduce their carbon
footprint even further, for example, by reducing the washing
temperature.
Our data consist of item-level transactions for detergents for
60,000 customers for a period of 104 weeks. The data consist of prices
for these detergents and categorical dummies for a number of product
attributes like the type of detergent, a supermarket brand dummy (i.e.,
a dummy that indicates if the detergent is of the same brand as the
supermarket chain) as well as other product attributes like the size of
the detergent. In addition, we also have detailed information on the
expenditure on the detergent and whether the detergent was bought on a
price discount or whether the price at which the detergent was bought
was marked down. (18) We note that it is particularly important to
control for promotions in our specification because the effect of
promotions is time-varying and typically varies across products. We also
note that if we did not have access to transactions data on individual
products then it would not be possible to control for promotions.
For tractability, we collapse (or aggregate) the transactions level
data to weekly level data. Besides balancing the data, the use of weekly
level data allows us to reduce the autocorrelation of price observations
considerably. Our data span from financial week 17 of 2007 to financial
week 15 of 2009 (both weeks inclusive). Therefore, we have data for a
period of 104 weeks (36 weeks in 2007, 52 weeks in 2008, and 16 weeks in
2009). Note that the carbon reduction label came into effect on week 10
in May 2008, which means that the carbon reduction label on the five
aforementioned carbon-labeled detergents was available only after week
10 in 2008. This fact is important because it allows us to use a DID
estimation approach and to control for time-invariant unobserved product
characteristics both for labeled and unlabeled detergents. Table 2
reports the summary statistics for the variables used in our analysis.
B. DID Regressions
Our aim is to investigate the effect that the carbon reduction
label has on the prices of detergents that have this label. As mentioned
earlier, we use two econometric techniques to test if carbon-labeled
detergents get a higher price than unlabeled detergents--the DIDs and
the synthetic control method.
Our first method, the DID approach, is an improvement over the
traditional hedonic method that is usually used in the extant literature
to isolate the effect that an environmental label has on the price of a
good. The conventional hedonic approach, using cross-sectional data,
isolates the effect that an environmental label has on the price of a
good by regressing the price of a good on a number of characteristics of
the good including a dummy for whether a good has a label. However, in
the cross-sectional setting, the hedonic method cannot generally be used
to estimate the causal impact of the label (or the environmental
quality) but only to obtain the degree of correlation between the label
and the price of a product (see, e.g., Bajari and Benkard 2005). This is
because, typically, there are unobserved factors (product
characteristics, etc.) that are correlated both with the product label
and with product prices making the label an endogenous characteristic.
(19)
Fortunately, the carbon label for detergents came into existence
some time after the period from when our data starts. Hence this
provides a market level quasi-experimental setting in which we can
observe labeled and unlabeled detergents both before and after the
carbon reduction labels were introduced and use these labeled and
unlabeled products as treatment and control groups in a standard DID
setup. As there is no change in any other product characteristics for
labeled and unlabeled detergents, we can use this quasi-experimental
setup to isolate the treatment effect or consumers' average
marginal willingness to pay for the carbon reduction label. Note that we
are actually measuring the average treatment effect for the treated
(ATT) which in the present setting measures the amount by which the
price of detergents with the carbon reduction label have changed
relative to what the prices of these detergents would have been without
the label. As usual, the DID estimator allows the treatment assignment
(i.e., which products are labeled) to correlate with time-invariant
product-specific factors. However, consistent estimation of treatment
effect rests on the assumption of independence of treatment assignment
and unobserved time-variant factors. We are not aware of any reasons
that would violate this assumption in the present application. As the
supermarket chain in question labeled different kinds of products with
different footprints, treatment assignment does not appear to be
systematic or favorable to the most potential (or effective) products.
Yet, it should be recognized that as only a small number of products
were labeled, the treatment effect estimate obtained with the DID method
might only be representative for the labeled products as well as for the
product category considered in the study.
C. Synthetic Control Method
In the DID specification, we test for the emergence of a price
premium in a simple label versus no label setup. The basic DID
specification is not flexible enough to allow for different labels to
have different effects on the prices of the carbon-labeled detergents.
To elucidate: in our data, the group of labeled detergent products
includes both high and low carbon footprint detergents (varying from 650
g of C[O.sub.2] emissions to 800 g of C[O.sub.2] per wash), but our DID
specification does not take this detailed information on the numerical
value of the carbon footprint into account while estimating the
treatment effect.
To allow for carbon reduction labels that have different carbon
footprints (i.e., show different numbers for the grams of C[O.sub.2]
emitted) to have different effects on detergent prices and to lend
robustness to our earlier results from the DID specification, we use the
synthetic control method following the approach outlined by Abadie,
Diamond, and Hainmueller (2010). (20) In the synthetic control method,
we construct, in turn, for each carbon-labeled detergent, an artificial
or "synthetic" product or detergent which in all other product
characteristics is as close as possible to the actual carbon-labeled
detergent except that this artificial detergent does not have the carbon
reduction label. This method is flexible enough to allow detergents with
different (low and high) carbon footprints to have different effects on
detergent prices. Another advantage of the synthetic method is that it
does not require us to assume that unobserved factors affecting price
are fixed over time or that the time trends of prices for labeled and
unlabeled detergents are the same pretreatment (as required by the DID
specification). In addition, the synthetic method is fully nonparametric
in the sense that no explicit functional form or distributional
assumptions are required.
The synthetic control method generates an artificial or synthetic
control unit using a weighted average or a convex combination of the
observed control units. (21) We treat the carbon-labeled detergent as
the treatment group (or treated unit) and the unlabeled detergents as
the control group. Our outcome of interest is the logarithmic
(normalized) price. Using the synthetic control method, we iteratively
produce synthetic controls (or construct synthetic products) for each of
the five carbon-labeled detergents. The group of detergents that
comprises the control group does not, of course, comprise any of the
five carbon-labeled detergents. After obtaining the synthetic control as
a convex combination of unlabeled detergents, we graphically plot and
compare the actual observed price trajectory (over time) of the
carbon-labeled detergent with the estimated counterfactual price
trajectory for the synthetic detergent (this is the price trajectory
that would have resulted for the carbon-labeled detergent if the
detergent had not been carbon labeled).
IV. RESULTS
A. DID Specifications
A common criticism of the DID approach is the uncertainty whether
the control group is able to faithfully reproduce the outcome that would
have been observed in the counterfactual situation in the absence of the
treatment. In our setting, this requirement translates to whether the
detergents which do not have the carbon reduction label are able to
mimic the counterfactual behavior of the carbon-labeled detergents had
these carbon-labeled detergents, not actually been carbon labeled. As we
are looking at the effect of the labeling (treatment) on detergent
prices (outcome), what we need to first ensure is that the unlabeled
detergents follow the same price trend pre-treatment as the
carbon-labeled detergents. The usual approach in the literature is to
use data from the pretreatment period(s) to show that the time trends of
the treatment (carbon-labeled detergents) and the control (unlabeled
detergents) groups are the same for the outcome variable in question. We
show such a graph in Figure 2 that plots the time trends for average
logarithmic prices (across weeks) for carbon-labeled and unlabeled
detergents. As shown in Figure 2, the price trends in the pretreatment
period are very similar for the carbon-labeled and unlabeled detergents.
(22) The graph also suggests that labeling does not have much of an
impact on the prices of the carbon-labeled detergents (the treatment
group) post-treatment.
