The Caloric Cost of Culture: Evidence from Indian Migrants∗
David Atkin†
February 2013
Abstract
This paper investigates whether malnutrition is partly influenced by culture. I document
that inter-state migrants within India consume fewer calories per Rupee of food expenditure
compared to their non-migrant neighbors, even for households with very low caloric intake.
I then provide two pieces of evidence in support of an explanation based on culture: that migrants make nutritionally-suboptimal food choices due to strong preferences for the traditional
foods of their origin states. First, I document that migrants bring their origin-state food preferences with them when they migrate. Second, I show that the gap in caloric intake between
locals and migrants is related to the suitability and intensity of these origin-state food preferences: the most adversely affected migrants (households in which both husband and wife
migrated to a village where their origin-state preferences are unsuited to the local price vector) would consume 7 percent more calories if they possessed the same preferences as their
neighbors.
∗ Thanks to Keith Chen, Angus Deaton, Adrian de Froment, Penny Goldberg, Gene Grossman, Dan Keniston, Nancy
Qian, Chris Udry and Jacqueline Yen.
† Department of Economics, Yale University, 27 Hillhouse Ave., CT 06511. E-mail: [email protected]
1
Introduction
Malnutrition is endemic in many poor countries. To emphasize the severity of this problem
and the need to resolve it, the Millennium Development Goals highlight ending hunger as one of
the key aims of the United Nations.1 It goes without saying that to design effective policies to end
hunger, we must first understand the determinants of hunger. Policies designed to reduce hunger
and malnutrition include the introduction of improved crop varieties, reducing tariffs to lower the
costs of food imports and food aid (UN Millenium Project Task Force on Hunger, 2005). However,
the effectiveness of these policies may be limited by the unwillingness of locals to switch away
from the types of food that they traditionally consume.
For example, many observers have noted that the strong preference for white maize over yellow maize in Africa has reduced the effectiveness of imported yellow maize as food aid, as well
as slowed the adoption of maize that is biofortified with Vitamin A (a process that turns the maize
yellow).2 Outside of the context of aid policies, Harris (1985) argues that beef avoidance taboos
established during the first millennium AD contributes to malnutrition amongst Hindus today.
Similarly, a well-known example from a historical context is when many hungry Irish refused
to eat the American maize provided by the British as food aid during the Irish potato famine
(Woodham-Smith, 1964). In all these cases, strongly-held preferences for local foods have deleterious consequences for nutrition.
Since the types of food that a group of people traditionally consumes embodies the preferences,
beliefs and social attitudes of the group, I will describe differences in food preferences across
groups as different cultures.3 That culture may also constrain caloric intake and contribute to
malnutrition is a question of first order importance.
The main goal of this paper is to determine whether low caloric intake is partly influenced by
culture. I focus on India, where the population is poor and hunger is prevalent and ask two simple
questions: i) to what extent does culture influence caloric intake?; and ii) does culture constrain
caloric intake even for households on the edge of malnutrition? The empirical results will allow
me to approximate the caloric “cost” of culture.
My analysis proceeds in four stages and draws on household survey data collected by the
National Sample Survey Organization of India in 1983 and 1987-88. These data prove ideal for
such an analysis because they contain extremely rich food consumption data for hundreds of
thousands of households, matched to detailed migration data on each household member.
1 See
http://www.un.org/millenniumgoals/.
2 See McCann (2005), Muzhingi et al. (2008) and De Groote and Kimenju (2008).
Barrett and Maxwell (2005) presents
the starker example from the 2000 drought in Northern Kenya. Maize was given to pastoralists with a strong preference
for milk, meat and tea and as a consequence was not eaten but turned into alcohol instead.
3 This follows the working definition proposed by Fernandez (2010) “consider differences in culture as systematic
variation in beliefs and preferences across time, space or social groups”. Note that definitions of “culture” vary slightly
across studies. Guiso, Sapienza and Zingales (2006) define culture as “those customary beliefs and values that ethnic,
religious, and social groups transmit fairly unchanged from generation to generation”. Given the strong persistence
in food preferences across generations described in Atkin (Forthcoming), both definitions of culture are applicable.
Section 2.2 discusses why differences in food preferences can be interpreted as cultural differences in more detail.
1
In the first stage of the analysis, I provide the background to my analysis. I focus on India during a period when childhood malnutrition rates were above 50 percent and 64 percent of households consumed fewer calories than the nutritional adequacy requirements used to determine
India’s poverty line. If undernourished households are not maximizing nutrition in this context,
it is important to understand why. The example of rice and wheat provides suggestive evidence
that households are not maximizing their nutrition and that culture is in part responsible. Despite these two cereals being nutritionally similar, there is dramatic regional variation in rice and
wheat consumption. For example, although the relative price of rice and wheat is similar in the
states of Bihar and Gujarat, Biharis consumed one and a half times more rice than wheat and
Gujaratis two and a half times more wheat than rice. As shown in Atkin (Forthcoming), these
regional preferences stem in part from historical consumption patterns (themselves determined
by the relative suitability of agro-climatic endowments for growing rice or wheat). In terms of the
costs of these cultural preferences for either rice or wheat, a crude counterfactual shows that mean
Indian caloric intake could be 6.2 percent higher if Indian households in certain regions switched
around the quantity of rice and wheat that they were consuming, and spent any cost savings on
the cheaper of the two grains.
Although suggestive, such an approach cannot deal with the possibility that foods which are
strong complements with either wheat or rice have different prices in different parts of the country.
The behavior of inter-state migrants provides more compelling evidence that culture can constrain
caloric intake. The key observation for my empirical strategy is that migrants face the same prices
as their neighbors but are likely to bring these cultural food preferences with them when they
migrate. A quirk of Indian data collection procedures ensures that households are surveyed in
groups of ten drawn from blocks of no more than 180 proximate households within a village,
town or city. This feature allows me to compare migrant and non-migrant households who face
very similar prices. The finding that migrants consume a different number of calories per Rupee of
food expenditure compared to otherwise similar locals provides prima facie evidence that culture
can constrain caloric intake.
The second stage of the analysis contains my main result: migrant households consume fewer
calories per person compared to non-migrant households in the same village, conditioning on
household food expenditures, characteristics and demographics in a flexible manner. The average level of this “caloric tax” (the percentage gap in caloric intake between locals and migrants)
is equal to 1.6 percent of caloric intake. Reassuringly, I find a similarly sized caloric tax when I
restrict attention only to households in which the wife moved village at the time of marriage, and
compare households where the wife migrated across a state boundary to households where the
wife moved within her own state.4 I also find no evidence that the caloric tax is restricted to wellnourished households for whom reductions in caloric intake have no nutritional consequences:
the magnitude of the tax ranges between 1 and 1.7 percent when I only consider households that
4 These two types of households are more similar in terms of observables, and such a comparison mitigates potential
selection problems that arise when household heads are choosing to migrate for better employment opportunities or
because they are particularly adaptable to different cultures.
2
are particularly poor or malnourished, or households with small children for whom caloric shortfalls are particularly harmful.
The third stage of the analysis investigates why migrants consume fewer calories than nonmigrants. I provide a chain of evidence that suggests migrants are making nutritionally-suboptimal
food choices due to cultural preferences for the traditional foods of their origin states. First, I document that migrants bring their origin-state food preferences with them. In particular, I show that
compared to other households in the same village, the food-budget shares of a migrant household
are more-closely correlated with the average food-budget shares of their origin state. Second, I
show that the heterogeneity in the size of the migrant caloric tax is related to the suitability and
intensity of their origin-state food preferences: the caloric tax is only present when the average
bundle of the migrant’s origin state provides fewer calories than the local bundle (both priced
at the village price vector), and increases in size when both husband and wife are migrants (as
opposed to just one of these two being migrants). In terms of magnitudes, the most adversely
affected households (where both husband and wife migrated to a village where the origin-state
reference bundle provides fewer calories than the local bundle) face a caloric tax of 7.2 percent.
This chain of results suggests that migrants in India consume fewer calories than non-migrants
because they prefer to purchase the traditional products from their origin state even when these
products are relatively expensive compared to alternatives available in their destination state.
