World Development 102 (2018) 292–303

Contents lists available at ScienceDirect

World Development
journal homepage: www.elsevier.com/locate/worlddev

Supermarket Shopping and Nutritional Outcomes: A Panel Data Analysis

for Urban Kenya
Kathrin M. Demmler a, Olivier Ecker b, Matin Qaim a
a
University of Goettingen, Germany
b
International Food Policy Research Institute (IFPRI), Washington, D.C., USA

a r t i c l e i n f o s u m m a r y

Article history: Overweight and obesity are growing health problems in many developing countries. Rising obesity rates
Accepted 10 July 2017 are the result of changes in people’s diets and lifestyles. Income growth and urbanization are factors that
Available online 5 August 2017 contribute to these changes. Modernizing food retail environments may also play a certain role. For
instance, the rapid spread of supermarkets in many developing countries could affect consumer food
Key words: choices and thus nutritional outcomes. However, concrete evidence about the effects of supermarkets
dietary choices on consumer diets and nutrition is thin. A few existing studies have analyzed related linkages with
overweight
cross-sectional survey data. We add to this literature by using panel data from households and individ-
obesity
supermarkets
uals in urban Kenya. Employing panel regression models with individual fixed effects and controlling for
panel data other factors we show that shopping in supermarkets significantly increases body mass index (BMI). We
Africa also analyze impact pathways. Shopping in supermarkets contributes to higher consumption of pro-
cessed and highly processed foods and lower consumption of unprocessed foods. These results confirm
that the retail environment affects people’s food choices and nutrition. However, the effects depend on
the types of foods offered. Rather than thwarting modernization in the retail sector, policies that incen-
tivize the sale of more healthy foods—such as fruits and vegetables—in supermarkets may be more
promising to promote desirable nutritional outcomes.
Ó 2017 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction of products than traditional markets and shops, such moderniza-

tion of the retail sector could possibly contribute to negative nutri-
Overweight and obesity are growing health problems world- tion and health outcomes (Hawkes, 2008; Popkin, 2014; Qaim,
wide. During1980–2013, the global proportion of overweight or 2017).
obese adults increased from 29% to 37% in men, and from 30% to Concrete evidence about the effects of supermarket shopping
38% in women (Ng et al., 2014). Developing countries are also on people’s diets in developing countries is thin. Very few studies
increasingly affected. The rapid rise in people’s body mass index analyzed related linkages, with mixed results. Tessier et al. (2008)
(BMI) strongly contributes to various non-communicable diseases showed that supermarket shopping is associated with improved
(NCDs), such as diabetes, hypertension, and some forms of cancer dietary quality in Tunis, Tunisia. However, average living standards
(NCD Risk Factor Collaboration, 2016). Obesity and NCDs are asso- in Tunisia are higher than in most other African countries. More-
ciated with morbidity and mortality, lost labor productivity, and over, data from a large city, such as Tunis, may not be representa-
high healthcare costs (Bommer et al., 2017; Herman, 2013; IFPRI, tive for other regions. Studies with data from Kenya and Guatemala
2016; Withrow & Alter, 2011; World Economic Forum, 2011). revealed that supermarkets contribute to higher overall energy
Rising rates of obesity are caused by income growth, urbaniza- consumption and a larger share of energy from processed foods
tion, and related changes in people’s lifestyles and diets. The ‘‘nu- (Asfaw, 2008; Kimenju, Rischke, Klasen, & Qaim, 2015; Rischke,
trition transition” is particularly characterized by higher Kimenju, Klasen, & Qaim, 2015). The same studies for Kenya and
consumption of processed foods that are dense in sugar, fat, and Guatemala also suggested that supermarket shopping increases
salt (Popkin, Adair, & Ng, 2012). Changes in the food retail environ- adult BMI and the likelihood of being overweight or obese. A study
ment may also play a role. In many developing countries, modern with data from Indonesia found no significant association between
supermarkets are spreading rapidly (Reardon, Timmer, Barrett, & supermarket shopping and BMI (Umberger, He, Minot, & Toiba,
Berdegué, 2003). As supermarkets sometimes offer different types 2015). These existing studies used cross-sectional survey data,

http://dx.doi.org/10.1016/j.worlddev.2017.07.018
0305-750X/Ó 2017 The Author(s). Published by Elsevier Ltd.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 K.M. Demmler et al. / World Development 102 (2018) 292–303 293

partly employing instrumental variable (IV) approaches to draw the types of foods that customers with rising incomes and appeal
causal inference. However, finding a valid instrument that is corre- for modernity demand. However, it is likely that supermarkets
lated with supermarket shopping but uncorrelated with diets and do not only react to changing consumer preferences but, in turn,
nutrition is very difficult. Hence, causal inferences based on also shape these preferences to some extent. Influence on con-
cross-section observational data remain tentative (Bound, Jaeger, sumer food choices can occur through locational factors, the range
& Baker, 1995). of products offered, the positioning of items in the shelves, packag-
We contribute to this research direction by using panel data and ing sizes, promotional campaigns, and general shopping atmo-
panel regression models for more robust causal inference. The sphere (Battersby & Peyton, 2014; Hawkes, 2008; Timmer, 2009).
main aim is to get a better understanding of the effects that the Compared to small traditional shops, supermarkets can better
spread of supermarkets in developing countries has on consumers’ exploit economies-of-scale. Hence, certain foods can be offered at
diets and nutrition. In particular, we use data collected in urban lower prices (Drewnowski, Aggarwal, Hurvitz, Monsivais, &
Kenya in 2012 and 2015 to analyze the effects of supermarket Moudon, 2012; Rischke et al., 2015). This is especially relevant
shopping on adult BMI and dietary composition. Kenya has one for non-perishable processed food items. In fact, outside of bigger
of the most prospering supermarket sectors in sub-Saharan Africa cities, supermarkets in developing countries often concentrate pri-
(Neven, Odera, Reardon, & Wang, 2009; Rischke et al., 2015). The marily on the sale of processed foods.1 Cheaper access to processed
share of grocery sales through supermarkets is about 10% at foods can improve food security and nutrition for very poor popula-
national level, but already much higher in large urban centers tion segments (Kimenju & Qaim, 2016; Reardon et al., 2003). How-
(Planet Retail., 2016). A rapid growth of supermarkets is also ever, heavy reliance on processed foods does not necessarily
expected in other parts of Africa. Better understanding the nutri- improve dietary quality and can intensify the obesity pandemic.
tion effects of modernizing retail environments can help to design Hence, the spread of supermarkets in developing countries can have
policies aimed at reducing negative health externalities. both positive and negative nutrition and health effects.