We now present the results of the DID regressions that we use to
investigate the effects of carbon labeling on the transaction prices for
carbon-labeled detergents. Our DID specification is the following:
(2) log (price)it = [[beta].sub.0] + [[gamma].sub.3]
([CarbonLabel.sub.i]
* [Treat Period.sub.t]] + ([[beta]' [X.sub.it] +
[[beta].sub.i] + [summation over (t)] [(WeekDummies).sub.t], +
[[epsilon].sub.it]
where [CarbonLabel.sub.i] and [TreatPeriod.sub.t] are defined as
follows:
[CarbonLabel.sub.i]
{1 if detergent is carbon-labeled product 0 otherwise.
TreatPeriod,
= {1 if Week > = Week 10 in 2008 0 otherwise.
Note that we use the logarithm of normalized price as the dependent
variable. Normalization is done by dividing the (money) price of the
detergent with the number of washes the detergent has on average. This
normalization gives us the price per wash which makes different-sized
detergent products comparable. (23) In addition, we use a logarithmic
transformation for the dependent variable for the ease of interpretation
(coefficients can be interpreted as percentage changes). (24) The week
dummies [[summation] (WeekDummies), in the specification above control
for any possible exogenous time trends (expected mean change) in the log
price of detergents during the sample period that affects all detergent
products. In some regressions, the vector [X.sub.it] consists of the
following control variables [X.sub.it] = {Price Discount [Dummy.sub.it],
Marked Down [Dummy.sub.it]} Note that in the DID specification given in
Equation (2), we include product fixed effects (for product i) denoted
in the above specification as [[delta].sub.i]. The coefficient of
interest is [[gamma].sub.3], the coefficient of the interaction term
([CarbonLabel.sub.i] * [TreatPeriod.sub.t]), which shows the
differential impact of carbon labeling on the price of the
carbon-labeled detergents using the corresponding changes for all other
unlabeled detergent products as control. (25)
The results of the DID regressions are reported in Table 3. We
first report the regression results for a simple OLS specification,
where product-specific fixed effects are not controlled for in column 1.
Column 2 reports the results of the same regression with additional
controls. Columns 3 and 4 report the results of regressions which
control for product fixed effects and which control for the standard
errors involved in the estimation process in different ways. Note that
the prices of individual products are quite heavily autocorrelated over
time and also correlated within product category (including time dummies
mitigates but does not totally remove the autocorrelation). It is
important to address the problems caused by these two issues in the
estimation process. Bertrand, Duflo, and Mullainathan (2004) show that
conventional standard errors often severely understate the standard
deviation of the estimators in the DID framework. They propose using
block-bootstrapped standard errors. So in Table 3 we report the results
of the DID regression with product fixed effects using both clustered
standard errors (in column 3) and bootstrapped standard errors (in
column 4). We note that for all results reported in Table 3, we use
heteroskedastic robust standard errors. Also, all reported standard
errors are clustered at the product level. (26)
The regression results in Table 4 show that the coefficient of
[[gamma].sub.3] (the coefficient of the interaction term
[CarbonLabel.sub.i] * [TreatTime.sub.t]) is negative and nearly zero in
all the four different specifications considered. The coefficient is not
statistically significant in any of specifications considered. In
addition, when we use the bootstrapped standard errors the results are
highly insignificant. Given the small magnitude of the coefficient in
all cases, we can conclude that there is no perceptible difference in
the prices between carbon-labeled and unlabeled products after the
carbon reduction label came into effect. In other words, our results
show that the labeling does not affect the prices of carbon-labeled
detergents relative to unlabeled detergents. (27)
Based on our earlier discussion (see Section II) we think that the
small magnitude of the coefficient and the insignificant treatment
effects (for most specifications) is not surprising. However, it is
important to emphasize that zero average impact does not conclusively
show (at least for now) that the carbon reduction labels do not have any
impact on prices, because it does not rule out the possibility that some
of the labels may have had a positive effect on price and some of the
labels may have had a negative effect on price. Therefore, we need to
investigate how the labels may have affected the prices of individual
detergents.
B. Synthetic Control Approach
The regressions results in the previous section suggest that on
average there is little to no change in the price of carbon-labeled
detergents compared to unlabeled detergents. Next, we use the synthetic
control method to investigate whether one or several of the five
carbon-labeled detergents have product-specific price changes that
differ from the price changes of similar unlabeled detergents. As
mentioned earlier, we construct the synthetic control for each
carbon-labeled detergent. To this end, we use the following set of
variables as given by the vector [??] (note that this vector excludes
the treatment dummy and the dummy for the treatment period and their
interaction):
[??] = {TabletDummy, LiquidDummy, Twoinone
Dummy, PriceDiscountDummy, MarkedDown
Dummy, Numberofwashes, OwnBrandDummy}
These variables are the criteria that we use to create convex
combinations of unlabeled detergents from the control group for each
carbon-labeled detergent (in turn). (28)
In odd-numbered Tables 5-13, we show the weights that each
detergent in the control group (not carbon labeled) has in the synthetic
approximation of the actual treatment detergent (carbon labeled). To
illustrate, detergent no. 4 (Own Brand Non-Bio Liquid Wash 1.5 L as
given in the fourth entry in the list of detergents in Table 1) is a
carbon-labeled detergent. The synthetic detergent 4 comprises a convex
combination of other control or unlabeled detergents with weights given
in Table 5. Detergent 3 gets a high weight of 0.973 in this convex
combination whereas detergent 9 gets a weight of only 0.006 in this
convex combination. Note that all weights are non-negative (most of the
weights being zero) and sum to one. Also note that none of the other
carbon-labeled detergents (nos. 7, 17, 32, and 41) are in the control
group that make up the synthetic detergent. Thus, the synthetic control
method constructs the counterfactual using only the most similar control
units.
We also list the pretreatment characteristics of the actual
carbon-labeled detergent along with its synthetic counterpart for each
carbon-labeled detergent (i.e., for detergent nos. 4, 7, 17, 32, and 41)
and show these in even-numbered Tables 6-14. So, for example, from Table
6 for detergent 4 we find that while the actual detergent has 17 washes,
the synthetic detergent has 17.03 washes (and a similar interpretation
holds for other characteristics as well). Therefore, the synthetic
detergent provides a reasonable approximation to the pretreatment
characteristics of the actual detergent. We also note from the other
even-numbered tables (Tables 8-14) that for all carbon-labeled
detergents, the synthetic detergent appears to mirror the pretreatment
characteristics of the actual detergent accurately.
Next, we plot the actual and counterfactual trajectories of the
outcome of interest, namely, the logarithmic price of the actual
carbon-labeled detergent and the synthetic detergent, which shows what
would have happened if the carbon-labeled detergent had not been
labeled. We repeat the exercise for all five detergents. We show these
actual and counterfactual price trajectories for the carbon-labeled
products in Figures 3-7.
These graphs show that in the pretreatment period the price
trajectories of the counterfactual product (synthetic control) are
almost identical for observed price changes for the actual labeled
products. The only exception is the second labeled product, but even for
this case the price difference between labeled and synthetic product
appear to stay constant before and after the treatment.
In agreement with the results of the DIDs approach, the price
trajectories of the actual detergent and its synthetic control move
together very closely both pre- and post-treatment (i.e., after the
carbon label actually came into effect on the 10th week of 2008 as shown
by a vertical dotted line). This result suggests that the carbon
footprint on the detergent products did not have any effect on the
prices of these products. Importantly, this is the case for all five
labeled products, which appear to indicate that there is no price
premium for any of the carbon-labeled detergents.