Finally, the fourth stage of the analysis rules out two competing explanations that do not rely
on migrant and non-migrant households having different preferences. The first competing explanation is that migrants have poor information about the local alternatives to their origin-state
foods. If this were the case, the caloric tax should not persist many years after migration; the size
of the tax should be smaller for more educated households; and the tax should not be present if
there are non-migrants in the household who are familiar with the local foods. I find no evidence
for any of these hypotheses. The second competing explanation is that migrants do not posses
the technologies, such as cooking equipment or recipes, needed to make delicious meals from the
locally-cheap foods. As above, if this were the case, the caloric tax should disappear for long-term
migrants or for migrants who moved into households containing non-migrants. Similarly, if a lack
of recipes or preparation techniques were the cause of the caloric tax, the tax should be smaller
when only the husband is a migrant compared to when only the wife is a migrant (since women
are typically in charge of food preparation in Indian households). In fact, I find the opposite result,
consistent with migrants bringing their origin-state preferences with them and husbands having
greater bargaining power in household decision making.
For policymakers, this paper shows that effective policies for relieving hunger should take
culture into account. I discuss policies in more detail in the conclusion.
This paper contributes to several literatures. First, it adds to the growing literature on the
importance of culture surveyed in Guiso, Sapienza and Zingales (2006) and Fernandez (2010).
In using the behavior of migrants to examine the influence of culture on household decisions, it
is particularly closely related to Fernandez, Fogli and Olivetti (2004) and Fernandez and Fogli’s
3
(2009) studies of female labor force participation and Giuliano’s (2007) study of family living arrangements. As nutrition has direct impacts on economic growth (Fogel, 1994), the paper is closely
related to Algan and Cahuc’s (2010) study of the relationship of culture and economic growth; and
Guiso, Sapienza and Zingales’s (2004) study on the role of culture and financial development.
Second, I add to the literature on the persistence of regional food preferences initiated by
Staehle (1934), with recent contributions by Bronnenberg, Dube and Gentzkow (2012) and Logan
and Rhode (2010). Although these papers document that migrants bring their food preferences
with them, none of them explore the nutritional consequences of such preferences. Finally, this
study is also related to Nunn and Qian’s (2011) study of the impact of New World potatoes on
Old World population. Their finding of large take up of a new crop over hundreds of years and
consequent nutritional improvements suggests that the persistence of food culture that emerges
from my analysis may be finite in duration.
In a companion paper, I provide theoretical and empirical evidence for the existence of regional
food preferences in India.5 The two papers differ in that Atkin (Forthcoming) lays out a model
in which the combination of agroclimatic endowments and habits generate regional food tastes
that favor the locally-abundant foods, and then explores the implications for the gains from trade
liberalization. In contrast, this paper takes India’s regional food preferences as given, interprets
these as cultural phenomena and investigates their influence on caloric intake.
I layout the remainder of the paper in the following manner. Section 2 introduces the dataset
and documents that households are basing food purchasing decisions in part on culture despite
being close to malnutrition. Section 3 explains how I use migrants in my empirical strategy to
identify the caloric costs of culture. Section 4 presents my main result, that migrant households
consume fewer calories compared to comparable non-migrant households. Section 5 presents
evidence that this finding is driven by migrants making nutritionally-suboptimal food choices due
to preferences for the favored foods of their origin states. Section 6 rules out the two competing
explanations. Finally, section 7 discusses the implications of my findings and concludes.
2
Background on consumption patterns and caloric intake across India
In this section, I present two findings that motivate the subsequent analysis. First, the median
Indian household is consuming far below the daily recommended caloric intake, consistent with
the extremely high levels of childhood malnutrition from health and nutrition surveys. Second,
despite this setting, households seem to be making food choices driven by culture even when
these choices result in reduced caloric consumption.
Since households in different regions face different prices, it is difficult to infer the costs imposed by these culturally-determined food preferences without strong assumptions about the
structure of preferences. In the next section, I argue that the behavior of inter-state migrants allows
me to obtain estimates of the caloric cost imposed by these cultural preferences. Such an approach
5 Both
this paper and Atkin (Forthcoming) formed part of my Ph.D. thesis.
4
requires a dataset that contains detailed information on household food consumption as well as
identifying whether household members are migrants, and if so where they migrated from. This
section also introduces the Indian National Sample Survey (NSS), a dataset that satisfies both these
criteria.
My analysis draws on two cross-sections of the Indian NSS: the 38th round (1983) and the
43rd round (1987-1988). These are the only two rounds of publicly available surveys in which the
same household is asked both about their consumption of a broad set of foods as well as about
their migration particulars.6 Each survey round contains approximately 80,000 rural households
(in 8,000 villages) and 45,000 urban households (in 4,500 urban blocks), resulting in a combined
data set containing 245,334 households.7 The NSS provides survey weights in order to make the
sample nationally representative, and all the statistics I report use these weights. The surveys
record expenditures and quantities purchased in the last 30 days (as well as quantities consumed
out of home-grown stock and gifts valued at the prevailing local prices) for up to 169 different
food items. I obtain calorie data for each household by multiplying each food’s caloric content,
estimated by the Indian National Sample Survey Organization (NSSO), by the quantity consumed
over the previous 30 days. I use this number (along with the size of the household from the
household roster) to calculate the daily caloric intake per household member. The surveys also
provide information on household demographics and characteristics, as well as expenditures on
non-food items.
2.1
Malnutrition in India
The first fact that jumps out of the data is that Indian households consume a small number
of calories. The mean caloric intake is 2224 per person per day across the two samples. In order
to get a sense of magnitudes, the equivalent figures for the United States and China in 2006-2008
were 3750 and 2990 calories respectively.8
India drew its poverty lines in 1978 based on the calorie norms required for a nutritionally
adequate diet. These norms were set at 2400 calories per person per day for rural India and 2100
calories per person per day for urban India.9 Using this simple indicator of household malnutrition, the caloric intake of 66.4 percent of rural households and 59.6 percent of urban households
in my combined sample fall below these calorie norms. Many households lie substantially below
these levels with 45.5 percent of households consuming fewer than 2000 calories and 35.6 percent
6 The
migration questions are part of the employment and unemployment schedule, schedule 10. In contrast, the
consumption data are collected in the Consumer Expenditure schedule, schedule 1. In more recent survey rounds,
these two schedules were no longer filled out by the same households.
7 Both the household and village/urban block identifiers are anonymized and so even if the same household or
village was surveyed in both rounds, it cannot be matched across rounds.
8 These
figures
come
from
the
Food
and
Agricultural
Organization
and
also
measure
the
amount
of
food
available
for
human
consumption
(and
so
do
not
account for wastage or food fed to household pets etc.).
These data are available at
http://www.fao.org/fileadmin/templates/ess/documents/food_security_statistics/FoodConsumptionNutrients_en.xls.
9 Poverty lines were set equal to the average per-capita expenditure of the households in the NSS that consumed
this quantity of calories. In subsequent years these poverty lines were simply updated using the rate of inflation.
5
consuming fewer than 1850 calories per person per day. The upper panel of Figure I shows the
full distribution of caloric intake in my sample.
While there is an imperfect mapping between measures of caloric intake and malnutrition,10
these low levels of caloric intake are consistent with the extremely high child-malnutrition rates
in India. The first wave of the National Family Health Survey was administered in 1992-1993. The
survey measured and weighed around 35,000 children under age 4 and found that 55.9 percent of
rural children and 45.2 percent of urban children were moderately to severely underweight, and
54.1 percent of rural children and 44.8 percent of urban children were moderately to severely
stunted.11 These numbers imply a higher prevalence of under nutrition than in Sub-Saharan
Africa (Deaton and Drèze, 2009), and suggest that a substantial number of Indian households
were living on the edge of malnutrition at the time of the surveys.
2.2
Cultural Preferences for Food
In this subsection, I provide suggestive evidence that despite this context, households seem
to be making nutritionally-suboptimal food choices in part due to cultural preferences for certain
foods.
The first thing to note is that almost all households could have purchased a bundle providing
the recommended caloric norms (2400 calories for rural and 2100 calories for urban households)
for a smaller outlay than they actually spent on food. For example, if every household spent
all their income on the item in their consumption bundle that provided the most calories per
Rupee, then the percentages of households below these thresholds would fall to 11.1 percent of
rural households and 4.0 percent of urban households (compared to the 66.4 percent and 59.6
percent in the actual data). Alternatively, if every household spent the same amount per calorie
as the most efficient household in their village, only 41.3 percent of rural households and 31.7
percent of urban households would be consuming nutritionally inadequate diets. However, both
of these bundles are likely to be deficient in other nutrients such as protein, fats and vitamins,
as well as highly monotonous. Therefore, the fact that households are not purchasing the most
calorically-efficient bundle does not imply that they are making food choices based on cultural
considerations. Households may simply be maximizing nutrition and a taste for variety.