2. Food environment and dietary choices 3. Materials and methods

Food choices are determined by various biological, socioeco- (a) Data

nomic, and psychological factors (Nestle et al., 1998). Food avail-
ability, price, type of display, quality, personal income, attitudes, We use data from a survey of households and individuals car-
taste, time constraints, and several other factors play a role when ried out in two rounds in Central Kenya. The first round was carried
people decide on what to eat (Dover & Lambert, 2016; Ventura & out in 2012, the second in 2015. The survey concentrated on small
Worobey, 2013). Economic development is typically associated towns (<70 thousand inhabitants), because this is the typical size
with profound changes in people’s diets. Income growth, urbaniza- of towns that supermarket chains currently enter in Kenya. All lar-
tion, technological change, advances in food preservation, and ger cities in the nation already have one or more supermarkets,
advertising through mass media, all contribute to higher consump- whereas in rural areas supermarkets are not yet observed. In
tion of relatively energy-dense processed foods and beverages. 2012, we purposively selected three towns in Central Kenya with
These dietary shifts are often referred to as the ‘‘nutrition transi- differences in the availability of supermarkets.2 The three towns
tion” (Popkin, 2014; Popkin et al., 2012). In most developed coun- are Ol Kalou and Njabini in Nyandarua County, and Mwea in Kiri-
tries, this nutrition transition already occurred several decades ago. nyaga County. Ol Kalou has had a supermarket since 2002. In Mwea,
In many developing countries, it is now happening at a relatively a supermarket was opened in 2011. Njabini had no supermarket,
fast pace. neither in 2012 nor in 2015. This provides a quasi-experimental set-
The nutrition transition can contribute to increases in body ting for the analysis of supermarket impacts on diets and nutrition.3
weight in two ways. First, consuming energy-dense foods will Except for these differences, the three towns are similar in terms of
likely lead to higher overall energy intakes. Second, nutrient com- infrastructure and other economic development indicators (Kenya
position and processing levels play important roles for the human National Bureau of Statistics, 2010).
body’s energy usage during food digestion and storage. On average, Systematic random sampling was used to select households for
the human body’s energy use for food digestion and storage makes interview within the urban and peri-urban areas of the three
up around 15% of total daily energy expenditures (Barr & Wright, towns. Since recent census data did not exist, we used available
2010). However, this value varies with dietary composition. For population statistics and the help of local administrators. At first,
instance, the body requires more energy for digesting proteins than all neighborhoods (residential estates) in each town were listed.
for carbohydrates and fats (Westerterp, 2004). Also, the digestion Then, household lists were compiled for each neighborhood, from
of fresh and whole foods with higher fiber contents requires more which we randomly selected the required number of households.
energy than the digestion of processed foods (Barr & Wright, 2010). We selected households from all neighborhoods, in order to avoid
Higher energy intakes and lower body energy expenditures may clustering and obtain a representative sample at town level.
have positive nutrition effects in situations where people suffer In each selected household, whenever available one male and
from energy deficiency. However, for people with sufficient energy one female adult (>18 years) were included in the study for inter-
consumption, the nutrition transition contributes to overweight views and anthropometric measurements. In 2012, we included
and obesity (Popkin et al., 2012). 432 randomly selected households and 601 adults. In 2015, we
Changing retail environments may possibly speed up the nutri- 1
In big cities, many supermarkets and hypermarkets also have large fresh fruit and
tion transition. In developing countries, supermarkets and other
vegetable sections, but in smaller cities and towns this is rare up till now, at least in
modern retail outlets are spreading rapidly, partly crowding out low-income countries of Asia and Africa (Rischke et al., 2015).
more traditional markets and small shops (Reardon et al., 2003). 2
The cross-sectional data collected in 2012 were also used by Kimenju et al. (2015)
Supermarkets tend to be larger than traditional outlets, and they and Rischke et al. (2015). This study builds up on this earlier research with panel data.
3
usually offer a bigger range of products under one roof. Another Living in a town with supermarket is not perfectly correlated with supermarket
use. Not all households in Ol Kalou and Mwea use supermarkets to buy food, and a
major difference is that supermarkets have self-service character, few households in Njabini occasionally buy food in supermarkets elsewhere.
providing greater freedom of choice for customers. Supermarkets However, this deliberate choice of towns provides exogenous variation in supermar-
respond to changing consumer preferences and lifestyles, offering ket use that is very useful for the impact evaluation.
294 K.M. Demmler et al. / World Development 102 (2018) 292–303

tried to reach the same households and individuals, but were only and thus helps to better understand the mechanisms for nutritional
able to track 219 households and 286 adult individuals of those outcomes.
that were also included in 2012. Unlike in rural areas, where The models in Eqns. (1) and (2) can be estimated with random
extended families often live in the same place for several genera- effects (RE) panel estimators. However, one potential issue is that
tions, in urban areas households are often much smaller and relo- the individual decision where to buy food is not random and
cate more frequently. Hence, higher attrition rates in urban panels may be influenced by unobserved factors. If such unobserved fac-
are commonplace. Attrition households were replaced with other tors are also correlated with the nutritional outcomes or the diet-
randomly selected ones in the same towns and neighborhoods. In ary dependent variables, the estimated supermarket effects would
total, in 2015 we collected data from 430 households and 598 adult be biased. This type of bias due to unobserved heterogeneity is also
individuals. Thus, the total sample includes 1,199 individual adult the main reason why IV approaches are commonly employed in
observations. impact evaluations with cross-sectional data. When panel data
Table 6 in the Appendix compares key variables for individuals are available, as in our case, estimators with individual fixed effects
that were included in both survey rounds (balanced panel) and (FE) can alternatively be used. FE estimators use differencing tech-
those that had to be excluded and newly included in 2015 due to niques, so that time-invariant heterogeneity is canceled out, even if
attrition. While small differences occur for age and gender, no sig- unobserved (Wooldridge, 2010). Time-variant heterogeneity may
nificant differences are found for consumption expenditures and still bias the results, which is why we control for living standards
other indicators of living standard. Against this background, we and levels of physical activity that can change over time. Much
use the unbalanced panel in the further analysis, even though we more difficult to capture are individual lifestyle factors and atti-
test key results for possible attrition bias. tudes that may also influence the decision where to buy food.
However, such unobserved factors are not expected to change
(b) Statistical methods within three years (the period in-between our two survey rounds),
so that they can be considered as time-invariant in this analysis.
Our main objective is to analyze the effects of supermarket Hence, we argue that FE estimators properly control for unob-
shopping on adult nutritional outcomes. For this purpose, we esti- served heterogeneity in our context without the need for
mate panel data regression models of the following type: instruments.
FE panel estimators require data variability within individuals
Nit ¼ b0 þ b1 Sit þ b2 X it þ eit ð1Þ over time. Hence, while unbalanced panel data can be used, the
FE specifications rely on those individuals that were included in
where Nit is the nutritional outcome variable for individual i at time both survey rounds. We run all models with both FE and RE esti-
t, such as BMI or being overweight or obese. The main explanatory mators and compare results using the Hausman test (Hausman,
variable of interest is Sit , a dummy variable that indicates whether 1978). A significant Hausman test statistic means that there is
or not the individual (or the household in which individual i lives) unobserved heterogeneity, so that the FE specification is preferred.
purchased any food in supermarkets (see below for details of vari- For all model estimations, we use standard errors that are cluster-
able definitions). X it is a vector of control variables, and eit is a ran- corrected at the household level, which is important because in
dom error term. We are particularly interested in the coefficient most households we observed more than one individual. All statis-
estimate for b1 . A positive and significant estimate for b1 would tical analyses are conducted using Stata version 13.
indicate that shopping in supermarkets has a net-increasing effect
on BMI, or on the likelihood of being overweight or obese. (c) Supermarket dummy variable
One important question is what type of control variables to
include in the vector X it . Especially relevant are variables that The main explanatory variable of interest in the regression
may be jointly correlated with N it and Sit , as omitting such vari- models is the supermarket dummy variable (Sit ), which takes a
ables could lead to biased estimates for b1 . We include a range of value of one if any food consumed in the household of individual
factors, such as individual age, gender, marital status, and physical i during the 30 days prior to the survey was purchased in a super-
activity levels, as well as household living standard (economic sta- market, and zero if all the food consumed was obtained from tra-
tus). In developing countries, living standard is often positively ditional sources. Traditional sources include traditional retailers,
correlated with BMI (Popkin et al., 2012). At the same time, richer such as daily markets, small shops, and kiosks, as well as food from
households are more likely to buy food in supermarkets, because own production or obtained through gifts. Table 7 in the Appendix
they can afford a wider range of processed and convenience foods. shows characteristics of the different sources of food (retail out-
Moreover, consumers in developing countries often associate lets), including typical food groups obtained from these sources.
supermarkets with western brands and modern lifestyles (Batra, Information on food consumption was obtained at the house-
Ramaswamy, Alden, Steenkamp, & Ramachander, 2000; Hawkes, hold level through a 30-day recall covering 168 food items. The
2008). Hence, not controlling for living standard would likely lead recall interviews were conducted with the household member that
to an overestimated coefficient b1 . Similarly, physical activity was mainly responsible for food purchases and food preparation. In
levels may also be jointly correlated with supermarket shopping addition to the quantities consumed, information on sources and
and nutritional outcomes. Finally, we include a time trend as part monetary expenditures was collected separately for each food
of vector X it , and town dummy variables to control for possible item.
regional differences. In the total sample with 1,199 observations, 668 individuals had
In addition to Eqn. (1) with nutritional outcomes as dependent consumed food purchased in supermarkets, whereas the other 531
variables, we estimate models with diet-related dependent vari- had not. The proportion of supermarket shoppers varies by town.
ables as follows: As one could expect, most non-supermarket shoppers live in Nja-
Dit ¼ c0 þ c1 Sit þ c2 X it þ eit ð2Þ bini, where no supermarket had been opened until 2015. A certain
proportion of non-supermarket shoppers is also found in the other
where Dit is a dietary indicator of individual i at time t, such as the two towns, Mwea and Njabini. There is also variation in supermar-
share of energy consumed from highly processed foods, or the ket shopping over time, which is important for efficient FE estima-
energy consumed from specific food groups. The coefficient c1 char- tions. As mentioned, in Mwea a supermarket was only established
acterizes the net effects of supermarket shopping on dietary choices in 2011, shortly before the first survey round was conducted in
 K.M. Demmler et al. / World Development 102 (2018) 292–303 295