V. FURTHER ANALYSIS AND DISCUSSION
A. Price Impacts
We think that the most plausible explanation for our results is
that customers find it difficult to notice, understand, and compare
carbon footprints of different detergents and therefore do not reward
carbon-labeled or less carbon intensive products with a price premium.
Our explanation is consistent with the finding of Teisl et al. (2008),
who show that price premiums are more difficult to find for labels which
have detailed information as this information is cognitively more
difficult for the consumer to process. Similarly, Wansink et al. (2004)
show that more information is not always better and their result
suggests that people generate better inferences from short claims than
from long claims on the front label. More recently, Muller and Ruffieux
(2011) show how the design of the label may affect consumer behavior. In
a laboratory experiment with 364 subjects, they find that consumer
responses to nutritional logos vary among different logos and on average
consumer response is better for those logos that simplify the message
most. Similar results are found in a report published by the European
Commission on the design of an environmental index. BIO Intelligence
Service (2012) studied consumer preferences for different kinds of label
designs using a survey of over 1,500 people in three different European
countries. The results suggest that consumers prefer a scale which can
be expressed as a color code system, such as a traffic light system.
(29) Upham, Dendler, and Bleda (2011) conducted interviews on a sample
of people asking them specific questions about their understanding of
the carbon footprint and found that people misunderstood or had
cognitive difficulties in processing the information on the label. The
results from all of these different studies support the idea that the
carbon reduction label is difficult to understand.
In the context of the carbon reduction label, the aforementioned
results would suggest that the label could be more effective if, instead
of simply indicating the level of C[O.sub.2] emissions in grams, it
would (instead) signal which detergents have a high carbon footprint and
which detergents have a low carbon footprint. This would make it more
likely for the consumer to be aware of the carbon reduction label and
also to have a scale in order to understand this information (and not
just the absolute value). These conclusions are also consistent with the
experimental findings of Michaud et al. (2013), who find that a much
simpler type of carbon label affects consumers' behavior in
experimental conditions.
Of course, it is possible that there are reasons other than
cognitive difficulties in understanding the carbon reduction label that
might explain our results. First, we note that the specific time frame
of our study is exceptional as the recorded purchases took place during
the credit crunch. The economic crisis may have tempered
pro-environmental behavior from the consumers as well as their budget
for green product purchases. Second, it is also possible that the
product category could affect the efficacy of labeling in the sense that
carbon labeling could be more effective for products with higher budget
shares or because detergents are like an "inventory" good for
which promotions and discounts play a key role. Third, and maybe most
importantly, it is possible that consumers have actually responded to
carbon labeling, but their response is not reflected in price but in the
quantity purchased. We find the last explanation quite plausible and
therefore we consider it in detail in the next subsection.
B. Demand Effects
So far we have focused exclusively on looking at the price impacts
of the carbon reduction label. It is possible that the carbon reduction
label could have had an impact on the demand of carbon-labeled products
that is not reflected in the price. Hence, it is interesting to look at
the direct demand effects of the labeling. Unfortunately, as we do not
have product-level aggregate sales data for different detergent products
but only for our sample of consumers (60,000 Clubcard account holders)
it might be challenging to uncover demand functions for the
carbon-labeled products using our data. Note that the demand estimation
is also complicated by the fact that we do not observe people's
purchases in the stores of other supermarket chains. This implies that
we do not, for example, observe whether there may have been systematic
changes in market shares of certain products or in the buying behavior
of customers. Because of these reasons, our data are less suitable for
estimating demand models than price models.
Despite these difficulties, as a robustness check we estimated
simple demand models for detergents. For these estimations, we once
again used the DID approach, but now our dependent variable is the
(logarithm of the) expenditure share of individual detergent products.
As regressors we use the same explanatory variables that we used in the
price models. Following standard demand models, we included own price,
the average price of substitutes (or detergents), and aggregate spending
on detergents as additional regressors. Note that we need to control for
these variables, because the treatment indicator is not necessarily
uncorrelated with these variables. However, our results are not
sensitive to the exclusion of these variables.
The regression results for the DID demand regressions are presented
in Table 4.
In the demand models that include fixed effects, the coefficient
estimates of price and expenditure variables are statistically
significant and have the expected signs (i.e., own price has a negative
effect and substitute price and expenditure have positive effects on the
quantity purchased). The treatment effect of the label on demand is
positive in all models, but is far from significant (when correct
standard errors are used). Moreover, numerically the estimate is small
which indicates that the demand impact on carbon-labeled detergents is
negligible. However, it is worth emphasizing that these estimation
results can be sensitive to our specific sample, which is not
necessarily a representative sample for all the customers of the
supermarket chain (but only for the Clubcard account holders). This is
not an issue in price regressions, because price effects should be
representative for all the customers and not just for Clubcard account
holders (at least when discounts are controlled for). This is why the
results of the demand estimation may be less reliable or robust than the
results we obtain on detergent prices. In any case, we think that it is
safe to say that these results strengthen our conclusion that
nonexistent price impacts originate from the consumer side and from
consumers' problems in understanding these labels.
C. Design of Carbon Reduction Labels
It is important to understand why the Carbon Trust Fund adopted the
carbon reduction label and the rationale behind the particular design of
this label. We briefly discuss below some of the reasons why this might
be so.
We believe that to reduce the carbon footprint of products, the
Carbon Trust Fund intended to design a label which supposedly would have
wide accessibility. As the carbon reduction label is a voluntary label,
it appears that the idea was that if the label was easily accessible
(and thus more attractive to firms) then it was more likely that it
would be adopted by many firms and used on a number of different
products. (30) In its current form, the carbon reduction label allows a
firm to use the carbon reduction label to certify all its products
whatever their level of C[O.sub.2] emissions. Thus, any firm can have
the label as long as it commits itself to reducing the C[O.sub.2]
emissions of its product within 2 years. In comparison, a simple label
of approval or a traffic light system can be much more financially
demanding for the firm and this high cost may become a barrier for the
adoption of these labels. (31) We think that the Carbon Trust Fund aimed
to spread the use of the carbon reduction label so that even if the
actual reduction in emissions for any product is small (as compared to,
say, a easier to understand traffic light label system) the cumulative
reduction in emissions achieved from all products (adopting this label)
taken together would mean a sufficient overall reduction in the total
level of carbon emissions.
However, we note that none of the other major supermarket chains in
the United Kingdom except this particular supermarket chain adopted
carbon reduction labels for their products. It appears that this general
lack of adoption of the label and its (consequent) lack of proliferation
has affected its efficacy. In fact, the supermarket chain in question
has recently gone on record complaining about how other supermarket
chains have not followed its example of implementing carbon reduction
labels and it is thinking of even giving up on the carbon reduction
label. (32) So why did the other supermarket chains not adopt this
label? Although the labeling process is very easy, it is still costly to
implement the label. Given this cost we believe that firms would be
willing to adopt the label only if they expect to obtain a price premium
and/or an increase in demand for the labeled products to make it
worthwhile for them to apply for the label and use it. (33) As
previously argued, a simple label of approval or a traffic light
labeling system in the front package is more likely to be noticed and is
therefore more likely to generate a price premium for the labeled
products. We believe that ambiguity as to whether a price premium would
actually emerge for labeled products has prevented other firms from
adopting the label. (34) The supermarket chain in question may have
committed itself too soon in adopting the label and so it is now keen to
roll back the label.