The consumption of rice and wheat provides more convincing evidence that households might
be making nutritionally-suboptimal food choices due to culture. Rice and wheat are the two dominant carbohydrates in India, together accounting for an enormous 65.5 of total caloric consumption in my sample period. There is also strong empirical evidence of regional differences in preferences for these two foods. The upper panel of Figure II plots the relative consumption of rice and
10 These
numbers are likely to overestimate actual caloric intake. Some of these calories are wasted (due to spoilage
or simply thrown away) or fed to servants and guests. Additionally, actual caloric intake will be lower than measured
caloric intake if household members have ailments such as diarrhea or gastroenteritis that prevent the body from fully
absorbing the calories.
11 The WHO defines moderately to severely underweight as two or more standard deviations below the median
weight-for-age of an international reference population. Stunting is defined in the same manner but using height-forage.
6
wheat calories against the relative price of rice and wheat calories for the 22,148 villages where I
observe both rice and wheat purchases.12 Although the relationship between relative price and
relative quantity is downward sloping, by drawing horizontal lines across the plot it is obvious
that for most of the observed price ratios there are some villages consuming relatively large quantities of rice and others relatively large quantities of wheat. This observation applies both when
rice is a relatively cheap calorie source, and when it is a relatively expensive calorie source.
Atkin (Forthcoming) provides an explanation for these patterns of regional preferences. The
paper combines habit formation in food consumption (in particular, adult preferences that favor
the foods consumed as a child) with a specific-factors model of agricultural production. Over
many generations households develop preferences for the foods that their region is well-suited to
producing, as these foods were relatively inexpensive in the past and so consumed in large quantities. As predicted by this model, Atkin (Forthcoming) finds that for India, regional preferences
for particular foods are positively related to food-specific measures of land suitability.
This explanation accords well with the common perception that Indians from the north and
north-west of the country prefer wheat and Indians from the south and east prefer rice. Wheat
grows best in well-drained soil and at moderate temperatures, while rice thrives at higher temperatures and submerged in water. In terms of ago-climatic conditions, much of the north and
north-west is classified as arid or semi-arid, while much of the south and east is classified as humid subtropical and tropical wet. Hence, the relative preferences of northern Indians for wheat
and south and eastern Indians for rice fits well with an explanation based on historic consumption
driven by land suitability.
These regional taste differences are apparent in the data when villages are grouped by state
in the lower panel of Figure II. I include only the villages in four states. Rajhastan in the north
consumes far more wheat than rice, and Andrah Pradesh in the south consumes far more rice
than wheat. However, in almost all villages in these two states the heavily-consumed food is
also relatively inexpensive. More convincingly, in Gujarat in the north-west and Bihar in the east,
relative prices are almost identical. Yet, Gujaratis consumed two and a half times more wheat than
rice, and Biharis consumed one and a half times more rice than wheat.
Regional taste differences of this kind fit squarely within the definitions of culture used in the
economics literature. Fernandez (2010) defines differences in culture as “systematic variation in
beliefs and preferences across time, space or social groups”. Since India is comprised of more than
two thousand ethnic groups generally concentrated in particular geographical areas, the systematic variation in food preferences across regions of India certainly fits within this definition. Guiso,
Sapienza and Zingales (2006) define culture as “those customary beliefs and values that ethnic, religious, and social groups transmit fairly unchanged from generation to generation”. Since adult
12 Only 2698 villages are excluded because either no rice or no wheat is purchased. These villages would likely
strengthen my findings if they were included since they possess extreme preferences for rice or wheat, however no
prices are available for unpurchased foods in these villages. Relative consumption is simply the log of the ratio of the
total calories of rice and wheat consumption amongst sample households in the village. The relative price ratio is the
log of the ratio of median prices per calorie across purchasing households in the village.
7
food preferences are in part determined by the foods consumed in childhood, and adults choose
which foods are fed to their children, the systematic variation in food preferences across regions
of India also fits this second definition of culture.13
Do cultural preferences for rice and wheat mean that households are consuming fewer calories
than they could if they were just maximizing nutrition and dietary variety? In order to answer this
question, I perform the following counterfactual. For every household, I calculate the amount it
would cost to purchase the same quantity of calories that they actually obtain from rice and wheat
but switching around the quantities of rice and wheat calories they purchase.14 For the subset of
households for whom this amount is less than they actually spend on rice and wheat, I calculate
the hypothetical caloric intake if the household allocated the cost savings to the cheaper of the
two foods. The gap between actual and hypothetical caloric intake provides a measure of the
number of calories households are forgoing in order to indulge in their cultural preferences for
rice or wheat. Of course, if many households did switch across rice and wheat, local prices would
change and the actual gain in caloric intake would likely be smaller.
This counterfactual solves several of the problems associated with the cruder counterfactuals
at the start of this subsection. First, by switching the quantities allocated to rice and wheat, the
amount of dietary variety is not changing (every household still obtains the same number of calories from the less-consumed food). Second, these two foods provide a similar number of calories
per Rupee and contain a similar set of nutrients. Therefore, the switching between rice and wheat
does not alter household nutrition along other dimensions.15
By switching across rice and wheat in this manner, 42.0 percent of households could increase
their caloric intake. The upper panel of figure I overlays the counterfactual distribution of caloric
intake for all households in villages where I observe both rice and wheat purchases. The lower
panel displays a histogram of the caloric gains reaped by the households who switched their
rice and wheat consumption in the counterfactual exercise. On average, households are forgoing
6.2 percent higher caloric intake (which if realized would reduce the percentage of households
consuming less than the recommended caloric norms by 6.3 percent in rural areas and 7.3 percent
in urban areas). Almost all of these potential caloric gains come from rice-loving households
switching into relatively cheap wheat since wheat is a cheaper calorie source on average (94.2
13 Birch (1999) surveys the evidence from the psychology and nutrition literatures indicating that adult food preferences form through consumption in childhood. For example, experimental evidence has shown that a mother’s diet
during pregnancy and lactation affects her child’s preferences for flavors and foods in later life (Mennella, Jagnow and
Beauchamp, 2001), and the preferences gained in childhood persist into adulthood in the available longitudinal data
(Kelder, 1994).
14 For households that do not consume both rice and wheat, I use the village median price for the price of the food
that they do not purchase. As previously, I restrict attention to the 22,148 villages where I observe both rice and wheat
purchases.
15 The average unit value for rice was 0.95 Rupees per 1000 calories in 1983 and 1.09 in 1987/88. The equivalent
numbers for wheat were 0.66 and 0.75. According to the NSS, rice contained 3460 calories per kilogram, 75 grams of
protein and 5 grams of fat. The equivalent numbers for wheat where 3410, 121 and 17. Wheat has slightly more fiber
and vitamins, particularly folic acid. Thus, if anything, wheat is more nutritious than rice. As shown below, the caloric
gains derive from households switching from rice to wheat consumption, and so nutrition is also likely to improve
along other dimensions with the switch.
8
percent of switchers consume relatively more rice than wheat).
This last exercise provides suggestive evidence that households in India make food choices
based on cultural preferences that are suboptimal from a nutritional sense. However, since every village has a different vector of relative prices, some of these consumption patterns may be
rationalized by complex substitution patterns. For example, if coconuts are much stronger complements to rice than they are to wheat, the relatively high consumption of rice in the south of
India may be due to cheap coconuts rather than a preference for rice. In the next section, I discuss
a methodology that sidesteps this issue by focusing on consumption differences between households that have different regional cultural preferences yet face the same set of relative prices.
3
Empirical methodology: examining the behavior of inter-state migrants
The behavior of inter-state migrants provides more compelling evidence that culture can con-
strain caloric intake. The key idea is that migrants bring the regional preferences of their originstate with them, yet they face the same relative prices as non-migrant households in their destination state.16 In terms of the concerns raised in the counterfactuals above, there is no reason to
think that migrant households value other dimensions of nutrition or variety more or less than
non-migrant households. Therefore, if migrant households consume bundles that resemble the
consumption bundle of households in their origin state, and these bundles provide fewer calories
than the typical local bundle, I can interpret this caloric difference as a “caloric tax” that migrants
pay to indulge in their cultural food preferences.
Inevitably, this methodology can only estimate the tax that actual migrants pay. The size of
this tax may be much larger for a non-migrant household that is forced to migrate. If potential
migrants are aware of this cost, the set of actual migrants are likely to posses particularly openminded or flexible preferences.