2012. As people first have to get used to this new retail format, expenditures in 2015 were deflated to 2012 using official con-
some of the households in Mwea that had not yet used the super- sumer price indices (Kenya National Bureau of Statistics., 2016).
market in 2012 had started to use it by 2015. Some variation in Finally, we control for individual physical activity, as this can
supermarket shopping over time was also observed in the other also influence food consumption and nutritional outcomes. In the
two towns. Out of those individuals that were included in both sur- survey, respondents were asked for the number of hours of physi-
vey rounds (n = 286), 44 (15%) had switched their supermarket cal activity during leisure time. These data were used to calculate
shopping status during 2012–15. leisure time physical activity ratios (PAR).4 PAR is a continuous
variable taking values larger than 1. Bigger values indicate higher
levels of physical activity.
(d) Nutritional outcomes and dietary variables

We use the body mass index (BMI) as the main indicator of 4. Results
nutritional outcomes for adults. BMI is the most common indicator
to classify overweight and obesity (Nelms, Sucher, & Lacey, 2011). (a) Descriptive statistics
Anthropometric measurements of individual weight and height
were obtained during both rounds of the survey according to inter- Descriptive statistics for key variables used in this analysis are
national standards (Centers for Disease Control & Prevention, shown in Table 1, for the total sample and also disaggregated for
2007). Using these measurements, we calculated BMI (BMI = body supermarket shoppers and non-shoppers. The upper part of the
weight in kg/body height in meters squared) for each individual. table shows the nutrition and dietary indicators.
Using common international thresholds for BMI, we also classified Even though Kenya is still facing problems of undernutrition
individuals according to their nutritional status (WHO, 2014). and child stunting, rates of adult overweight and obesity are high.
Adults with a BMI  25 kg/m2 and <30 kg/m2 are defined as over- In our sample, 47% of the adults were overweight or obese. This is
weight. With a BMI  30 kg/m2 individuals are defined as obese. higher than the average of 26% found in recent statistics for Kenya
We club the two categories and define individuals with (IFPRI, 2016; Kenya National Bureau of Statistics, 2014; WHO,
BMI  25 kg/m2 as overweight/obese. 2015). However, these national statistics refer to all of the coun-
For the dietary analysis, we used the food consumption data try’s regions, including poor rural areas where undernutrition is
from the 30-day recall. Quantities of each food item consumed still more widespread. Regionally disaggregated official statistics
by the household were converted to amounts of energy using are only available for women. For Central Kenya, where the three
national food composition tables for Kenya and other countries towns included in this study are located, the prevalence of over-
in Africa (FAO, 2010, 2012; Sehmi, 1993). Energy consumption weight/obesity among female adults was estimated at 47% in
from each food item at the household level was divided by 30 to 2014 (Kenya National Bureau of Statistics, 2014). Hence, the nutri-
obtain daily values and then converted to individual levels with tional outcomes measured in our survey seem to be reasonable for
the help of adult equivalent scales. Adult equivalents (AE) were urban areas in Central Kenya.
calculated based on average energy requirements, taking individ- Looking at the disaggregated groups in Table 1, we see that
ual age, sex, and body height into account (FAO, 2004). those shopping in supermarkets have a significantly higher mean
In addition to total energy consumption per person (expressed BMI and are also more likely to be overweight or obese than those
in kcal/AE/day), we also look at energy consumption from specific not shopping in supermarkets. Figure 1 breaks these comparisons
food groups that may be affected by supermarket shopping. As down by survey year. During 2012–15, BMI of both groups
supermarkets in small towns offer very few fresh and unprocessed increased considerably, but the increase was more pronounced
foods, we are particularly interested in effects on energy from for those shopping in supermarkets.5 The data in Table 1 also show
unprocessed staples (grains, pulses, roots, and tubers) and fruits that supermarket shoppers have significantly higher total energy
and vegetables. These groups are generally considered as ‘‘healthy” consumption than non-supermarket shoppers and a larger share of
foods, because they are high in dietary fiber. Fruits and vegetables this energy comes from animal products and highly processed foods.
are also rich in vitamins and minerals. Other food groups, such as However, these comparisons do not control for other factors that
meats and fish, dairy and eggs, and vegetable oils, are more may also influence diets and nutrition. As can be seen in the lower
energy-dense and often further processed. High consumption of part of Table 1, there are also significant differences in living stan-
such energy-dense foods can more easily contribute to overweight dard and other sociodemographic variables. Below, we control for
and obesity (Swinburn, Caterson, Seidell, & James, 2004). Further- such differences through estimation of panel regression models.
more, we look at the share of highly processed foods (see Table 8
in the Appendix) in total daily energy consumption, as this may
(b) Supermarket effects on BMI
also be influenced by supermarket shopping.
Table 2 shows results of panel regression models with BMI as
(e) Control variables dependent variable. Model (1) refers to the unbalanced panel with
all observations included. Two versions are shown, one with FE and
In the individual-level regression models to explain nutritional the other with RE specifications. The Hausman test statistic, which
outcomes and diets we control for typical sociodemographic fac- is shown in the lower part of the table, suggests that the FE spec-
tors such as age, sex, and marital status. In addition, we include ification is preferred. Shopping in supermarkets increases individ-
a year dummy variable for observations in 2015 and town vari- ual BMI by 0.64 kg/m2. The finding of a net-increasing effect of
ables for Ol Kalou and Njabini (Mwea is the reference category). supermarkets on BMI is consistent with Asfaw (2008) and
It should be noted that all time-invariant variables drop out in Kimenju et al. (2015), who had used cross-sectional data. However,
the FE specifications. In all models, we also control for household
4
living standard, measured in terms of per capita consumption PAR is defined as a multiple of the basal metabolic rate. In the nutritional
sciences, PAR is often used to calculate physical activity levels (PAL), which are one
expenditures in Kenyan Shillings (KES). These expenditures com-
ingredient in determining individual energy requirements (FAO, 2004).
prise the value of all food and non-food goods and services con- 5
While the growth rates in BMI and in the prevalence of overweight/obesity during
sumed over a period of 30 days, including home-produced foods. 2012–15 are higher for supermarket shoppers, the growth rate differences between
To make monetary values comparable between survey years, the two groups are not statistically significant.
296 K.M. Demmler et al. / World Development 102 (2018) 292–303

Table 1
Sample descriptive statistics

Variable Total Shopping in supermarkets Not shopping in supermarkets

2 ***
Body mass index (kg/m ) 25.33 25.80 24.73
(5.07) (5.08) (5.00)
Overweight/obese (1,0) 0.47 0.52*** 0.40
(0.50) (0.50) (0.49)
Energy consumption (kcal/AE/day) 3164.61 3300.71*** (1388.74) 2993.41
(1439.11) (1483.75)
Energy from unprocessed staples (kcal/AE/day) 408.66 387.46** 435.34
(386.15) (421.46) (335.01)
Energy from fruits and vegetables (kcal/AE/day) 375.32 392.05*** (245.02) 354.26
(250.35) (255.58)
Energy from meats and fish (kcal/AE/day) 121.84 148.28*** (123.06) 88.59
(112.00) (85.49)
Energy from dairy and egg (kcal/AE/day) 39.75 47.60*** 29.89
(45.90) (51.67) (35.02)
Energy from oils (kcal/AE/day) 133.26 187.68*** (208.80) 64.79
(190.58) (137.12)
Share of energy from highly processed foods (%) 7.60 8.57*** 6.37
(5.59) (5.25) (5.76)
Expenditure per capita (1000 KES) 11.90 14.02*** 9.24
(9.19) (10.67) (5.88)
Age (years) 36.54 34.60*** 38.99
(12.20) (9.92) (14.21)
Female (1,0) 0.65 0.67 0.63
(0.48) (0.47) (0.48)
Married (1,0) 0.74 0.76** 0.70
(0.44) (0.43) (0.46)
Physical activity ratio (PAR) 2.23 2.21** 2.27
(0.49) (0.47) (0.51)
Ol Kalou (1,0) 0.32 0.50*** 0.09
(0.47) (0.50) (0.29)
Mwea (1,0) 0.29 0.41*** 0.14
(0.46) (0.49) (0.35)
Njabini (1,0) 0.39 0.08*** 0.77
(0.49) (0.28) (0.42)
Share of supermarket purchase (%) 8.39 15.06*** 0.00
(11.24) (11.25) (0.00)
Number of observations 1199 668 531

Notes: Mean values are shown with standard deviations in parentheses. **Difference between those shopping and not shopping in supermarkets is significant at 5% level;
***
Difference between those shopping and not shopping in supermarkets is significant at 1% level.