Another reason why the Carbon Trust Fund might have adopted the
carbon reduction label in its current form, that is, as a label which
discloses the exact level of C[O.sub.2] emissions generated by a product
(instead of having a simple label of approval or adopting a traffic
light system) could be just to educate consumers. If consumers observe
the exact number of grams of C[O.sub.2] emissions from a product, they
may become aware about the impact of their carbon consumption on the
level of C[O.sub.2] emissions released. This is similar to, say, a Guide
Daily Amount (GDA) scale, which is used to educate consumers about the
nutritional characteristic of a product. Moreover, observing the
C[O.sub.2] emissions for each product allows the consumer to compare not
only products within the same category but also products across
categories. We note though that it would probably take quite a long time
before consumers become accustomed to evaluating information about
carbon emissions in the products they consume in this way. This is
especially hard since comparison across product categories is
complicated. For example, 100g of C[O.sub.2] emissions could be the
signal of a green product in the detergent category but could signal a
brown product for apples. The value of the level of C[O.sub.2] emissions
cannot be understood only by itself but needs to be compared along a
range of other values. Therefore, we think that the use of a scale or a
traffic light could complement the disclosing of the exact amount of
C[O.sub.2] emissions. Ideally, a short front package logo could
complement more detailed information at the back and could be easier to
notice and understand. Actually, the Carbon Trust has recently
introduced two new labels which are simpler and convey less information.
These labels are more like labels of approval. (35) We think that
decreasing the cognitive cost of label comprehension could increase the
likelihood of its purchase and lead to the emergence of a price premium
while at the same time achieving consumer education.
VI. CONCLUSIONS
We have studied the impact of the carbon reduction label for prices
of detergents. We utilized detailed scanner data from a noted UK
supermarket chain recording consumers' transaction prices before
and after the introduction of the carbon reduction labels to evaluate
the effects of the labeling. Our regression results, based on a DID
approach, indicate that the carbon reduction label has no impact on
prices, that is, on average there is no premium for detergents that have
a carbon reduction label compared with detergents that do not have a
carbon reduction label. We also did not find any demand impact for the
carbon reduction label, although the results of simple demand models
need to be interpreted with caution. We also used the synthetic control
method to allow for the effect of the carbon reduction label to be
different for products with different carbon footprints. We did not find
any evidence to indicate that prices would have been different for
individual labeled products with low/high levels of carbon footprint
than for the counterfactual synthetic products without the label.
Therefore, the results from the DID regression as well as the synthetic
control method appear to outline a consistent story. The evidence
appears to be quite strong that there does not exist a price premium for
carbon-labeled detergents.
As we discussed in this study, our results may appear somewhat
surprising since one would expect that the presence of an environmental
label should lead to an increase in price when consumers value the
environmental attribute. This appears to be the case for carbon labels
in general according to several surveys. However, we believe that the
specific design of this carbon label is responsible for its lack of
success. The specific form of the label used includes detailed
information on carbon emissions and it is difficult for consumers to
process this information. It is therefore important to investigate the
effectiveness of simpler carbon labels in the future. Since our analysis
concentrated on detergents, in future research it would be also
important to study the effectiveness of carbon label for other products
or product categories.
doi: 10.1111/ecin.12278
Online Early publication October 16, 2015
ABBREVIATIONS
AIDS: Almost Ideal Demand System
ATT: Average Treatment Effect for the Treated
DID: Difference-in-Difference
GDA: Guide Daily Amount
REFERENCES
Abadie, A., and J. Gardeazabal. "Economic Costs of Conflict: A
Case Study of the Basque Country." American Economic Review, 93(1),
2003, 113-32.
Abadie, A., A. Diamond, and J. Hainmueller. "Synthetic Control
Methods for Comparative Case Studies: Estimating the Effect of
California's Tobacco Control Program." Journal of the American
Statistical Association, 105(490), 2010, 493-505.
Andreoni, J. "Impure Altruism and Donations to Public Goods: A
Theory of Warm-Glow Giving." The Economic Journal, 100,
1990,464-77.
Bajari, P., and C. Benkard. "Demand Estimation with
Heterogeneous Consumers and Unobserved Product Characteristics: A
Hedonic Approach." Journal of Political Economy, 113(6), 2005,
1239-76.
Bertrand, M., E. Duflo, and S. Mullainathan. "How Much Should
We Trust Differences-in-Differences Estimates?" The Quarterly
Journal of Economics, 119(1), 2004, 249-75.
BIO Intelligence Service. "Study on Different Options for
Communicating Environmental Information for Products." Final
Report, European Commission-DG Environment, 2012.
Bonroy, O., and C. Constantatos. "On the Economics of Labels:
How Their Introduction Affects the Functioning of Markets and the
Welfare of All Participants." American Journal of Agricultural
Economics, 97(1), 2015, 239-59.
Bougherara, D., and P. Combris. "Eco-Labelled Food Products:
What Are Consumers Paying For?" European Review of Agricultural
Economics, 36(3), 2009, 321-41.
Chen, C. "Design for the Environment: A Quality-Based Model
for Green Product Development." Management Science, 47(2), 2001,
250-63.
Cohen, M. A., and M. P. Vandenbergh. "The Potential Role of
Carbon Labeling in a Green Economy." Energy Economics, 34(1), 2012,
S53-63.
Conley, T. G., and C. R. Taber. "Inference with
'Difference in Differences' with a Small Number of Policy
Changes." Review of Economics and Statistics, 93(1), 2011, 113-25.
Donald, S. G., and K. Lang. "Inference with
Difference-in-Differences and Other Panel Data." Review of
Economics and Statistics, 89(2), 2007, 221-33.
Fletcher, J., and P. Downing. "Consumer Understanding of Green
Terms: A Supplementary Report on Consumer Responses to Environmental
Labels." A Report to the Department for Environment, Food and Rural
Affairs, Brook Lyndhurst & Icaro Consulting, London, 2011.
Gabszewicz, J. J., and J.-F. Thisse. "Price Competition,
Quality and Income Disparities." Journal of Economic Theory, 20(3),
1979, 340-59.
Gadema, Z., and D. Oglethorpe. "The Use and Usefulness of
Carbon Labelling Food: A Policy Perspective from a Survey of UK
Supermarket Shoppers." Food Policy, 36(6), 2011, 815-22.
Greenstone, M., and T. Gayer. "Quasi-Experimental and
Experimental Approaches to Environmental Economics." Journal of
Environmental Economics and Management, 57, 2009, 27-44.
Griffith, R., and L. Nesheim. "Hedonic Methods for Baskets of
Goods." Economics Letters, 120(2), 2013, 284-87.
Hamilton, S. F., and D. Zilberman. "Green Markets,
Eco-Certification, and Equilibrium Fraud." Journal of Environmental
Economics and Management, 52(3), 2006, 627-44.
Harbaugh, R., J. W. Maxwell, and B. Roussillon. "Label
Confusion: The Groucho Effect of Uncertain Standard." Management
Science, 57, 2011, 1512-27.
Hartikainen. H., T. Roininen, J.-M. Katajajuuri, and H. Pulkkinen.
"Finnish Consumer Perceptions of Carbon Footprints and Carbon
Labelling of Food Products." Journal of Cleaner Production, 73,
2014, 285-93.