At this juncture it is useful to describe the migration information contained in the 38th round
and 43rd rounds of the NSS. By design, the survey only records permanent migrations as opposed
to temporary migrations for seasonal work opportunities. The survey asks whether the enumeration village (or urban block) differs from the household member’s “last usual residence”. If so,
the household member is asked the reason for migration, how long ago they migrated and the
state in which their last usual residence was located. I define inter-state migrants as households
in which either the household head or their spouse emigrated from another one of the 31 states
in India. Except where noted otherwise, I use the household head’s migration information if both
the household head and the spouse emigrated. Since in 99.7 percent of cases the household head
is male and the spouse is female, in this paper I use the terms household head and husband inter16 There is a long history of using migrants to identify the effects of culture.
Fernandez (2010) describes the approach
in detail and reviews this literature. The use of migrants to highlight the persistence of regional food preferences was
pioneered by Staehle (1934).
9
changeably, and the terms spouse and wife interchangeably.
Table I provides summary statistics for the dataset. Under the migrant definitions above, about
6.1 percent of households are classified as migrant households. Migrant households are disproportionately urban with more than half of migrant households living in urban areas compared to
under a quarter of households in the population as a whole. Most of these households are longterm migrants, with only 15.2 percent having been in their new location for fewer than 5 years and
41.3 percent having been in their new location for 20 or more years. Finally, a large proportion of
migrant households (41.3 percent) are those in which only the spouse of the household head is a
migrant. In a further 32.5 percent of migrant households, both the head and spouse are migrants
from the same origin-state. In contrast, in only 13.2 percent of cases is the the household head the
only migrant, consistent with norms of “patrilocality” in India.17 Later in this section, I describe
these norms and explain how I exploit them to control for potential confounding factors.
The identification argument in the sections that follow relies on the assumption that migrants
and non-migrants living in the same geographic location face the same prices and external environment, have the same desire for good nutrition and dietary variety and differ only in their
food preferences (after controlling for various households expenditure measures, household demographics and observable household characteristics). Fortunately, the survey collection methodology employed by the NSSO ensures that this comparison can be carried out at an extremely
high level of geographic disaggregation, making such an assumption more tenable. In each survey round the NSSO draws a sample of around 8,000 rural villages and 4,500 urban blocks from
the 1981 census rolls and surveys 10 households in each village/block. In order to reduce the
work load for the survey enumerators, any village or urban block with more than 1200 inhabitants (approximately 180 households) is subdivided into smaller geographical subgroups and
only one subgroup is surveyed. Thus, in a village with one surveyed migrant, I compare the
migrant household to the other nine non-migrant households surveyed in the same enumeration
subgroup (a subgroup which never encompasses more than about 180 proximate households).
Table II compares household characteristics across migrant and non-migrant households within
the same village. Each row contains the characteristic mean as well as the coefficient on a migranthousehold dummy when the characteristic is regressed on a village-round fixed effect and a
migrant-household dummy. Compared to other households in their village/urban block, migrant households have about 6.2 percent higher per-capita expenditures on all goods, as well as
4.5 percent higher per-capita food expenditure. Along with the higher food expenditure, they
consume 1.3 percent more calories per person. Migrant households are slightly smaller, contain
a larger proportion of prime aged males, are less likely to be categorized as an agricultural laborer household, and are more likely to be categorized as urban self-employed. In my empirical
specifications, I will include explicit controls for all of these characteristics. Finally, the last two
rows confirm that migrants do face the same prices as non-migrants, at least after controlling for
17 The
two omitted categories are migrant households were there is no spouse and migrant households where husband and wife are migrants from different states.
10
observables. The first row shows the coefficient on a migrant-status dummy when the log price
per calorie is regressed on a product-village-round fixed effect and a migrant-status dummy. The
second row shows the same coefficient but with the full set of controls for food expenditure and
household characteristics described later in the paper (essentially the other variables in Table II).18
Migrants do seem to pay about one third of one percent more than non-migrant households in
the same village buying the same product, presumably by buying slightly higher quality levels
or different varietals. However, this difference disappears once I control for the fact that migrant
households have higher expenditures, as well as different demographics and characteristics, compared to other households in their village.
My empirical work explicitly controls for the fact that migrant and non-migrant households
differ slightly on observables even within the same village. However, as a robustness check I draw
on two alternative samples in which migrant and non-migrant households appear far more similar
along observable dimensions. These two samples also mitigate the potential selection problems
that arise when household heads are choosing to migrate for better employment opportunities or
because they are particularly adaptable to different cultures.19
I take advantage of the fact that a substantial proportion of migration in India is driven by
women moving to their husband’s village at the time of marriage (Srinivas, 1980). This norm of
“patrilocality” is so prevalent that 57 percent of the wives of household heads report that both
their current village is not their last usual residence and list the reason for leaving that location
as “on marriage”. Most of these moves occur within the same state, with wives crossing a state
border at the time of marriage in only 6 percent of cases. I exploit this variation by focusing just on
households in which the wife moved for marriage, and comparing households in which the wife
moved inter-state (migrant households) to households in which the wife moved intra-state (nonmigrant households). (Throughout the paper, I use the term “migrant household” to exclusively
refer to households in which a member has moved between states of India.)
In particular, the “wife moved for marriage sample A” focuses on households in which the wife
of the household head moved specifically for the purpose of marriage (either within or between
states) and I exclude the households in which both the husband and the wife moved at the same
time. In order to ensure that I am not simply picking up men moving for work opportunities and
then later bringing a wife from their home village, the "wife moved for marriage sample B" further
restricts the set of observations by excluding households in which the husband also moved (either
inter- or intra-state) at any point in their life. Hence, these wives are moving to their husbands
birth village at the time of marriage.
18 Since there are multiple products for each household, there are 5,601,266 observations in this regression (5,576,186
once the controls are included). Unlike the previous characteristic regressions which are weighted by household survey
weights, these regressions are weighted by the household survey weights multiplied by the food-budget share spent
on that product. This weighting ensures that more important prices are weighted more heavily, as well as ensuring that
the sum of weights for each household is equal to the household survey weight from the NSS.
19 To produce my findings on caloric intake, migrants need to consume higher price per calorie foods than nonmigrants with similar incomes for reasons unrelated to the tastes of their origin state. The prior may be that the selection
bias works in the other direction. For example, migrants may be more likely to be manual laborers who consume diets
heavy in cheap carbohydrates, or migrants may have unusually adaptable and adventurous tastes.
11
Table I presents descriptive statistics for these two subsamples. In contrast to the main sample, migrant households in these two subsamples are substantially more likely to be rural and
substantially less likely to have migrated less than five years previously. Table II confirms the
conjecture that migrant and non-migrant households are more similar when I restrict attention to
the sample of wives moving at the time of marriage. Although migrant households still spend
slightly more on both food and all goods, the difference in expenditures between migrant and
non-migrant households declines by between one third and two thirds. Migrant households are
no longer smaller, nor do they contain a larger proportion of prime aged males. Finally, migrants
pay less not more than non-migrants for the same product, although these differences are miniscule and insignificant either with and without controls.
4
Migrant households consume fewer calories than comparable nonmigrant households
In this section, I present the main finding of the paper: that migrant households pay a “caloric
tax”. In particular, I test the hypothesis that migrants consume fewer calories per Rupee of food
expenditure compared to the non-migrant households living around them.
In order to test this hypothesis, I use the survey consumption data to generate ln caloriesi , the
log of caloric intake per person per day, for every household (where i indexes households). I
regress this measure of caloric consumption on migranti , a dummy variable for a migrant household. Additionally, I include village-round fixed effects dvt (where v denotes village or urban block
and t denotes the survey round), and a vector Xi of household-level controls:
ln caloriesi = β 1 migranti + dvt + Πt Xi + ε i .
(1)
The vector Xi contains a third-order polynomial in the log of the per-capita food expenditure
over the previous 30 days, as well as a comprehensive set of household demographics and characteristics that follow the specification used by Subramanian and Deaton (1996). Household demographics are captured by log household size as well as the proportion household members that
fall into five sex-specific age buckets.20 The included household characteristics are indicator variables for the household’s primary activity among the following categories: rural self-employed
in agriculture, rural self-employed in non-agriculture, rural agricultural labor, rural other labor,
rural other, urban self-employed, urban wage earner, urban casual labor and urban other. I allow the coefficients on all these control to differ by survey round by interacting each control with
a survey-round fixed effect. Subramanian and Deaton (1996) also include indicators for religion
and whether the household is a member of a scheduled caste. Since religious affiliation and caste
membership may be a cultural determinant of food preferences, I do not include these as controls.
20 These buckets are males aged 0-4, males aged 5-9, males aged 10-15, males aged 15-55, males aged over 55, females
aged 0-4, females aged 5-9, females aged 10-15 and females aged 15-55. The omitted demographic category is females
aged over 55.
12
However, results are essentially unchanged by their inclusion.