our estimate is smaller in magnitude. For instance, Kimenju et al. variables from the RE specification of model (1), the supermarket
(2015), who used the same data from Central Kenya collected in effect on BMI increases to 0.72.
2012, estimated that supermarket shopping increases BMI by We carry out a few additional tests to check the robustness of
1.69 kg/m2. As argued above, the FE panel estimator used here is the results. A first test relates to the possible effects of sample attri-
more reliable because it does not depend on assumptions about tion. Model (2) in Table 2 shows FE and RE specifications of the BMI
the validity of an instrument. However, in spite of the smaller model with only the observations from the balanced panel
effect found here, we confirm the hypothesis that supermarkets included. Except for the constant term, the FE results are identical
contribute to BMI increases, even after controlling for unobserved to those in model (1), which is not surprising. Although all obser-
heterogeneity and other confounding factors. vations were included in model (1), FE estimation of the treatment
The other results of model (1) in Table 2 show that being mar- effect only considers individuals that were included in both survey
ried also contributes to higher BMI. Furthermore, the RE specifica- rounds, as the FE estimator exploits the variation within individu-
tion, which includes the time-invariant characteristics that drop als over time. But also for the RE specifications, results of models
out from the FE specification, suggests that females have a much (1) and (2) are quite similar, which we take as evidence that sam-
higher BMI than males. This is consistent with existing statistics ple attrition does not lead to systematic bias.
from Kenya and elsewhere (Kenya National Bureau of Statistics, A second test relates to the relatively small number of super-
2014; Ng et al., 2014). BMI is also positively associated with age market switchers. As mentioned in Section 3, there are only 44
and living standard, as one would expect. Looking at the town vari- individuals in the sample who were included in both survey
ables, we see that people living in Ol Kalou have a higher BMI than rounds and switched their supermarket shopping status during
those living in Mwea, which is the reference town in this model. As 2012–15 (88 observations). The FE estimates rely on these
mentioned, Ol Kalou is the town where a supermarket had already switchers, so it is important to know how representative they
opened in 2002. On the other hand, people in Njabini, where no are for the rest of the sample. Table 9 in the Appendix compares
supermarket had been opened until 2015, have a significantly key socioeconomic characteristics of these switchers with the
lower BMI. This correlation between the town variables and nutri- total sample. The switchers are more likely to be female. In terms
tional status is likely the result of our sampling strategy where we of the other variables, including household living standards, no
deliberately chose towns with differences in supermarket access. It significant differences are observed. Of course, a larger number
implies that the town variables may possibly capture some of the of switching observations could lead to more efficient FE esti-
effects of supermarket shopping. Indeed, when excluding the town mates. But the similarity of the switchers with the rest of the
 K.M. Demmler et al. / World Development 102 (2018) 292–303 297

(A) Body mass index (BMI) specifications of model (1) show positive coefficients for supermar-
ket shopping, but these are not statistically significant. This is sur-
27
prising because Figure 1 shows that supermarket shoppers are
26.34***
significantly more likely to be overweight/obese than individuals
26
who obtained all of their food from traditional sources. Interesting
25.31**
25.06 to see in Table 3, however, is that people in Njabini are significantly
25
24.39
less likely to be overweight/obese than people in Mwea, even after
controlling for other factors. Njabini is the town where no supermar-
24 Shopping in supermarkets ket had opened until 2015. In model (2) of Table 3, we exclude the
BMI

Not shopping in
town variables and suddenly see a significant positive coefficient
23
supermarkets for supermarket shopping. According to this model, shopping in
supermarkets increases the probability of being overweight/obese
22
by 7 percentage points.7
We admit that the evidence of an overweight/obesity increasing
21
net effect of supermarket shopping in our data is not very strong,
20
also because the RE specifications do not control for unobserved
2012 2015 heterogeneity. That the supermarket effect is not showing up more
clearly is due to the fact that many adults have a BMI around 25 kg/
(B) Prevalence of overweight/obesity m2. Of course, supermarkets are not the only factors contributing
to BMI increases, so that crossing the overweight/obesity threshold
0.7
occurs in both groups, supermarket shoppers and non-shoppers
(Figure 1). However, the finding that supermarket shopping signif-
0.6 0.58***
icantly increases BMI as such already implies that this will also
contribute to more overweight/obesity. We presume that this
Proportion of individuals

0.5
0.46*** 0.45 would be more visible with a larger number of switching observa-
0.4 tions in the balanced panel.
0.35 Shopping in supermarkets

0.3 Not shopping in (d) Supermarket effects on dietary choices

supermarkets

0.2 To better understand how supermarkets contribute to rising

BMI, we analyze effects on consumers’ dietary choices. Several
0.1 studies had used cross-sectional data to show that supermarket
shopping contributes to higher total energy consumption (Asfaw,
0 2008; Kimenju et al., 2015; Rischke et al., 2015; Toiba, Umberger,
2012 2015
& Minot, 2015). Rischke et al. (2015) showed that the average price
Figure 1. Differences in nutritional outcomes between individuals shopping and of calories purchased in supermarkets is lower than the price per
not shopping in supermarkets. **Difference between those shopping and not calorie purchased in traditional outlets. This could explain some
shopping in supermarkets is significant at 5% level; ***Difference between those of the calorie consumption effects. Our descriptive statistics con-
shopping and not shopping in supermarkets is significant at 1% level.
firm that supermarket shoppers consume significantly more calo-
ries than people who obtain all of their food from traditional
sample suggests at least that the FE estimates do not suffer from sources (Table 1). However, panel model estimates that we tried
significant selection bias. revealed that these differences in total energy consumption cannot
A third test relates to the possible role of traditional retail out- be interpreted as a net effect of supermarket shopping. After con-
lets, which are not uniform. As shown in Table 7 in the Appendix, trolling for other factors, supermarket shopping does not increase
traditional retailers include daily markets, kiosks, and small shops. total energy consumption significantly.
In terms of some characteristics, small shops are similar to super- However, beyond total energy consumption we find significant
markets: while supermarkets are larger and offer a wider variety of effects of supermarkets on dietary composition. The FE specifica-
processed foods, some small shops also have a self-service option. tion in Table 4 shows that shopping in supermarkets increases
To analyze the possible role of small shops, we include an addi- the share of energy from highly processed foods in total energy
tional dummy variable for shopping in these small shops in the consumption by about 3 percentage points. This increase is plausi-
BMI models. Results are shown in Table 10 in the Appendix. Shop- ble given that supermarkets in the small towns considered here
ping in small shops does not seem to affect individual BMI, neither primarily sell processed and highly processed foods. Higher con-
in the FE nor in the RE specification. At the same time, the super- sumption of highly processed foods with more sugar, fat, and lower
market effects remain significant and similar in magnitude to those fiber content can contribute to rising BMI even without significant
in Table 2. effects on total energy consumption.
A tendency of supermarkets to contribute to dietary shifts
toward more processed foods was also found by Asfaw (2008),
(c) Supermarket effects on the prevalence of overweight/obesity Kimenju et al. (2015), and Rischke et al. (2015). Coefficient esti-
mates are not directly comparable across studies, because of differ-
Table 3 shows results of model estimates where being over- ences in the exact specification of the dependent variables and
weight/obese is used as a dummy dependent variable. We use lin- functional forms. Yet, in general, the earlier studies with cross-
ear probability models for these estimates.6 The FE and RE
7
This is in line with findings by Asfaw (2008) and Kimenju et al. (2015), even
6
Alternatively, one could have estimated probit models. The reason why we prefer though the estimated effects in these earlier cross-sectional studies were larger. For
linear probability models is that these also allow fixed effects specifications, which is instance, Kimenju et al. (2015) estimated that supermarket shopping increases the
not possible with probit models in most software packages. probability of being overweight/obese by 13 percentage points.
298 K.M. Demmler et al. / World Development 102 (2018) 292–303

Table 2
Effects of supermarket shopping on body mass index

Body mass index (kg/m2)

(1) (2)
FE RE FE RE
Shopping in supermarkets (1,0) 0.64* (0.38) 0.61** (0.29) 0.64* (0.38) 0.70** (0.36)
Married (1,0) 1.07* (0.56) 1.06*** (0.30) 1.07* (0.56) 0.93** (0.44)
Physical activity ratio 0.22 (0.18) 0.25 (0.16) 0.22 (0.18) 0.27 (0.17)
Female (1,0) 3.29*** (0.28) 3.29*** (0.49)
Age (years) 0.02 (0.04) 0.10*** (0.02) 0.02 (0.04) 0.08*** (0.02)
Expenditure per capita (1000 KES) 0.01 (0.02) 0.06*** (0.02) 0.01 (0.02) 0.03 (0.02)
Ol Kalou (1,0) 0.84** (0.39) 0.46 (0.75)
Njabini (1,0) 0.82* (0.43) 1.01 (0.76)
Year 2015 0.38** (0.17) 0.00 (0.13) 0.38** (0.17) 0.03 (0.14)
Constant 25.26*** (1.50) 18.63*** (0.74) 25.89*** (1.62) 20.30*** (1.15)
Waldv2 236.38*** 75.25***
F-value 2.50** 2.48**
Hausman testv2 58.43*** 48.39***
Number of observations 1199 1199 572 572

Notes: Coefficient estimates are shown with standard errors cluster-corrected at household level in parentheses. Model (1) uses the unbalanced panel with all observations.
Model (2) only uses observations from the balanced panel. FE, fixed effects; RE, random effects. *Significant at 10% level; **Significant at 5% level; ***Significant at 1% level.