Houde, S. "How Consumers Respond to Environmental
Certification and the Value of Energy Information." Working Paper
No. 20019, NBER, 2014.
Kiesel, K., and S. B. Villas-Boas. "Can Information Costs
Affect Consumer Choice? Nutritional Labels in a Supermarket
Experiment." International Journal of Industrial Organization,
31(2), 2013, 153-63.
Koos, S. "Varieties of Environmental Labelling, Market
Structures, and Sustainable Consumption across Europe: A Comparative
Analysis of Organizational and Market Supply Determinants of
Environmental-Labelled Goods." Journal of Consumer Policy, 34(1),
2011, 127-51.
Kuminoff, N., C. Parmeter, and J. Pope. "Which Hedonic Models
Can We Trust to Recover the Marginal Willingness to Pay for
Environmental Amenities?" Journal of Environmental Economics and
Management, 60, 2010, 145-60.
Mason, C. F. "The Economics of Eco-Labeling: Theory and
Empirical Implications." International Review of Environmental and
Resource Economics, 6(4), 2013, 605-17.
Mattoo, A., and H. V. Singh. "Eco-Labelling: Policy
Considerations." Kyklos, 47(1), 1994, 53-65.
Michaud, C., D. Llerena, and I. Joly. "Willingness to Pay for
Environmental Attributes of Non-Food Agricultural Products: A Real
Choice Experiment." European Review of Agricultural Economics,
40(2), 2013, 313-29.
Milgrom, P., and J. Roberts. "Relying on the Information of
Interested Parties." The RAND Journal of Economics, 17(1), 1986,
18-32.
Muller, L. and B. Ruffieux. "Consumer Responses to Various
Nutrition Front of Pack Logos: A Framed Field Experiment." Working
Papers No. 2011-05, Grenoble Applied Economics Laboratory (GAEL), 2011.
Mussa, M., and S. Rosen. "Monopoly and Product Quality."
Journal of Economic Theory, 18(2), 1978, 301-17.
Nimon, W., and J. Beghin. "Are Eco-Labels Valuable? Evidence
from the Apparel Industry." American Journal of Agricultural
Economics, 81(4), 1999, 801-11.
Noussair, C., S. Robin, and B. Ruffieux. "Do Consumers Really
Refuse to Buy Genetically Modified Food?" The Economic Journal,
114(492), 2004, 102-20.
Onozaka, Y., and D. T. Mcfadden. "Does Local Labeling
Complement or Compete with Other Sustainable Labels? A Conjoint Analysis
of Direct and Joint Values for Fresh Produce Claim." American
Journal of Agricultural Economics, 93(3), 2011, 689-702.
Rubik, F., and P. Frankl, eds. The Future of Eco-Labelling: Making
Environmental Product Information Systems Effective. Sheffield, UK:
Greenleaf Publishing, 2005.
Sedjo, R. A., and S. K. Swallow. "Voluntary Eco-Labeling and
the Price Premium." Land Economics, 78, 2002, 272-84.
Seiler, S. "The Impact of Search Costs on Consumer Behavior: A
Dynamic Approach." Quantitative Marketing and Economics, 11(2),
2013, 155-203.
Teisl, M. F., B. Roe, and R. L. Hicks. "Can Eco-Labels Tune a
Market? Evidence from Dolphin-Safe Labeling." Journal of
Environmental Economics and Management, 43, 2002, 339-59.
Teisl, M. F., J. Rubin, and C. L. Noblet. "Non-Dirty Dancing?
Interactions between Eco-Labels and Consumers." Journal of Economic
Psychology, 29(2), 2008, 140-59.
Thogersen, J. "Psychological Determinants of Paying Attention
to Eco-Labels in Purchase Decisions: Model Development and Multinational
Validation." Journal of Consumer Policy, 23(3), 2000, 285-313.
Upham, P., L. Dendler, and M. Bleda. "Carbon Labelling of
Grocery Products: Public Perceptions and Potential Emissions
Reductions." Journal of Cleaner Production, 19(4), 2011, 348-55.
Vanclay, J., J. Shortiss, S. Aulsebrook, A. Gillespie, B. Howell,
R. Johanni, M. Maher, K. Mitchell, M. Stewart, and J. Yates.
"Customer Response to Carbon Labelling of Groceries." Journal
of Consumer Policy, 34, 2011, 153-60.
Wansink, B., and P. Chandon. "Can 'Low-Fat'
Nutrition Labels Lead to Obesity?" Journal of Marketing Research,
43, 2006, 605-17.
Wansink, B., S. Sonka, and C. Hasler. "Front-Label Health
Claims: When Less Is More." Food Policy, 29, 2004, 656-67.
Zago, A. M., and D. Pick. "Labeling Policies in Food Markets:
Private Incentives, Public Intervention, and Welfare Effects."
Journal of Agricultural and Resource Economics, 29(1), 2004, 150-65.
SUPPORTING INFORMATION
Additional Supporting Information may be found in the online
version of this article:
Appendix S1. Additional results and graphs
MIKA KORTELAINEN, JIBONAYAN RAYCHAUDHURI and BEATRICE ROUSSILLON *
* We thank Chris Birchenhall, Corrado Di Maria, Farasat Bokhari,
Reyer Gerlagh, Shaun Hargreaves Heap, Anni Huhtala, Marita Laukkanen,
Franco Mariuzzo, Luca Panzone, Nicola Pavanini, Grischa Perino, Dan
Rigby, and Ada Wossink for helpful comments. We also thank conference
participants at EAERE 2012 in Prague, at EEA 2012 in Malaga, and at
EARIE 2012 in Rome as well as the seminar audience in Grenoble and
Manchester for useful discussions. The usual caveat applies. The data
used in this article are confidential, but the authors' access is
not exclusive.
Kortelainen: VATT Institute for Economic Research, Helsinki
FI-00101, Finland; Department of Economics, University of Manchester,
Manchester, UK. Phone 358295519421, Fax 358295519599, E-mail
mika.kortelainen@vatt.fi
Raychaudhuri: School of Economics, University of East Anglia,
Norwich, UK. Phone 7791740729, Fax 7791740729, E-mail
j.raychaudhuri@uea.ac.uk
Roussillon: GAEL, Univ. Grenoble Alpes; INRA, Grenoble, F-38000,
France. Phone 33476827882, Fax 3347682 5455, E-mail
beatrice.roussillon@upmf-grenoble.fr
This is an open access article under the terms of the Creative
Commons Attribution-Non-Commercial-NoDerivs License, which permits use
and distribution in any medium, provided the original work is properly
cited, the use is non-commercial and no modifications or adaptations are
made.
(1.) See the recent report published by the European Environment
Agency, which can be downloaded from the website:
http://www.eea.europa.eu/publications/end-use energy-emissions.
(2.) From the website of the Carbon Trust at: http://
www.carbontrust.com/client-services/footprinting/footprintcertification.
(3.) For a detailed discussion on carbon labeling and its potential
usefulness in reducing carbon dioxide emissions, see Cohen and
Vandenbergh (2012) and references cited therein.
(4.) Another option could be to use the DID setup and estimate
separate treatment effects for each labeled product (using interactions
or separate subsamples of these products). However, as Conley and Taber
(2011) show (see also Donald and Lang 2007), the conventional
statistical inference methods are not consistent for DID when there are
a small number of treated units and a large number of control units.