As Xi includes a polynomial in log per-capita food expenditure, I am testing whether migrants
consume fewer calories than their neighbors in the same village, conditional on their food expenditure and other household-level characteristics (the hypothesis β 1 < 0). Given my specification,
this test is exactly equivalent to asking if migrants purchase more expensive calories than nonmigrants in their village, conditioning on Xi . The characteristic and demographic variables control for the possibility that, compared to other households in the same village, migrant households
may work in less physically intensive jobs or have different demographic structures.
Column 1 of Table III shows the results of this regression. I reject the null hypothesis, that migrants and non-migrants consume an equal or greater number of calories for a given level of food
expenditure, at the 1 percent level: inter-state migrant households are consuming 1.59 percent
fewer calories than their non-migrant neighbors, controlling for food expenditure. In monetary
terms, this caloric “tax” on migrants is commensurate with the caloric decline due to a 2.65 percent reduction in food expenditure for the average migrant household in 1987/88 or a 2.27 percent
reduction in 1983.21
As discussed in section 2, caloric intake is not equivalent to nutrition, and households may
trade off calorie-rich foods for protein- or vitamin-rich foods. However, there is no reason to
believe that migrants would trade-off these components of a nutritious diet in a different manner
to non-migrants who face the same vector of prices and spend a similar amount of money on food.
Thus, a caloric tax on migrants likely implies lower nutrition as well.
The magnitudes of the coefficients do not imply that preference differences between migrants
and non-migrants can only generate small caloric costs. First, under a preference-difference story,
the size of the caloric tax should depend on how costly it is for a migrant to indulge in their originstate food preferences. If migrant preference are well-suited to the local price-vector, migrants
may actually consume more calories for a given level of food expenditure. For example, a migrant
from wheat-loving Rajasthan moving to rice-loving West Bengal will consume more calories than
comparable locals if wheat is a cheaper calorie source than rice in the destination village. The
coefficient on migranti merely summarizes the average caloric tax faced by migrants traveling
along a multitude of routes and facing very different caloric taxes.22 Second, recall that for many
of these households only one member of the household (usually the wife) migrated from another
state. Under a preference-difference story, any effects will be more exaggerated if both husband
and wife are migrants since a greater proportion of the household possess non-local preferences.
Sections 5.2 and 5.3 provide support for the preference-difference story by exploring these two
dimensions of heterogeneity.
21 I
calculate these numbers using the elasticities implied by the round-specific coefficients on the expenditure controls combined with the mean log per-capita food expenditure of migrants in each round.
22 The average caloric tax faced by migrants will be negative if preferences adapt through the process of habit formation to favor whichever foods are locally inexpensive as in Atkin (Forthcoming).
13
4.1
Robustness checks
The remaining columns of Table III report a variety of robustness checks that support my find-
ing that migrant households consume fewer calories conditional on food expenditure. Columns 2
and 3 of Table III run the specification in equation 1 on the two “wife moved for marriage” subsamples (described in section 3). In these subsamples, I am comparing households where wives
moved intra-state at the time of marriage (a non-migrant household) with those households where
the wife moved inter-state (a migrant household). Column 2 restricts attention to wives whose
husbands did not also move at the time of marriage (sample A), while column 3 further restricts
attention to wives whose husbands are still living in their birth village (sample B). The caloric tax
on migrants is significantly negative for both these subsamples, but is attenuated by between 24
and 31 percent. The decline in the size of the coefficient is not surprising. Since these wives are
moving into their husband’s households (often containing other extended family members such
as the husband’s parents), any regional preferences that the migrant wife brought from her origin
state are likely to have a smaller impact on household spending decisions compared to the scenario where both the husband and wife migrated from that origin state. This attenuation is larger
in sample B as this sample also excludes households where the husbands are migrants but moved
at a different time from their wives.23
Columns 4 to 6 of Table III reproduce this caloric tax finding using alternative sets of expenditure controls in place of the polynomial in log per-capita food expenditure. Column 4 uses a
third-order polynomial in log per-capita expenditure on all goods. I find a caloric tax on migrants
of about 1.36 percent. Since the coefficients are of similar magnitude when I control for either food
expenditure or total expenditure, it is not the case that migrant households are simply substituting
from non-food to food expenditure in order to indulge their origin-state food preferences. Results
are also significant and similar in magnitude in both column 5, which uses a polynomial in log percapita real food expenditure, and in column 6, which uses a flexible specification to accommodate
for the fact that price indexes may differ across state.24
I report two further robustness specifications. In the first of these, I instrument the polynomial
of log per-capita food expenditure with the polynomial of log per-capita non-food expenditure
to control for potentially correlated measurement error (since calories and food expenditure are
calculated using the same raw data). Food expenditure may also be endogenous. For example, a
shock that increases the demand for calories, such as changing work patterns, will also affect food
expenditure and result in a positive correlation between food expenditure and the error term,
biasing the coefficients on food expenditure upwards. However, there will be a negative or no
correlation with non-food expenditure, and so the true value of the coefficient will be bounded
23 I would also find β < 0 if wives from more distant villages are more prized and fed higher quality foods. Alter1
natively, I would find β 1 < 0 if wives are both fed (or choose) cheaper calorie sources than other household members,
combined with wives from further away being fed less food (or controlling a smaller share of the household budget).
24 Column 5 uses log per-capita food expenditure divided by a state-specific Stone price index (the sum of log prices
weighted by state budget shares). Column 6 interacts log per-capita food expenditure with state-round fixed effects,
thereby allowing any relationship between the dependent variable and income to vary by state.
14
between the instrumented and uninstrumented estimates. These results are shown in column
7 of Table III. Since β 1 is only attenuated by about a quarter and still significantly negative in
the instrumented specification, both measurement error and the potential endogeneity of food
expenditure do not seem to be a major problem in this context.25
Finally, column 8 replaces ln caloriesi with ln calories_per_Rupeei , the household’s caloric intake per Rupee spent on food, and includes controls for per-capita expenditure on all goods (recall
that the coefficient on migranti would be identical to column 1 if food expenditure controls were
included instead). As in my main specification, I find that migrant households obtain 1.45 percent
fewer calories per Rupee of food expenditure than non-migrant neighbors after conditioning on
total expenditure.
4.2
Migrants consume fewer calories even when on the edge of malnutrition
One possibility is that migrants are willing to indulge in their cultural preferences but only if
they are sufficiently rich and well-nourished such that any foregone calories are irrelevant (and
may even be beneficial). Accordingly, Table IV repeats the basic specification using various subpopulations that are poor and under-nourished (i.e. I compare the caloric intake of poor and
undernourished migrant households to that of poor and undernourished non-migrant households
in the same village).26
Column 1 repeats the baseline specification. Columns 2 through 4 restrict attention to undernourished households (those consuming fewer than either 1850 or 2000 calories per person
per day, or those consuming fewer calories than the calorie norms used to calculate the Indian
poverty lines). Columns 5 through 8 restrict attention to poorer households (those spending less
than either the median or the 25th percentile of either per-capita expenditure or per-capita food
expenditure in that survey round). Although the size of the caloric tax paid by migrants is slightly
smaller for these seven subpopulations, it still lies between 1 and 1.7 percent. The coefficients are
significantly negative in all cases except when considering the bottom quartiles of total expenditure or food expenditure. In these two cases, the sample size is dramatically smaller, and the
standard errors are much higher (rather than the magnitude of the coefficients being dramatically
lower).
Nutritional shortfalls at young ages have substantial scarring effects on productivity, cognitive
ability, earnings and health in adulthood (Almond and Currie, 2010). Thus, adequate nutrition is
particularly important for households with children. Columns 9 through 12 restrict attention to
families with children below the age of 5, and families with children below the age of 15. Since
childhood malnutrition is more likely among poorer households, I refine these two samples further by restricting attention only to households who have children and are spending less than the
25 The
first stage is very strong with a Cragg-Donald F-statistic of 36.8.
alternative approach would be to interact income (or other variables) with the migrant dummy. However, this
methodology has two shortcomings. First, it does not allow the coefficients on the controls to vary by income. Second,
the coefficient on the interaction term shows whether migrants have different elasticities of caloric consumption (rather
than whether migrants consume fewer calories for a given level of food expenditure even when poor).
26 An
15
median per-capita food expenditure in that survey round. In all four of these subpopulations, I
find that migrant households consume significantly fewer calories than non-migrant households
conditional on food expenditure, with magnitudes ranging between 1.3 and 1.7 percent. Given
that India’s child malnutrition rates were in excess of 50 percent around this time period (see
section 2), these subpopulations are very likely to be on the edge of malnutrition. In summary,
migrant households consume fewer calories than non-migrant households, conditional on food
expenditure, even when on the edge of malnutrition.