Table 3
Effects of supermarket shopping on the probability of being overweight/obese

Being overweight/obese (1,0)

(1) (2)
FE RE RE
Shopping in supermarkets (1,0) 0.01 (0.04) 0.03 (0.03) 0.07** (0.03)
Married (1,0) 0.07 (0.05) 0.09*** (0.03) 0.09*** (0.03)
Physical activity ratio 0.04 (0.03) 0.04** (0.02) 0.04** (0.02)
Female (1,0) 0.25*** (0.03) 0.26*** (0.03)
Age (years) 0.01 (0.01) 0.01*** (0.00) 0.01*** (0.00)
Expenditure per capita (1000 KES) 0.00 (0.00) 0.01*** (0.00) 0.01*** (0.00)
Ol Kalou (1,0) 0.06 (0.04)
Njabini (1,0) 0.10** (0.04)
Year 2015 0.09*** (0.03) 0.04** (0.02) 0.05** (0.02)
Constant 0.80*** (0.30) 0.07 (0.08) 0.15* (0.08)
Waldv2 215.99*** 201.00***
F-value 2.17**
Hausman testv2 26.32***
Number of observations 1199 1199 1199

Notes: Coefficient estimates of linear probability models are shown with standard errors cluster-corrected at household level in parentheses. Being overweight/obese
includes individuals with BMI > 25 kg/m2. FE, fixed effects; RE, random effects. *Significant at 10% level; **Significant at 5% level; ***Significant at 1% level.

sectional data suggested larger effects on dietary composition,

Table 4 underlining again the importance of panel data for identifying reli-
Effects of supermarket shopping on the share of energy consumed from highly able net impacts of supermarket shopping.
processed foods
Table 5 analyzes further details of supermarket effects on peo-
Share of energy from highly ple’s diets beyond highly processed foods. The models shown have
processed foods (%) absolute energy consumption from different food groups as depen-
FE RE dent variables. In all models, the supermarket dummy variable has
Shopping in supermarkets (1,0) 3.07*** (1.13) 0.45 (0.87) significant coefficients, either in the FE or RE specifications. The FE
Married (1,0) 3.08 (2.62) 1.61** (0.78) specifications suggest that supermarket shopping reduces energy
Physical activity ratio 0.65 (0.57) 0.20 (0.48) consumption from unprocessed staples by 112 kcal/AE/day, and
Female (1,0) 1.46** (0.59)
from fresh fruits and vegetables by 124 kcal/AE/day. These are sub-
Age (years) 0.11 (0.13) 0.23*** (0.02)
Expenditure per capita (1000 KES) 0.06 (0.06) 0.18*** (0.04) stantial effects, accounting for more than one-third of total average
Ol Kalou (1,0) 0.68 (0.80) energy consumption from these two food groups.
Njabini (1,0) 1.90* (1.07) For the other food groups in Table 5, the supermarket dummy
Year 2015 2.33*** (0.60) 2.76*** (0.45) variable is only significant in the RE specifications. Yet the Haus-
Constant 4.71 (4.95) 19.77*** (2.09)
man test statistics suggest that unobserved heterogeneity is not
Waldv2 177.89***
F-value 5.96*** an issue in these models, so that the RE estimator produces unbi-
Hausman testv2 23.10*** ased estimates.
Number of observations 1199 1199 Supermarket shopping increases the consumption of meats and
Notes: Coefficient estimates are shown with standard errors cluster-corrected at
fish by 24 kcal/AE/day, of dairy and eggs by 9 kcal/AE/day, and of
household level in parentheses. FE, fixed effects; RE, random effects. *Significant at vegetable oils by 60 kcal/AE/day. Together with highly processed
10% level; **Significant at 5% level; ***Significant at 1% level. foods, these are also the food groups that supermarket shoppers
 K.M. Demmler et al. / World Development 102 (2018) 292–303 299

Significant at 5% level;
16

118.73*** (16.97)
112.32*** (17.71)
27.66** (13.26)
59.81*** (15.31)

21.06*** (7.39)

35.76*** (9.67)
25.26 (40.36)
7.75*** (1.38)
1.24*** (0.46)
2.82 (11.25)
14

248.89***

1199
RE
Vegetable oils 12
Supermarkets

Consumption (kg)
10 Traditional sources

**
37.27 (63.46)

34.11** (14.10)
78.65 (117.63)
5.80 (19.16)

9.70*** (2.42)
1.16 (2.00)
9.03 (27.39)

Notes: Coefficient estimates are shown with standard errors cluster-corrected at household level in parentheses. AE, adult equivalent; FE, fixed effects; RE, random effects. Significant at 10% level;
54.99***
6

1199
8.43
FE

24.18** (11.35)
0.26** (0.13)
8.94*** (3.45)

6.26*** (2.37)
1.69*** (0.42)
3.63 (2.33)
0.86 (3.21)
5.34 (4.10)

8.71* (4.60)
6.20 (3.90)

0
51.21***

1199
SM NSM SM NSM SM NSM SM NSM
RE

*
Dairy and egg

Highly processed Meats and fish Dairy and eggs Vegetable oils

Figure 2. Quantity of food consumed from different food groups and food sources.
20.66 (17.11)

Notes: Quantities refer to consumption at the household level over a 30-day period.
18.44 (23.67)
1.55*** (0.55)

6.26** (2.93)
7.88 (6.16)

1.99 (4.17)

0.17 (0.44)

Total quantity consumed per household is split up by quantity purchased in

3.25***

1199
5.75

supermarkets and quantity obtained from traditional sources. SM, refers to

Energy consumption from different food groups (kcal/AE/day)

FE

individuals who purchased some of their food in supermarkets; NSM, refers to

individuals who did not use supermarkets at all. Pooled data for 2012 and 2015.
24.17*** (7.30)

47.73* (28.89)
6.23*** (1.48)
3.80 (6.43)
5.02 (8.01)

0.35 (0.26)

14.06 (9.23)
3.87 (10.21)
1.13 (4.94)

9.10 (5.77)

94.13***

actually purchase most in supermarkets (Figure 2). Table 5 and Fig-

1199
RE

ure 2 also reveal a few other interesting phenomena.

Meats and fish

Households that use supermarkets purchase only some of their

food in supermarkets. Of course, certain foods that are hardly sold
10.54 (10.84)

in supermarkets but that people still want to consume have to be

41.23 (32.21)

34.82 (57.97)
6.12*** (1.25)
5.70 (11.28)

5.13 (7.63)
0.04 (1.14)

obtained from traditional sources. Cases in point are unprocessed

5.81***

1199
6.41
FE

staples and fresh fruits and vegetables. Results in Table 5 show that
supermarket shoppers reduce the consumption of these groups,
but they do not abandon them completely. But even for the types
86.66*** (21.44)
68.36*** (24.85)

151.57*** (51.71)

of foods that are sold in supermarkets, traditional sources continue

28.78* (16.81)

72.35*** (15.38)
16.53 (21.34)

31.96** (13.06)

11.26*** (1.76)
24.12** (9.74)
1.40** (0.62)

to play an important role for all consumers. Interestingly, the

119.49***

1199
RE

quantities of highly processed foods, dairy, and vegetable oils con-

Fruits and vegetables

sumed from traditional sources are more or less the same for those
shopping and not shopping in supermarkets. Only that supermar-
ket shoppers consume extra quantities of these foods that they
124.30** (56.82)

331.75* (169.25)
78.92*** (23.63)
97.29 (93.22)

purchase in supermarkets (Figure 2). Hence, the quantities of these

18.92*** (3.07)
13.04 (24.79)

2.99 (4.60)

foods obtained from supermarkets seem to be of additional nature.