This is clearly a problem in our setting, as allowing heterogeneous
treatment effects would allow us to have only one unit in the treatment
group (for each treated product considered) and many products in the
control group. In addition, we note that as the synthetic control method
does not require the common trend assumption or any additional
assumptions about the specific type of parametric functional form used,
it is more flexible and robust than the DID setup when studying
heterogeneous treatment effects.
(5.) For instance, Michaud et al. (2013) consider only three
different product characteristics in the experimental design: a product
price, an eco-label, and a carbon footprint with two different levels
(high emissions vs. low emissions). The carbon label used in their
experiments is much simpler than the labels typically used in the real
market.
(6.) One particular weakness related to the design is that they do
not look at the changes in sales of unlabeled products.
(7.) It is typically assumed that consumers obtain a higher utility
from the consumption of green product because of a warm glow effect
(Andreoni 1990). They feel better knowing that they have contributed to
the environment protection.
(8.) See the literature review by Bonroy and Constantatos (2015)
for details. Two classic variants of this utility function are the
consideration of a degenerated distribution or the perception of the
product quality q. In the former case, we only have two groups of
consumers: one valuing the environment and the other not.
(9.) Recently, Houde (2014) has shown using a structural demand
model that firms use the certification to price discriminate their
products.
(10.) As mentioned by Bonroy and Constantatos (2015), this idea is
very close to the model of vertical product differentiation of
Gabszewicz and Thisse (1979) with the labeled product representing the
high quality product and the pollutant product representing the low
quality product.
(11.) Note that the existence of a price premium does not mean that
the introduction of a label will be welfare enhancing.
(12.) Other relevant papers include Teisl, Roe, and Hicks (2002)
and Houde (2014). Teisl et al. (2002) use the almost ideal demand system
(AIDS) model to estimate the impact of the dolphin-safe label. Their
results indicate that the dolphin-safe label affected consumer behavior
and the market share of canned tuna. However, it is very difficult to
compare this paper (or its results) to our setting, because the label
was a label of approval and it was adopted in the study period by the
three largest tuna companies in the world. Moreover, their paper also
deviates methodologically from more recent papers that look at the
impact of environmental labels (including our paper). Houde (2014)
instead studies the effects of Energy Star certification in the U.S.
appliance market. Again, the label studied is quite different in
comparison to the carbon label in the United Kingdom, because the Energy
Star label is a label of approval and identifies the most energy
efficient products in the marketplace. Using a structural demand model,
he finds that consumers respond to certification in different ways.
(13.) Note that the approach used in these studies is quite
different from ours, because experimental studies control for salience
and understanding of the label as well and they also assume away search
costs, all of which are likely to be important in our case.
(14.) One could be skeptical about whether information about the
carbon footprint of a comparable product affects consumers'
purchases. For the conscientious consumer who reads the detailed
information on each labeled product, the information about the benchmark
product does not add any new information at all, while for the consumer
who wants to save time by just looking at the logo for the product the
information about the benchmark product is probably written too small to
be noticed or they may simply not use this information anyway.
(15.) This idea (or scenario) is based on a naive version of
Milgrom's unraveling argument: the consumer observes that the firm
is disclosing something, and therefore expects or assumes the labeled
product to be better quality than the other unlabeled product (see
Milgrom and Roberts 1986).
(16.) We replace wherever appropriate in the product names given in
Table 1, the name of the supermarket chain with the phrase "Own
Brand."
(17.) During the sample period, we consider in our analysis the
supermarket chain already had six different types of products
certified/labeled: toilet paper, kitchen rolls, laundry detergents,
chilled and long-life orange juice, light bulbs, and Jaffa oranges/soft
fruit. However, only a small number of products had been labeled in each
of these product categories. The number of labeled products was smaller
for other product categories than for detergents, which is why we
decided to concentrate on detergents.
(18.) Some of these variables are used in our analysis, although we
note that in DID models time-invariant control variables or
characteristics (such as detergent type) become redundant.
(19.) For more detailed discussion on endogeneity problems in these
kind of hedonic regressions, see for example, Greenstone and Gayer
(2009) and Kuminoff, Parmeter, and Pope (2010).
(20.) Another option would be to use the DID setup and interact the
treatment group and period indicators with an indicator for each labeled
product. However, this approach has a few weaknesses at least in the
context of our application. First, it requires stronger assumptions than
the synthetic control method (common trend and functional form
assumptions). Second, the problem with this kind of regression in our
setting is that we would then have five treatments (different labels),
but only one product for each treatment. Although this kind of
regression can be estimated, statistical inference on the interaction
terms is not reliable as discussed in the introduction.
(21.) The idea behind the synthetic control method is that a
(convex) combination of control units provides a better counter-factual
for the treated unit than any single control unit alone. In our case,
labeled detergents form the treatment group while unlabeled detergents
form the control group. For K unlabeled detergents, we assign weights
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] to each of these
control detergents. The weights are chosen so that the synthetic
detergent resembles the actual carbon-labeled detergent as much as
possible. We refer the interested reader to Abadie et al. (2010) for
additional technical details and to Abadie and Gardeazabal (2003) for an
economic application.
(22.) We also drew the same graph using only the supermarket's
own products (labeled and unlabeled detergents). The price trends are
very similar for these two groups of products in this case also. For
more details, please see Appendix S1.
(23.) We also estimate the main models without using any
normalization. The results are very similar and consistent with our
results presented later. For more details, please see Appendix S1.
(24.) Again, the results and conclusions are not sensitive to this
choice.
(25.) We also consider a simple OLS regression (i.e., without
product fixed effects), where the DID specification is the conventional
specification used in the literature: log [(price).sub.it] =
[[beta].sub.0] + [[gamma].sub.1][CarbonLabel.sub.i] + [[gamma].sub.2]
[TreatPeriod.sub.t] + [[gamma].sub.3] ([CarbonLabel.sub.i] *
[TreatPeriod.sub.t]) + [[summation].sub.t], [(WeekDummies).sub.t], +
[[epsilon].sub.it]
(26.) To address the autocorrelation issue, we also use an
alternative approach suggested by Bertrand et al. (2004). We collapse
the data to two time periods (before and after the introduction of the
labels) and estimate the regression model using the collapsed data. The
results were very similar to the results in Table 3. For more details,
please see Appendix S1
(27.) For robustness, we also rerun regression 2 earlier by
limiting our sample only to those detergents that belong to the
supermarket chain. We find that the interaction term [CarbonLabel.sub.i]
* [TreatTime.sub.t] is insignificant in all specifications considered
(these are the same specifications that we used in the columns of Table
3) indicating that the impact of the carbon label is insignificant. So
our main result is again robust to using this alternative control group.
For more details, please see Appendix S1.
(28.) For the synthetic control method, we have had to drop a few
detergents for which we did not have data for all 104 weeks. For
example, for detergent number 2 we did not have data from week 98, for
detergent 12 we did not have data from weeks 24 to 47, and so on. So we
had to drop detergent numbers 2, 12, 33, 35, and 38 from the data set
used in the analysis. We also had to drop data for some periods (weeks)
for which we had data missing on the outcome of interest (log of the
average price per wash). For example, in week 1, we have data for only
37 detergents and similarly for week 2 we also have data only on 37
detergents. So we had to drop week numbers 1, 2, 3, 75, 83, 91, and 95.
(29.) It is stated in the report that: "Labels that present
the performance of a product on a comparative scale, such as star,
letters or numbers, or a color code system are vastly preferred and are
more easily understood and motivating than those that present technical
information only."