5
Why are migrant households consuming fewer calories than nonmigrant households?
The previous section documented that migrant households consume fewer calories than com-
parable non-migrant households, and that this result holds even for households on the edge of
malnutrition. In this section, I provide several pieces of evidence in support of an explanation for
this finding based on preference differences: migrants make nutritionally-suboptimal food choices
due to strong preferences for the favored foods of their origin states. First, I document that migrants bring their origin-state food preferences with them when they migrate. Second, I show that
the heterogeneity in the size of the migrant caloric tax is related to the suitability and intensity of
these origin-state food preferences: the caloric tax is only present when the average bundle of the
migrant’s origin state provides fewer calories than the local bundle (both priced at the local price
vector), and increases in size when both husband and wife are migrants (as opposed to just one of
these two being migrants). These pieces of evidence suggest that Indian migrants consume fewer
calories than non-migrants because they prefer to purchase the favored products from their origin
state even when these products are relatively expensive compared to local alternatives.
5.1
Migrants bring their food preferences with them
In this subsection, I present the first piece of evidence, that migrants bring their food prefer-
ences with them. In particular, I test two hypotheses: 1) compared to other households living
in the same village, a migrant’s consumption bundle more closely resembles the average bundle
of their origin state, and 2) compared to other households living in the same village, a migrant’s
consumption bundle less closely resembles the average bundle of their current state of residence.
In order to test these hypotheses, I require a measure of preference similarity. I propose an
intuitive and transparent measure based on correlations between household consumption bundles
and a reference consumption bundle for each state. For each of the 31 states s, I calculate the
average food-budget shares spent on each of 169 foods in each of the two survey rounds t. I
exclude the migrant populations of these states when calculating the average. These average foodbudget shares are then stacked in a vector bsharest which serves as the reference consumption
bundle for each state in each round.
16
With these reference bundles in hand, I compute the correlation between the vector of household is food-budget shares, denoted by bsharei , and the reference bundle of state s in the same
round t that household i was surveyed in: ρis = corr (bsharei , bsharest ). This exercise produces 31
household-state correlations for each household. I test whether migrant and non-migrant households posses the same preferences by comparing these correlations across households that face
the same price vector.27
At the time of enumeration, every household lives in a particular village or urban block, v,
contained within the state d (the “destination” state for migrant households). From the migration
records, I know the origin state o of every household (simply the current state for non-migrants).
Using these two pieces of information, I regress the ρis correlations described above on three indicator variables that identify the relationship between household i and the comparison state s:
the household currently lives in that state Id=s , the household does not currently live in that state
nor did they originally migrate from that state Id6=s,o6=s , and the household does not currently live
in that state but originally migrated from that state Id6=s,o=s . These three indicator variables are
interacted with a migranti dummy variable. I also introduce village by round by comparison-state
fixed effects, dvts , with the fixed effects absorbing the correlation of non-migrant bundles with
each comparison-state s. Finally, I include the same vector Xi of household-level controls used in
section 4:
ρis = β 1 migranti × Id=s + β 2 migranti × Id6=s,o6=s + β 3 migranti × Id6=s,o=s + dvts + Πt Xi + ε ist . (2)
The village by round by comparison-state fixed effects capture the average correlation between
food bundles in village v and state s. Therefore, the coefficients on the three interaction terms identify to what degree migrant bundles differ from the village average. Thus, this specification allows
me to test the two hypotheses formulated in the first paragraph of this section. First, compared to
other households in the same village facing the same prices, are food-budget shares of a migrant
from origin-state o more-closely correlated with the average food-bundle shares of origin-state o
(the hypothesis β 3 > 0)? Second, compared to other households in the same village facing the
same prices, are the food-budget shares of a migrant less-closely correlated with the average food
bundle of their current state of residence, d (the hypothesis β 1 < 0)?
As shown in column 1 of Table V, the data support both these sign predictions for the main
27 The
use of budget shares to calculate my preference similarity measure can be justified by assuming that houseθ
g
holds have Cobb-Douglas preferences over each food, u = Π169
g=1 c g where c g is the consumption of food g and θ g is
the taste parameter for food g and ∑169
g=1 θ g = 1. The preference similarity measure described above is the correlation between a household’s θ g taste parameters and the average taste parameters of non-migrant households in state
s. Atkin (Forthcoming) proposes a more complicated methodology to extract taste estimates from consumption data
using a modified Almost Ideal Demand System which can account for both non-homotheticities and responses of budget shares to local relative prices. Such a methodology is less feasible in this context since it is difficult to estimate
migrant-specific tastes when there are very few migrants from a particular origin state in a particular destination state.
Additionally, such a methodology is less necessary in this context where I explore within-village differences in preferences since all households in a village face the same set of prices. In the robustness analysis, I present a modification
that mitigates worries associated with non-homotheticities in consumption.
17
migrant sample. In both cases, I can reject the null hypotheses, that migrants do not differ from
non-migrants in the manner described above, at the 1 percent level. Reassuringly, the β 2 coefficient
is over five times smaller than the β 3 coefficient, and this difference is significant at the 1 percent
level (e.g. migrant bundles are more similar to their origin state, compared to non-migrants, than
they are to other comparison states, again compared to non-migrants).
In order to get a sense of magnitudes, the average correlation between a non-migrant’s bundle
and the reference bundle of the origin state of a migrant from their village is 0.6347. The estimated
β 3 coefficient implies that this correlation increases by 0.0213 for the migrant household. Similarly,
the average correlation between a non-migrant’s bundle and the bundle of their current state, is
0.7769 and the estimated β 1 corresponds to a decline of 0.0137 for the migrant household. Therefore, for non-migrant households, the correlation with their own state is 0.1422 higher than with
the migrant’s origin-state. For a migrant, the difference between the two correlations shrinks to
0.1072, closing about one quarter of this gap.
Columns 2 through 6 of Table V run the regression in equation 2 for the “wife moved for
marriage” subsamples and the alternative expenditure specifications detailed in section 4.1.28 In
addition to these robustness checks, my findings are robust to using three alternative similarity
measures. First, although budget shares have the appealing feature that they map directly into
parameters of the utility function if food preferences are of the Cobb-Douglas form, column 7
instead uses vectors of caloric shares to calculate the correlation coefficients. Second, migrant
households are likely to come from very different parts of the income distribution than the average
household in their origin state.29 Since migrants are never observed before their migration, any
correction is necessarily imperfect. Column 8 presents one such correction. Instead of using the
average bundle of non-migrant households as the state reference bundle, I calculate a reference
bundle that only uses non-migrant households in the same national income quartile as household
i (again in the round that the household was surveyed). Therefore, the regression compares the
correlation between a household’s bundle and the bundle consumed in state s by households at
similar income levels. Finally, the negative coefficient on migranti × Id=s is partly mechanical. As
the mean budget shares of state d were calculated using only the non-migrant households, the
correlation between household consumption bundles and the average state d bundle is likely to
be larger for non-migrant households. Although this bias is likely to be small (the average stateround sample contains 4,000 households), column 9 calculates the average bundle of state s using
all households. Results are remarkably similar across all the robustness specifications, with β 1
significantly negative and β 3 significantly positive at the one percent level in all specifications
bar the wife moved for marriage sample B subsample. Once again, the decline in the size of the
coefficient in the wife move samples is not surprising and is consistent with my caloric findings.
28 Additionally, Table IV reports these correlation results for the various subsamples of poor and undernourished
households detailed in section 4.2.
29 For example, the mean difference between the log per-capita expenditure of a migrant household and the log percapita expenditure of the average household in their origin state is a sizable 0.26. Although some of this difference
may be the expenditure gain due to migration, the migrant households would likely have been richer than the average
origin-state household and consumed a different bundle due to Engel effects.
18
These households typically only have one migrant member (the wife) who may have limited say
in household purchasing decisions, particularly when moving to their husbands birth village in
the B sample.
In summary, migrant households bring with them the idiosyncratic food preferences of their
origin state.
5.2
The size of the caloric tax depends on the suitability of the migrant’s preferences
This subsection links together my previous findings that migrants both bring their origin-state
tastes with them and consume fewer calories per Rupee than locals: I show that the size of the
caloric tax paid by migrants is related to how well-suited their origin-state preferences are to the
local price vector. In particular, I test the hypothesis that the size of the caloric tax is larger if migrants move to a village where their origin-state preferences place them at a caloric disadvantage
relative to locals.