21.42***
9.42***

1199
FE

This may possibly be explained by supermarkets selling popular

brands that are not available in traditional outlets. Larger packag-
Effects of supermarket shopping on energy consumption from different food groups

ing sizes, product placement, pricing, advertising, and the self-

service character of supermarkets may also incentivize customers
170.79*** (24.16)
130.68*** (35.16)

217.03*** (66.29)
47.46* (27.56)

49.31*** (15.59)
22.43 (30.58)

80.82** (34.40)

to buy additional quantities.

2.83*** (1.04)
7.92*** (2.05)
8.07 (17.65)

109.05***

The establishment of supermarkets in small towns of Kenya is a

1199
RE

relatively recent development, and the range of products offered in

Unprocessed staples

these supermarkets is still limited, at least when compared to

much larger stores in big cities. Our data do not allow us to analyze
how dietary behavior of small-town consumers may change when
199.37*** (53.87)
111.61* (59.27)
56.69 (154.93)

272.37 (379.24)
15.13*** (5.00)

the number of supermarkets, as well as store sizes, continue to

21.69 (41.86)

3.04 (9.48)

grow. However, even at this early stage, the results clearly support
5.40***

1199
4.23
FE

the hypothesis that supermarkets contribute to the nutrition tran-

sition, rather than only reacting to shifting consumer preferences.
Expenditure per cap. (1000 KES)
Shopping in supermarkets (1,0)

5. Conclusion
Number of observations

Many developing countries currently experience profound

Physical activity ratio

Significant at 1% level.

transformations in the food retail sector, with modern supermar-

kets massively gaining in importance. While developments are
Hausman testv2
Ol Kalou (1,0)
Married (1,0)

already more advanced in some parts of Asia and Latin America,

Njabini (1,0)
Female (1,0)
Age (years)

Wald-chi2
Year 2015

the share of supermarkets in food retailing is still relatively low

Constant

F-value

in most sub-Saharan African countries, even though it is increasing

Table 5

rapidly. Possible dietary and nutrition implications are not yet suf-
ficiently understood. We have analyzed effects on food consumers
***
300 K.M. Demmler et al. / World Development 102 (2018) 292–303

in Kenya, which is among the countries with the fastest growth of supermarkets can also improve the nutrition of consumers. A few
supermarkets in Africa. Using panel data from small towns in Cen- studies showed that better access to supermarkets is associated
tral Kenya, we have shown that supermarkets significantly affect with healthier diets in some regions in the US (Drewnowski
nutritional outcomes. After controlling for other relevant factors, et al., 2012; Laraia, Siega-Riz, Kaufman, & Jones, 2004; Morland,
our results suggest that shopping food in supermarkets increases Diez Roux, & Wing, 2006). In these situations, supermarkets offer
adult BMI by 0.64 kg/m2. That supermarkets tend to increase con- fresh foods that are otherwise more difficult to access, especially
sumer BMI in developing countries was also shown in a few previ- for lower income consumers living in so-called ‘‘food desert”
ous studies (Asfaw, 2008; Kimenju et al., 2015). These previous neighborhoods (Michimi & Wimberly, 2010). This is different from
studies had even suggested larger effects, but they built on cross- typical situations in Africa, but these examples underline that
section observational data where controlling for possible bias due modern retail is not inevitably associated with negative nutrition
to unobserved heterogeneity is more difficult. We argue that our and health implications.
estimates with panel data models are more realistic and reliable. The expansion of supermarkets in Africa and other parts of the
However, regardless of the exact magnitude of effects, results con- developing world will likely continue. Hence, from a food policy
firm that the growth of supermarkets contributes to the nutrition perspective it is important to understand the diet and nutrition
transition in Africa. implications and intervene where necessary to avoid undesirable
To better understand the underlying mechanisms, we have also outcomes. Intervening does not imply banning supermarkets. But
analyzed effects of supermarkets on consumer dietary choices. certain types of regulations and economic incentives may be
Unlike a few previous studies (Asfaw, 2008; Rischke et al., 2015; appropriate in some situations. For instance, supermarkets in small
Toiba et al., 2015), we did not find that supermarkets contribute African towns so far hardly sell fresh fruits and vegetables, because
to net increases in total calorie consumption. However, our panel this does not yet seem to be profitable. Regulations that incentivize
data models revealed significant shifts in dietary composition. supermarket stores to also offer certain fresh products at reason-
Supermarket shopping contributes to a sizeable decrease in energy able prices could be a possible policy intervention. Alternatively,
consumption from unprocessed staples and from fresh fruits and traditional fruit and vegetable vendors could be encouraged to
vegetables. These food groups are hardly sold in the small-town set up stalls near the supermarket entrances, possibly through con-
supermarkets in Central Kenya that primarily concentrate on pro- tractual arrangements. Other measures to promote dietary diver-
cessed foods. Accordingly, we found significant increases of super- sity and nutrition-sensitive food environments are also worth
market shopping on energy consumption from dairy, vegetable oil, considering. Apart from regulations, this may also include con-
processed meat products (sausages etc.), and highly processed sumer awareness building for the importance of fruits and vegeta-
foods (bread, pasta, snacks, soft drinks etc.). These shifts toward bles in healthy diets.
processed and highly processed foods lead to less healthy diets, Finally, we would like to point out a few limitations of our
with higher sugar, fat, and salt contents, and probably lower study. First, while the use of panel data has clear advantages over
amounts of micronutrients and dietary fibers. Some of the effects cross-sectional data, our panel suffered from significant attrition.
are still relatively small in magnitude, but they may increase with While we tested for attrition bias to the extent possible, a bal-
supermarkets further gaining in importance. The observed changes anced panel with a larger number of observations would be ben-
in dietary composition can also explain the increasing effect on eficial to analyze further details. Especially a sample with a larger
BMI, even without a rise in total calorie consumption. The reason number of individuals switching their supermarket shopping
is that the human body requires less energy for the digestion of behavior over time would be useful for more robust causal infer-
processed and highly processed foods. ence with fixed effects estimators. Second, the geographic range
These results are alarming from a nutrition and health perspec- of our data is limited and the time period considered relatively
tive. Even though we failed to establish a clear effect of supermar- short. More comprehensive and longer term data may help to
ket shopping on the likelihood of being overweight or obese, rising better understand impact heterogeneity and dynamics. Third,
BMI will inevitably aggravate nutrition status in situations where the 30-day food consumption recall at the household level that
many people are already near or above the BMI threshold of we used has certain drawbacks in terms of data accuracy
25 kg/m2, as is the case for adults in Central Kenya. Overweight (Schoeller, 1995). We chose this relatively long recall period
and obesity are responsible for various non-communicable dis- because some of the more durable food items are only purchased
eases that cause high economic costs, human suffering, and lost once a month. However, shorter and repeated recalls at individual
quality of life. level are preferable when the focus is on analyzing actual food
It would be wrong to attribute the obesity pandemic in develop- and nutrient intakes (Shim, Oh, & Kim, 2014). Hence, there is
ing countries to the expansion of supermarkets alone. There are clearly scope for follow-up research to better understand the
many factors that contribute to the nutrition transition. However, nutrition and health effects of the modernizing retail sector in
our results suggest that supermarkets are not only a symptom of various developing-country situations.
this transition, but they influence dietary habits to a significant
extent. Nevertheless, a modernizing retail sector should not be Acknowledgments
condemned, because—if properly managed—it can also have
important positive nutrition effects. For instance, in a recent study This research was financially support by the German Research
in Kenya, Chege, Andersson, and Qaim (2015) showed that small- Foundation (DFG) as part of the GlobalFood Program (RTG 1666).
holder farmers benefit from marketing contracts with supermar- We thank Ramona Rischke and Simon Kimenju for sharing the
kets in terms of higher incomes that also contribute to better first-round survey data collected in Kenya in 2012. We also thank
quality diets in these farm households. Depending on initial nutri- Stephan Klasen, Sebastian Vollmer, and two anonymous reviewers
tion status and access to food diversity, the establishment of new of this journal for valuable comments.
 K.M. Demmler et al. / World Development 102 (2018) 292–303 301

Appendix A.