(30.) Koos (2011) shows that a larger supply of
environmental-labeled goods within the market increases the likelihood
of purchasing these goods. Indeed, the availability of these labeled
products in the supermarket is a necessary condition for the purchase of
the labeled good. Moreover, his results indicate that the larger the
share of major retailers using the label, the more likely the labeled
product is bought.
(31.) With a simple label of approval, a firm might have to make
improvements or investments in its production process to raise the
environmental quality of its products above the level imposed by the
label and this could be costly. With a traffic light label, a
firm's products could end up being classified as environmentally
unfriendly and therefore the firm could be reluctant to apply for such a
label.
(32.) See the report on the supermarket chain in the article by
Adam Vaughan in the guardian.co.uk, Monday, January 30, 2012, 15.02 GMT.
(33.) Firms endure some certification costs related to the
monitoring and assessment of the C[O.sub.2] emissions disclosed as well
as packaging costs. For instance, the noted supermarket chain claims
"a minimum of several months' work" to calculate the
carbon footprint of a product.
(34.) Harbaugh et al. (2011) show that the quality and the number
of products having a label may impact the size of the potential price
premium.
(35.) See the following website for more information:
http://www.carbontrust.com/client-services/footprinting/
footprint-certification/carbon-footprint-label
TABLE 1
List of Detergent Products
(1) Fairy Liquitabs Non-Bio 11 Wash/385 G
(2) Fairy Non-Bio Liquid Wash 1.37 L
(3) Own Brand Bio Liquid Wash 1.5 L
(4) Own Brand Non-Bio Liquid Wash 1.5 L (Carbon
Labeled)
(5) Persil Powder Non-Bio 28 Wash/2.38 kg
(6) Own Brand Powder Bio 800 G
(7) Own Brand Non-Bio 1.2 kg (Carbon Labeled)
(8) Own Brand Powder Color 800 G
(9) Own Brand Value Bio Cone Liquid Wash 1 L
(10) Fairy Powder Non-Bio 10 Wash/800 G
(11) Persil Powder Non-Bio 10 Wash/850 G
(12) Own Brand Non-Bio Tablets 24 Pk 12 Washes/900 G
(13) Persil Tablets Non-Bio 24 Pack 12 Wash/912 G
(14) Own Brand Powder Non-Bio 30 Wash/2.4 kg
(15) Own Brand Color Liquid Capsules 10 Wash/500M1
(16) Own Brand Bio Tablets 48 Pk 24 Washes/1.8 kg
(17) Own Brand Non-Bio Tablets 48 Pk 24 Washes/1.8
kg (Carbon Labeled)
(18) Own Brand Color Tablets 48 Pk 24 Washes/1.8 kg
(19) Persil Non-Bio Capsules 20 Pk 10 Wash
(20) Fairy Non-Bio Tablets 56 Pk 28 Wash/1.848 kg
(21) Persil Non-Bio Capsules 40 Pk 20 Wash
(22) Own Brand 2Inl Freshtablets 48 Pk 24 Washes/1.8 kg
(23) Persil Bio Liquigel 1.5 L
(24) Persil Non-Bio Liquigel 1.5 L
(25) Fairy Liquitabs Non-Bio 22 Wash/770 G
(26) Persil Tablets Non-Bio 48 Pack 24 Wsh 1.74 kg
(27) Own Brand Powder 2Inl Lavender 800 G
(28) Own Brand Lav 2Inl Liqd Wash 1.5 L
(29) Own Brand 2Inl Lav Tablets 48 Pk 24 Washes/1.8 kg
(30) Persil Non-Bio Small & Mighty 730M1
(31) Surf Tropical Small & Mighty 730M1
(32) Own brand Non-Bio Liquid Capsules 20 Wash/1 L
(Carbon Labeled)
(33) Own Brand Bio Liquid Capsules 20 Wash/1 L
(34) Own Brand Color Liquid Capsules 20 Wash/1 L
(35) Own Brand 2 In 1 Lavliquid Capsules 20 Wash/1 L
(36) Own Brand 2Inl Oceantablets 48 Pk 24 Wash/1.8 kg
(37) Surf Sunshine Small & Mighty 730M1
(38) Persil Non-Bio Small & Mighty 1.47 L
(39) Own Brand Super Cone Color Liqd 700M1/20 Wsh
(40) Own Brand Super Cone Bio Liquid 700M1/20 Wsh
(41) Own Brand Super Cone Non-Bio Liqd Wash
700M1/20 Wsh (Carbon Labeled)
(42) Own Brand Super Conc2In1 Lav Liqd 730M1/20 Wsh
(43) Own Brand Powder Non-Bio 42 Wash/3.36 kg
TABLE 2
Summary Statistics: Detergent Data
Variable Mean Std. Dev. Min. Max. N
Carbon label 0.118 0.323 0 1 4,369
Own brand 0.61 0.488 0 1 4,369
Powder 0.296 0.457 0 1 4,369
Tablet 0.446 0.497 0 1 4,369
Liquid 0.258 0.438 0 1 4,369
Size 1.234 0.604 0.385 3.36 4,369
Price discount 0.078 0.252 0 1 4,369
Marked down 0.002 0.009 0 0.2 4,369
No. of washes 19.44 6.884 10 42 4,369
Two-in-one 0.166 0.372 0 1 4,369
Average price per 0.168 0.057 0.044 0.362 4,369
wash
TABLE 3
Price Regressions
(2)
(1) OLS w/ controls
OLS (clustered SE) (clustered SE)
CarbonLabel * -.065 -.064
TreatPeriod (.071) (.070)
CarbonLabel -.133 -.138
(.097) (.099)
TreatPeriod .033 .019
(.031) (.031)
Price Discount -.237 ***
(.061)
Marked Down 1.381
(1.222)
Product Fixed Effects No No
Week Dummies Yes Yes
No. of Obvs. 4,369 4,369
(3) (4)
FE (clustered SE) FE (bootstrap SE)
CarbonLabel * -.069 -.069
TreatPeriod (.068) (.066)
CarbonLabel
TreatPeriod
Price Discount -.189 *** -.189 ***
(.022) (.021)
Marked Down -.695 * -.695 *
(.284) (.285)
Product Fixed Effects Yes Yes
Week Dummies Yes Yes
No. of Obvs. 4,369 4,369
Notes: Dependent variable is the logarithm of normalized
price. Normalization is done by dividing the (money) price
of the detergent with the number of washes the detergent has
on average. Independent variables are given in the rows.
Price Discount is a dummy for detergents that are offered on
a price discount. Marked Down is a dummy for detergents that
are marked down.
CarbonLabel is a dummy variable which is 1 for detergents
that are carbon-labeled and 0 for detergents that are not
carbon labeled. TreatPeriod is a dummy which is 1 for the
post-treatment period or the period after May 2008, the date
at which the carbon label came into effect, and 0 for
periods before this date or the pretreatment period. The DID
estimator is the coefficient on the interaction term
CarbonLabel * TreatPeriod. r-Statistics reported under each
coefficient in parenthesis. OLS, ordinary least squares
regression; FE, fixed effects regression; SE, standard
errors.
Significance at (+) p <0.10; * p <0.05; ** p <0.01; *** p <0.001.