In order to test this hypothesis, I use the information on the actual migration route traveled
and the preferences of households in the migrant’s origin state. As in the previous subsection,
I proxy the migrant’s origin-state preferences with their origin-state reference bundle, a vector
of average food-budget shares of non-migrants in their origin state. I then calculate the calories
derived from 1 Rupee allocated in the same proportions as this origin-state reference bundle but
with foods priced at the destination-village price vector. Similarly, I calculate the calories derived
from 1 Rupee allocated in the same proportions as the average bundle of non-migrant households
in the migrant’s destination village (also at destination-village prices). The log difference between
the calories derived from each of these 1 Rupee bundles measures the caloric advantage of a migrant’s origin-state preferences over the local preferences. Migrants who move to villages where
their origin-state bundle is a relatively expensive method of obtaining calories compared to the
local bundle have a negative value for this log difference. These migrant households have particularly disadvantageous preferences and should face a larger caloric tax compared to a migrant
household for whom this log difference is positive.
To implement this test, I rerun my migrant calorie regression, equation 1, except I now interact
the migrant dummy with an indicator variable that takes the value of 1 for negative values of log
difference described above: 1[ln K (bshareot , Pvt ) < ln K (bsharevt , Pvt )], where ln K (., .) are the log
calories derived from a 1 Rupee reference bundle (either of the migrant’s origin state, bshareot , or
the migrant’s destination village, bsharevt ) priced at the destination-village price vector Pvt in the
19
round t in which the household was surveyed:30
ln caloriesi = β 1 migranti + β 2 migranti × 1[ln K (bshareot , Pvt ) < ln K (bsharevt , Pvt )]
+ dvt + Πt Xi + ε i . (3)
The values of the ln K (., .) difference ranges from -0.62 for the 1st percentile of migrants to 0.92
for the 99th percentile, with negative values for 33 percent of migrant households. As before, I
include village-round fixed effects, dvt , and the same vector Xi of controls for expenditure and
household demographics described in section 4.
The hypothesis at the start of the subsection corresponds to the prediction β 2 < 0: the caloric
tax is more negative if a migrant’s origin-state bundle provides relatively few calories per Rupee
compared to the local bundle. Column 1 of Table VII presents the results of this regression. I can
reject the null hypothesis at the one percent level. The estimated β 2 coefficient is significantly negative and equal to -0.0283. The main effect, β 1 , is insignificant and close to zero. Migrants only pay
a caloric tax if they live in a village where purchasing their origin-state reference bundle provides
fewer calories than the local bundle. Summing the two coefficients, I find that migrant households living in villages where their preference are badly suited to the local price vector consume
3.33 percent fewer calories than comparable non-migrant households.
The categorization of migrant households into two groups based on the sign of the ln K (., .)
difference is a strong parametric assumption given that the origin-state bundle is an imperfect
measure of a migrant’s origin-state preferences. Column 2 of Table VII takes a more flexible stance
and divides the ln K (., .) difference into quartiles and allows the caloric tax to vary for migrants
in each quartile.31 The largest caloric tax of 3.93 percent is faced by migrant households in the
bottom quartile, those who moved to villages where their origin-state preferences provide the
fewest calories relative to the local preferences. The size of the caloric tax progressively declines
for households in the second and third quartiles before becoming significantly positive for the top
quartile.32 This top quartile of migrants who have the most advantageous origin-state preferences
receive a caloric dividend rather than pay a caloric tax. In terms of magnitudes, these migrant
households consume 2.27 percent more calories per Rupee spent compared to their non-migrant
neighbors.
30 I obtain the vector P
vτ by treating unit values (the expenditure on a food divided by the quantity purchased) as
price data and generate vectors of local prices from the median unit value paid for each food in each enumeration
village (or urban block). Unit values are not actual prices since quality varies. In part because of this concern, I use
median village prices as my price measure. These prices are robust to outliers and are less contaminated by quality
effects. Deaton (1988) details an alternative methodology to correct for quality differences. If none of the village sample
purchase a good, I use the median price at an incrementally higher level of aggregation.
31 The three quartile boundaries are -0.071, 0.118 and 0.322.
32 The finding that the caloric tax is negative even when the origin-state bundle provides more calories than the
local bundle is not surprising. The origin-state bundle is a only a perfect measure of migrant preferences if preferences
are unchanging, Cobb-Douglas in the manner described in footnote 27 and identical within each state. One departure
that would produce this finding is if richer households obtain fewer calories per Rupee due to income effects. In this
scenario, the true ln K (., .) difference would be more negative than the measured difference if migrant income rises
due to migration. Hence, migrants would pay a caloric tax even for small negative values of the measured ln K (., .)
difference.
20
I also directly interact the continuous measure of a migrant’s caloric advantage over locals. If
the origin-state reference bundle perfectly described the preferences of migrant households (for
example if the utility function took a Cobb-Douglas form, were fixed for life and were the same
for every person born in that state), the log difference between the caloric intake of migrants and
non-migrant households would be exactly equal to ln K (bshareot , Pvt ) − ln K (bsharevt , Pvt ). Under this scenario, the coefficient on the interaction of migranti with the ln K (., .) difference would
take a value of one. Column 3 of Table VII reports the coefficient from this regression. I find a
positive and significant coefficient of 0.0629: the size of the caloric tax increases with the caloric
disadvantage of the migrant’s origin-state preferences over the local preferences. The fact that the
coefficient is substantially smaller than one implies that either migrants adapt their preferences
after migrating, not all household members possess such preferences, or preferences are not of
the Cobb-Douglas/identical-within-state form (or some combination of these three explanations).
I will provide some explicit evidence for the first two of these explanations in later sections. If
preferences are not Cobb-Douglas, there may be partial substitution towards relatively inexpensive calorie sources mitigating the size of the caloric decline attributable to different preference
parameters.
As in previous sections, I perform a variety of robustness checks and present results for a number of subpopulations. Columns 4 through 12 of Table VII reproduce my findings using the wife
moved for marriage subsamples, the alternative expenditure controls, the caloric share or incomequartile adjusted reference baskets, instrumented food expenditure, and replacing caloric intake
with calories per Rupee.33 Table VIII contains the results for the subpopulations of undernourished households, poor households and households with children. In all nineteen of the robustness and subpopulation specifications, I find that the β 2 coefficient is significantly less than zero.
The magnitudes for migrants with disadvantageous preferences (e.g. for whom ln K (bshareot , Pvt ) <
ln K (bsharevt , Pvt )) range from a caloric tax of 1.87 percent for the subpopulation of households
below median income, to a caloric tax of 3.46 percent for the subpopulation of households consuming fewer than 2000 calories per person per day. I find significant caloric taxes for migrants
with disadvantageous preferences even for households in the bottom quartiles of total expenditure or food expenditure (a tax of 2.66 percent and 2.08 percent respectively).34 Even on the edge of
malnutrition, migrants consume substantially fewer calories per Rupee compared to locals when
their preferences are out-of-sync with the local price vector.
I perform two additional robustness tests that address particular concerns with this exercise.
One reasonable concern is that the ln K (., .) difference is related to the distance migrants have trav33 The
caloric share specifications use the log difference of the cost to purchase a 1000 calorie bundle using the
average caloric shares of non-migrant households in the migrant’s origin state priced at the destination-village price
vector relative to the cost to purchase a 1000 calorie bundle using the average caloric shares of non-migrant households
in the migrant’s destination village priced at the destination-village price vector. Therefore, the sign predictions on β 1
and β 2 are reversed (the caloric tax on migrants is less negative when the origin-state bundle providing 1000 calories is
relatively cheap compared to the destination-village bundle providing 1000 calories).
34 The sum of the β and β coefficients is significantly different from zero at the 5 level for the poorest quartile of
2
1
total expenditure and at the 10 percent level for the poorest quartile of food expenditure.
21
eled and migrants from far-off places differ in some unobservable way to other types of migrants.
I address this concern head on by including an additional interaction between migranti and the log
distance between the migrant’s destination region (a subset of their state) and their origin state.
These results are shown in column 13 of Table VII. The coefficient on the distance interaction is
insignificant while the β 2 coefficient is essentially unchanged. A second reasonable concern is that
the negative β 2 coefficient is driven by outlier households at the village level. For example, if there
is a non-migrant household that consumes an unusually small number of calories per Rupee, the
village-level ln K (bsharevt , Pvt ) will be low and the average difference between migrant and nonmigrant calories per Rupee will be high. Accordingly, columns 14 and 15 calculate the ln K (., .)
difference between origin-state and destination bundles utilizing the average bundles at either
the district level (bsharedt ) or the state level (bsharest ) instead of at the village level (pricing the
bundles in both cases at destination-village prices). The β 2 coefficients are negative and significant
in both these specifications.