Table 6
Comparison of balanced panel with excluded and newly included observations in 2015

(1) (2) (3) (4)

Total sample Balanced panel Excluded and newly included in 2015 Difference between (2) and (3)
Female (1,0) 0.65 (0.48) 0.68 (0.47) 0.63 (0.48) 0.06** (0.03)
Age, y 36.54 (12.20) 39.44 (12.77) 33.89 (11.02) 5.55*** (0.69)
Married (1,0) 0.74 (0.44) 0.76 (0.43) 0.72 (0.45) 0.04* (0.03)
Physical activity ratio 2.23 (0.49) 2.25 (0.50) 2.22 (0.48) 0.02 (0.03)
Energy availability (kcal/AE/day) 3164.61 (1439.11) 3205.28 (1513.14) 3127.51 (1368.26) 77.77 (83.60)
Expenditure per capita (1000 KES) 11.90 (9.19) 12.04 (8.28) 11.78 (9.94) 0.26 (0.53)
Number of observations 1199 572 627 1199

Notes: Mean values are shown with standard deviations in parentheses. *Significant at 10% level; **Significant at 5% level; ***
Significant at 1% level.

Table 7
Different sources of food and their characteristics

Source of food Characteristics Main food groups obtained from this source Average share of Number of
total energy observations
consumption (%) using source
Supermarket (modern retail) Self-service; Bread, pasta, cereals, instant noodles, 12.7 668
Large variety of foods and brands; snacks, fats, oils, dairy products, sugar
Highly processed foods;
Refrigerated and frozen food;
Limited offer of fresh foods;
Non-food products;
No credit possibility
Small shop (traditional retail) Semi self-service; Rice, flour, sugar, fats 5.4 485
Limited variety of foods and brands;
Some refrigerated foods;
Sometimes credit possibility
Market/kiosk (traditional retail) Over the counter service; Maize, other staple foods, fruits, 65.7 1199
Very limited variety of brands; vegetables, meat, milk
Fresh fruits and vegetables;
Unprocessed staples;
Credit possibility
Own production/gift Own plot or garden; Maize, potatoes, poultry, eggs, milk 16.3 1014
In a few cases own farms;
Gifts from friends

Table 8
Food groups by level of processing

Food groups Examples

Unprocessed
Eggs & milk Eggs, fresh whole milk, natural yoghurt
Fruits & vegetables Mango, orange, green leafy vegetables, tomatoes, onions
Meats Beef, pork meat, fresh chicken, fresh fish
Pulses Lentils, black beans, cowpea etc.
Roots, tuber, plantain Arrow roots, cassava, yams, potato, cooking bananas
Traditional staples Amaranth, sorghum, green maize
Medium processed
Fats & oils Butter, margarine, vegetable oils
Meats Frozen fish, frozen chicken, dried fish
Staples Rice, maize flour, wheat flour, oats
Sugars Sugar, jiggery
Highly processed
Bread & pasta Bread, cornflakes, pasta
Dairy Flavored yoghurt/milk, tinned baby milk
Fats & oils Peanut butter
Meats Sausages, bacon, ham
Miscellaneous Mandazi, samosa, ketchup
Sugars Glucose powder
Sweet drinks and snacks Chips, soft drinks, cake, popcorn

Notes: The food items mentioned are only examples. In total, 168 food items were included in the survey. All of them were
classified by level of processing following the same principle.
302 K.M. Demmler et al. / World Development 102 (2018) 292–303

Table 9
Comparison of total sample with supermarket switchers

Variable Total sample Supermarket switchers Difference

Female (1,0) 0.65 (0.48) 0.77 (0.42) 0.13*** (0.05)
Age, y 36.54 (12.20) 36.99 (11.02) 0.48 (1.23)
Married (1,0) 0.74 (0.44) 0.75 (0.44) 0.01 (0.05)
Physical activity ratio 2.23 (0.49) 2.24 (0.45) 0.01 (0.05)
Expenditure per capita (1000 KES) 11.90 (9.19) 12.63 (6.02) 0.78 (0.70)
Number of observations 1199 88

Notes: Mean values are shown with standard deviations in parentheses (standard errors in the last column). Supermarket switchers are those who changed their supermarket
shopping status during 2012–15. ***Difference significant at 1% level.