TABLE 4
Demand Regressions
(1) (2)
OLS OLS w/ controls
(clustered SE) (clustered SE)
CarbonLabel * TreatPeriod .157 .021
(.197) (.140)
CarbonLabel .034 .091
(.387) (.128)
TreatPeriod -.104 .104
(.092) (.109)
Price Discount -.075
(.055)
Marked Down -1.884
(1.402)
Average Price .067 *
(.033)
Sum Expenditure .003' **
(.000)
Mean Detergent Price .060
(.106)
Product Fixed Effects No No
Week Dummies Yes Yes
No. of Obvs. 4,369 4,369
(3) (4)
FE FEw/
(clustered SE) (bootstrap SE)
CarbonLabel * TreatPeriod .014 .014
(.130) (.123)
CarbonLabel
TreatPeriod
Price Discount -.025 -.025
(.044) (.041)
Marked Down -1.200 -1.200
(1.134) (1.197)
Average Price -.141 * -.141 *
(.064) (.065)
Sum Expenditure .002 *** .002 ***
(.000) (.000)
Mean Detergent Price .338 * .659 *
(.160) (.327)
Product Fixed Effects Yes Yes
Week Dummies Yes Yes
No. of Obvs. 4,369 4,369
Notes: Dependent variable is the logarithm of the ratio of
spending on detergents for a week over the total spending on
all products for a week. Independent variables are given in
the rows. Price Discount is a dummy for detergents that are
offered on a price discount. Marked Down is a dummy for
detergents that are marked down. Average price denotes the
own price of the detergent (averaged by week). Mean
Detergent price denotes the average price of substitutes.
Sum Expenditure denotes the aggregate spending on detergents
in that week. CarbonLabel is a dummy variable which is 1 for
detergents that are carbon labeled and 0 for detergents that
are not carbon labeled. TreatPeriod is a dummy which is 1
for the post-treatment period or the period after May 2008,
the date at which the carbon label came into effect and 0
for periods before this date or the pretreatment period. The
DID estimator is the coefficient on the interaction term
CarbonLabel * TreatPeriod. t-Statistics reported under each
coefficient in parenthesis. OLS, ordinary least squares
regression; FE, fixed effects regression; SE, standard
errors.
Significance at (+) p<0.10; * p < 0.05; ** p <0.01;
*** p < 0.001.
TABLE 5 Detergent Weights in Synthetic Unit for Detergent No. 4
Treatment Detergent No. 4
Control Control
Detergent No. Weight Detergent No. Weight
1 0 23 0.002
3 0.973 24 0
5 0 25 0
6 0 26 0
8 0 27 0
9 0.006 28 0
10 0 29 0
11 0 30 0
13 0 31 0
14 0 34 0
15 0 36 0
16 0 37 0
18 0 39 0
19 0 40 0.018
20 0 42 0
21 0 43 0
22 0
TABLE 6
Log(price) Predictor Means for Detergent No. 4
Log(price) Predictor Means
Treatment Detergent No. 4
Variables Real Synthetic
Number of washes 17 17.031
Two in one dummy 0 0
Own brand dummy 1 0.997
Powder dummy 0 0
Liquid dummy 1 0.999
Tablet dummy 0 0
Discount (average) 0.0333569 0.0329327
Mark down (average) 0.0003054 0.0000147
TABLE 7
Detergent Weights in Synthetic Unit for
Detergent No. 7
Treatment Detergent No.7
Control Control
Detergent No. Weight Detergent No. Weight
1 0 23 0
3 0 24 0
5 0 25 0
6 0.662 26 0
8 0.182 27 0
9 0 28 0
10 0 29 0
11 0 30 0
13 0 31 0
14 0 34 0
15 0 36 0
16 0 37 0
18 0 39 0
19 0 40 0
20 0 42 0
21 0 43 0.156
22 0
TABLE 8
Log(price) Predictor Means for Detergent No. 7
Log(price) Predictor Means
Treatment Detergent No. 7
Variables Real Synthetic
Number of washes 15 14.992
Two in one dummy 0 0
Own brand dummy 1 1
Powder dummy 1 1
Liquid dummy 0 0
Tablet dummy 0 0
Discount (average) 0 0
Mark down (average) 0.0061858 0.006187
TABLE 9
Detergent Weights in Synthetic Unit for
Detergent No. 17
Treatment Detergent No.17
Control Control
Detergent No. Weight Detergent No. Weight
1 0 23 0
3 0 24 0
5 0 25 0
6 0 26 0
8 0 27 0
9 0 28 0
10 0 29 0
11 0 30 0
13 0 31 0
14 0 34 0
15 0 36 0
16 0.458 37 0
18 0.541 39 0
19 0 40 0
20 0 42 0
21 0 43 0
22 0
TABLE 10
Log(price) Predictor Means for Detergent
No. 17
Log(price) Predictor Means
Treatment Detergent No. 17
Variables Real Synthetic
Number of washes 24 23.976
Two in one dummy 0 0
Supermarket store dummy 1 0.999
Powder dummy 0 0
Liquid dummy 0 0
Tablet dummy 1 0.999
Discount (average) 0 0
Mark down (average) 0.0035762 0.0035729
TABLE 11
Detergent Weights in Synthetic Unit for
Detergent No. 32
Treatment Detergent No.32
Control Control
Detergent No. Weight Detergent No. Weight
1 0 23 0
3 0 24 0
5 0 25 0
6 0 26 0
8 0 27 0
9 0 28 0
10 0 29 0.005
11 0 30 0
13 0 31 0
14 0 34 0.971
15 0.014 36 0
16 0 37 0
18 0 39 0
19 0 40 0
20 0.01 42 0
21 0 43 0
22 0
TABLE 12
Log(price) Predictor Means for Detergent
No. 32
Log(price) Predictor Means
Treatment Detergent No. 32
Variables Real Synthetic
Number of washes 20 19.96
Two in one dummy 0 0.005
Own brand dummy 1 0.99
Powder dummy 0 0
Liquid dummy 0 0
Tablet dummy 1 1
Discount (average) 0 0.0004813
Mark down (average) 0.001897 0.002492
TABLE 13
Detergent Weights in Synthetic Unit for
Detergent No. 41
Treatment Detergent No.41
Control Control
Detergent No. Weight Detergent No. Weight
1 0 23 0
3 0 24 0
5 0 25 0
6 0 26 0
8 0 27 0
9 0 28 0
10 0 29 0
11 0 30 0
13 0 31 0
14 0 34 0
15 0 36 0
16 0 37 0
18 0 39 0.449
19 0 40 0.55
20 0 42 0
21 0 43 0
22 0
TABLE 14
Log(price) Predictor Means for Detergent No.
41
Log(price) Predictor Means
Treatment Detergent No. 41
Variables Real Synthetic
Number of washes 20 19.98
Two in one dummy 0 0
Own brand dummy 1 0.999
Powder dummy 0 0
Liquid dummy 1 0.999
Tablet dummy 0 0
Discount (average) 0 0
Mark down (average) 0.0005066 0.0005059
FIGURE 1 An Example of a Carbon Reduction Label
working with The carbon footprint of this
the Carbon Trust product is 850g per wash and
we have committed to reduce
850g this
CO2
per wash By comparison the carbon
footprint of non-biological
washing liquid is 600g per
wash
Help to reduce this footprint.
Washing at 30[degrees]C rather
than 40[degrees]C saves 160g
CO2 per wash
COPYRIGHT 2016 Western Economic Association International
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2016 Gale, Cengage Learning. All rights reserved.