Food and Agriculture Organization of the United Nations (FAO) (2012). West African
Food Composition Table. Rome, Italy: Food and Agriculture Organization of the
Table 10 United Nations.
Effects of supermarket shopping on body mass index with additional controls Food and Agriculture Organization of the United Nations (FAO), United Nations
University (UNU), & World Health Organization (WHO) (2004). Human energy
Body mass index (kg/m2) requirements: Report of a Joint FAO/WHO/UNU Expert Consultation: Rome, 17–24
FE RE October 2001. Rome, Italy: Food and Agricultural Organization of the United
Nations.
Shopping in supermarkets (1,0) 0.65* (0.38) 0.61** (0.29) Hausman, J. A. (1978). Specification tests in econometrics. Econometrica, 46(6),
Shopping in small shops (1,0) 0.14 (0.20) 0.03 (0.19) 1251–1271.
Married (1,0) 1.07* (0.56) 1.06*** (0.30) Hawkes, C. (2008). Dietary implications of supermarket development: A global
Physical activity ratio 0.22 (0.18) 0.25 (0.16) perspective. Development Policy Review, 26(6), 657–692. http://dx.doi.org/
Female (1,0) 3.29*** (0.28) 10.1111/j.1467-7679.2008.00428.x.
Age (years) 0.02 (0.04) 0.10*** (0.02) Herman, W. H. (2013). The economic costs of diabetes: Is it time for a new
Expenditure per capita (1000 KES) 0.01 (0.02) 0.06*** (0.02) treatment paradigm? Diabetes Care, 36(4), 775–776. http://dx.doi.org/10.2337/
dc13-0270.
Ol Kalou (1,0) 0.85** (0.40)
International Food Policy Research Institute (IFPRI) (2016). Global Nutrition Report
Njabini (1,0) 0.83* (0.44)
2016 From Promise to Impact Ending Malnutrition by 2030. Washington, DC:
Year 2015 0.38** (0.17) 0.01 (0.14)
International Food Policy Research Institute.
Constant 25.34*** (1.53) 18.63*** (0.74) Kenya National Bureau of Statistics (2010). Population Distribution by Administrative
Waldv2 247.67*** Units. Nairobi, Kenya: Kenya National Bureau of Statistics.
F-value 2.17** Kenya National Bureau of Statistics (2014). Demographic and Health Survey 2014.
Hausman testv2 59.85*** Nairobi, Kenya: Kenya National Bureau of Statistics.
Number of observations 1199 1199 Kenya National Bureau of Statistics. (2016). Kenya National Bureau of Statistics.
Consumer Price Indices (CPI). Retrieved November 7, 2016, from http://www.
Notes: Coefficient estimates are shown with standard errors cluster-corrected at knbs.or.ke/index.php.
household level in parentheses. FE, fixed effects; RE, random effects. *Significant at Kimenju, S. C., & Qaim, M. (2016). The nutrition transition and indicators of child
10% level; **Significant at 5% level; ***Significant at 1% level. malnutrition. Food Security, 8(3), 571–583. http://dx.doi.org/10.1007/s12571-
016-0566-x.
Kimenju, S. C., Rischke, R., Klasen, S., & Qaim, M. (2015). Do supermarkets contribute
to the obesity pandemic in developing countries? Public Health Nutrition, 18(17),
References 3224–3233. http://dx.doi.org/10.1017/S1368980015000919.
Laraia, B. A., Siega-Riz, A. M., Kaufman, J. S., & Jones, S. J. (2004). Proximity of
Asfaw, A. (2008). Does supermarket purchase affect the dietary practices of supermarkets is positively associated with diet quality index for pregnancy.
households? Some empirical evidence from Guatemala. Development Policy Preventive Medicine, 39(5), 869–875. http://dx.doi.org/10.1016/j.
Review, 26(2), 227–243. http://dx.doi.org/10.1111/j.1467-7679.2008.00407.x. ypmed.2004.03.018.
Barr, S. B., & Wright, J. C. (2010). Postprandial energy expenditure in whole-food and Michimi, A., & Wimberly, M. C. (2010). Associations of supermarket accessibility
processed-food meals: Implications for daily energy expenditure. Food & with obesity and fruit and vegetable consumption in the conterminous United
Nutrition Research, 54(1), 5144. http://dx.doi.org/10.3402/fnr.v54i0.5144. States. International Journal of Health Geographics, 9, 49–63. http://dx.doi.org/
Batra, R., Ramaswamy, V., Alden, D. L., Steenkamp, J.-B. E. M., & Ramachander, S. 10.1186/1476-072X-9-49.
(2000). Effects of brand local and nonlocal origin on consumer attitudes in Morland, K., Diez Roux, A. V., & Wing, S. (2006). Supermarkets, other food stores,
developing countries. Journal of Consumer Psychology, 9(2), 83–95. http://dx.doi. and obesity: The atherosclerosis risk in communities study. American Journal of
org/10.1207/S15327663JCP0902_3. Preventive Medicine, 30(4), 333–339. http://dx.doi.org/10.1016/j.
Battersby, J., & Peyton, S. (2014). The geography of supermarkets in Cape Town: amepre.2005.11.003.
Supermarket expansion and food access. Urban Forum, 25(2), 153–164. http:// NCD Risk Factor Collaboration. (2016). Trends in adult body-mass index in 200
dx.doi.org/10.1007/s12132-014-9217-5. countries from 1975 to 2014: A pooled analysis of 1698 population-based
Bommer, C., Heesemann, E., Sagalova, V., Manne-Goehler, J., Atun, R., Bärnighausen, measurement studies with 19.2 million participants. The Lancet, 387(10026),
T., & Vollmer, S. (2017). The global economic burden of diabetes in adults aged 1377–1396. http://dx.doi.org/10.1016/S0140-6736(16)30054-X.
20–79 years: a cost-of-illness study. The Lancet Diabetes & Endocrinology, 5(6), Nelms, M., Sucher, K., & Lacey, K. (2011). Chapter 12 – Energy balance and body
423–430. http://dx.doi.org/10.1016/S2213-8587(17)30097-9. weight. In Nutrition therapy and pathophysiology. Belmont, CA, Wadsworth:
Bound, J., Jaeger, D. A., & Baker, R. M. (1995). Problems with instrumental variables Cengage Learning.
estimation when the correlation between the instruments and the endogenous Nestle, M., Wing, R., Birch, L., DiSogra, L., Drewnowski, A., Middleton, S., et al. (1998).
explanatory variable is weak. Journal of the American Statistical Association, 90 Behavioral and social influences on food choice. Nutrition Reviews, 56(5), 50–64.
(430), 443–450. http://dx.doi.org/10.1080/01621459.1995.10476536. http://dx.doi.org/10.1111/j.1753-4887.1998.tb01732.x.
Centers for Disease Control and Prevention (2007). National Health and Nutrition Neven, D., Odera, M. M., Reardon, T., & Wang, H. (2009). Kenyan supermarkets,
Examination Survey (NHANES). Anthropometry Procedures Manual. Atlanta, GA: emerging middle-class horticultural farmers, and employment impacts on the
Centers for Disease Control and Prevention. rural poor. World Development, 37(11), 1802–1811. http://dx.doi.org/10.1016/
Chege, C. G. K., Andersson, C. I. M., & Qaim, M. (2015). Impacts of supermarkets on j.worlddev.2008.08.026.
farm household nutrition in Kenya. World Development, 72, 394–407. http://dx. Ng, M., Fleming, T., Robinson, M., Thomson, B., Graetz, N., Bennett, D. A., et al.
doi.org/10.1016/j.worlddev.2015.03.016. (2014). Global, regional, and national prevalence of overweight and obesity in
Dover, R. V. H., & Lambert, E. V. (2016). ‘‘Choice Set” for health behavior in choice- children and adults during 1980–2013: A systematic analysis for the Global
constrained settings to frame research and inform policy: Examples of food Burden of Disease Study 2013. The Lancet, 384(9945), 766–781. http://dx.doi.
consumption, obesity and food security. International Journal for Equity in org/10.1016/S0140-6736(14)60460-8.
Health, 15(48), 1–8. http://dx.doi.org/10.1186/s12939-016-0336-6. Planet Retail. (2016). Country report Kenya. Retrieved February 28, 2016, from
Drewnowski, A., Aggarwal, A., Hurvitz, P. M., Monsivais, P., & Moudon, A. V. (2012). http://www.planetretail.net/Markets/Country/91.
Obesity and supermarket access: proximity or price? American Journal of Public Popkin, B. M. (2014). Nutrition, agriculture and the global food system in low and
Health, 102(8), e74–e80. http://dx.doi.org/10.2105/AJPH.2012.300660. middle income countries. Food Policy, 47, 91–96. http://dx.doi.org/10.1016/
Food and Agriculture Organization of the United Nations (FAO) (2010). World Food j.foodpol.2014.05.001.
Dietary Assessment System, Version 2.0. International Network of Food Data Popkin, B. M., Adair, L. S., & Ng, S. W. (2012). Global nutrition transition and the
Systems of FAO. Rome, Italy: Food and Agriculture Organization of the United pandemic of obesity in developing countries. Nutrition Reviews, 70(1), 3–21.
Nations. http://dx.doi.org/10.1111/j.1753-4887.2011.00456.x.
 K.M. Demmler et al. / World Development 102 (2018) 292–303 303

Qaim, M. (2017). Globalisation of agrifood systems and sustainable nutrition. Toiba, H., Umberger, W. J., & Minot, N. (2015). Diet transition and supermarket
Proceedings of the Nutrition Society, 76(1), 12–21. http://dx.doi.org/10.1017/ shopping behaviour: Is there a link? Bulletin of Indonesian Economic Studies, 51
S0029665116000598. (3), 389–403. http://dx.doi.org/10.1080/00074918.2015.1111997.
Reardon, T., Timmer, C. P., Barrett, C. B., & Berdegué, J. (2003). The rise of Umberger, W. J., He, X., Minot, N., & Toiba, H. (2015). Examining the relationship
supermarkets in Africa, Asia, and Latin America. American Journal of Agricultural between the use of supermarkets and over-nutrition in Indonesia. American
Economics, 85(5), 1140–1146. http://dx.doi.org/10.1111/j.0092- Journal of Agricultural Economics, 97(2), 510–525. http://dx.doi.org/10.1093/
5853.2003.00520.x. ajae/aau111.
Rischke, R., Kimenju, S. C., Klasen, S., & Qaim, M. (2015). Supermarkets and food Ventura, A. K., & Worobey, J. (2013). Early influences on the development of food
consumption patterns: The case of small towns in Kenya. Food Policy, 52, 9–21. preferences. Current Biology, 23(9), R401–R408. http://dx.doi.org/10.1016/j.
http://dx.doi.org/10.1016/j.foodpol.2015.02.001. cub.2013.02.037.
Schoeller, D. A. (1995). Limitations in the assessment of dietary energy intake by Westerterp, K. R. (2004). Diet induced thermogenesis. Nutrition & Metabolism, 1,
self-report. Metabolism, 44, 18–22. http://dx.doi.org/10.1016/0026-0495(95) 1–5. http://dx.doi.org/10.1186/1743-7075-1-5.
90204-X. Withrow, D., & Alter, D. A. (2011). The economic burden of obesity worldwide: A
Sehmi, J. (1993). National food composition tables and the planning of satisfactory diets systematic review of the direct costs of obesity. Obesity Reviews, 12(2), 131–141.
in Kenya. Nairobi, Kenya: Government Printers. http://dx.doi.org/10.1111/j.1467-789X.2009.00712.x.
Shim, J.-S., Oh, K., & Kim, H. C. (2014). Dietary assessment methods in epidemiologic Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (second
studies. Epidemiology and Health, 36, 1–8. http://dx.doi.org/10.4178/epih/ ed.). Cambridge, Massachusetts: MIT Press.
e2014009. World Economic Forum (2011). The global economic burden of non-communicable
Swinburn, B., Caterson, I., Seidell, J., & James, W. (2004). Diet, nutrition and the diseases. Geneva, Switzerland.
prevention of excess weight gain and obesity. Public Health Nutrition, 7(1a), World Health Organization (WHO). (2014). WHO/Europe | Nutrition – Body mass
123–146. http://dx.doi.org/10.1079/PHN2003585. index. Retrieved August 19, 2016, from http://www.euro.who.int/en/health-
Tessier, S., Traissac, P., Maire, B., Bricas, N., Eymard-Duvernay, S., El Ati, J., et al. topics/disease-prevention/nutrition/a-healthy-lifestyle/body-mass-index-bmi.
(2008). Regular users of supermarkets in Greater Tunis have a slightly improved World Health Organization (WHO). (2015). Global health observatory data
diet quality. Journal of Nutrition, 138(4), 768–774. repository. Retrieved October 26, 2016, from http://apps.who.int/gho/data/
Timmer, C. P. (2009). Do supermarkets change the food policy agenda? World node.main.A897 A?lang = en.
Development, 37(11), 1812–1819. http://dx.doi.org/10.1016/
j.worlddev.2008.08.022.