In summary, the size of the caloric tax paid by migrant households is systematically related to
how well-suited their origin-state preferences are to the local price vector.
5.3
The size of the caloric tax depends on the intensity of the migrant’s preferences
In this subsection, I provide further support for a preference-difference explanation by show-
ing that the caloric tax paid by migrants is related to the intensity of their preferences. In particular, I test the hypothesis that the size of the caloric tax paid by migrants depends on how many
household members are migrants. The prior is that any impacts associated with being a migrant
household should be more exaggerated if both husband and wife are migrants compared to if
only one of the two is a migrant (since in the former scenario both of the primary decision makers
posses non-local preferences).
Panel 1 of Table IX explores heterogeneity across these different within-household migrant
structures. I interact the migranti dummy in equation 1 with an exhaustive set of three dummies for the the migrant status of the household head and their spouse: (1) there is no spouse
(nospousei ), (2) only one of either the head or the spouse is a migrant (onlyonei ), and (3) both the
head and the spouse are migrants (bothi ):
ln caloriesi = β 1 migranti × nospousei + β 2 migranti × onlyonei + β 3 migranti × bothi + dvt + Πt Xi + ε i .
(4)
The β coefficients from this specification provide separate estimates for the caloric tax faced by migrants for each of these three structures. The hypothesis at the start of the subsection corresponds
to the prediction that β 3 > β 2 : the caloric tax is larger if both husband and wife are migrants
rather than only one of the two. I find support for this hypothesis. When only only one of the
husband and wife is a migrant, I obtain a coefficient on the migrant dummy of -0.0130. In contrast
the size of the caloric tax is significantly more negative when both husband and wife are migrants
22
(a coefficient of -0.0234).35
Panel 2 of Table IX performs a similar breakdown for equation 3 of the previous subsection.36
For all three different migrant structures, I find that the size of the migrant tax is larger when
the migrant’s origin-state reference bundle provides fewer calories than the local bundle. The
ordering of the size of the tax is also consistent with the hypotheses above.37 The most adversely
affected households (households in which both husband and wife migrated to a village where
the origin-state reference bundle provides fewer calories than the local bundle) face a caloric tax
of 7.2 percent. The magnitude of this caloric cost is substantial. The median caloric intake for
this migrant subgroup is 2156 calories per person per day with 58.4 percent of rural households
and 53.9 percent of urban households consuming less than the recommended calorie norms. If
these migrants had the same preferences as locals, the median would rise to 2318 calories and the
percentage below the caloric norms would fall to 48.3 percent of rural households and 41.9 percent
of urban households.
Finally, Panel 3 of Table IX breaks out the correlation result of equation 2 in section 5.1. I
find strikingly similar results to the caloric intake analysis. Migrant consumption bundles most
resemble their origin-state reference bundle when when both husband and wife are migrants, with
smaller effects when only one of the head and spouse are migrants.38
In summary, the size of the caloric tax paid by migrants depends on how many household
members possess preferences that differ from those of the local population.
I can also evaluate an additional hypothesis in this manner. In traditional societies such as
India, men typically have greater bargaining power in household decision making. Additionally,
as discussed in section 3, patrilocality is the norm in most of India and so wives often movein with the extended family of the husband. Therefore, under a preference-difference story, the
caloric impacts associated with having a migrant in the household might be more exaggerated if
the husband is a migrant as opposed to the wife.
Panels 4 to 6 of Table IX present similar specifications to the Panels 1 to 3 but break the “only
one” category into two categories: only the spouse is a migrant, and only the head is a migrant.
Across each of the three regressions, the ordering of the effects size across these two categories
is in accordance with the prior above that the impacts are more exaggerated if the husband is a
migrant as opposed to the wife. For example, in Panel 4, the size of the caloric tax if only the
35 I can reject the null that the coefficients on
migranti × onlyonei and migranti × bothi are equal with a p-value of 5.6
percent.
36 If both head and spouse are migrants but come from different origin states, I take the average value of
ln K (bshareoτ , Pvτ ) across the two origin states. Since there are very few such households, results are essentially unchanged if these households are dropped.
37 Once more, I can reject at the 1 percent level that the null that the size of the caloric taxes for migrants with unsuited
preferences are identical if only one of the head and spouse are migrants, or if both are migrants.
38 If both head and spouse are migrants but come from different origin states, I replace the I
d6=s,o =s indicator variable
with the value of one half for each of the two origin states (and similarly for Id6=s,o6=s ). Since there are very few such
households, results are essentially unchanged if these households are dropped. As before, I can reject the null that the
migranti × Id6=s,o=s interactions are identical if only one of the head and spouse are migrants, or if both are migrants
(at the one percent level). Similarly, I can also reject that the migranti × Id=s interactions are identical across these two
groups at the one percent level.
23
spouse is a migrant is around 1.1 percent and rises to 1.9 percent when only the husband is a
migrant. However, since both the coefficient differences and sample sizes are smaller for this
hypothesis test, in none of the three regressions are these differences statistically significant at the
10 percent level.
6
Competing explanations
This paper establishes three facts. First, I find that on average migrant households are paying a
caloric tax as they consume fewer calories than non-migrants for a given level of food expenditure,
and this holds even for households on the edge of malnutrition. Second, I find that compared to
non-migrants, migrant households posses food preferences that more closely resemble the food
preferences of households in their origin-state. Third, I find that the size of this caloric tax is
systematically related to how well suited a migrant’s origin-state preferences are to the price vector
of their destination-village and how intense these preferences are. This chain of results strongly
suggests that migrants in India are making nutritionally-suboptimal food choices due to strong
preferences for the favored foods of their origin states. Prior to discussing the implications of
these findings, I turn to dismissing two competing explanations.
6.1
An information story
The first competing explanation is that migrants have poor information about the local al-
ternatives to their origin-state foods. Similarly, compared to locals, migrants may have a worse
knowledge of local prices and the caloric content of local foods. Under this scenario, migrants
would consume fewer calories per Rupee than non-migrants as they are not aware of cheaper alternatives. Migrants would also continue to consume bundles that more closely resembled their
origin-state reference bundle since they are more familiar with the foods that are common in their
origin state. In this subsection, I provide four pieces of evidence that contradict this informationbased explanation for my findings.
First, I document that the caloric tax is persistent and remains many years after migration.
Even if migrants are initially uninformed, after many years in the destination they would become
familiar with the local foods and prices. I exploit the data on time since migration and rerun my
three main regression specifications on subpopulations that exclude recent migrants. Specifically, I
exclude migrant households where the most recent migrant arrived less than 5, 10 or 20 years prior
to the survey. Columns 13 to 15 of Table IV presents these three subpopulation results for the first
empirical specification, equation 1. The caloric tax remains significantly negative for the first two
long-term migrant specifications, although the size of the tax declines. When I exclude migrants
who arrived less than 20 years previously, the tax disappears altogether, with the coefficient on
the migrant dummy both positive and insignificant. However, Tables VI and VIII tell a more
complete story. Long-term migrants still consume bundles more closely related to their originstate bundle than locals do (columns 13 to 15 of Table VI). Furthermore, when long-term migrants
24
live in locations where their origin-state bundle provides fewer calories than the local bundle,
there is a significant caloric tax ranging from 2.74 percent when I exclude migrants who moved less
than 5 years ago to 1.79 percent when I exclude migrants who arrived less than 20 years previously
(columns 13 to 15 of Table VIII). This caloric tax for migrants who arrived 20 or more years ago is
only negated in the aggregate because the long-term migrants who live in locations where their
origin-state bundle provides more calories than the local bundle enjoy a caloric dividend of 1.49
percent. Therefore, even migrants who have had many years to learn about local foods and prices
pay a caloric tax when their origin-state preferences are unsuited to the local price vector.
Second, if the explanation is that migrants have poor information, the migrant tax is likely to
be smaller among literate segments of the population who can acquire information more easily. I
find the opposite relationship in the data. Column 16 of Tables IV, VI and VIII restrict attention
to households in which the household head is literate. The size of the caloric tax actually grows
larger when I focus on this subpopulation (rising to 1.71 percent compared to 1.59 percent in the
basic specification).
Third, I find evidence of a caloric tax on migrants when only one of the husband or wife are
migrants (Panel 1 of Table IX), or even when wives are moving to their husband’s original birth
village (the wife moved for marriage sample B, column 3 of Table III). In these cases, other household
members already posses information about local foods and prices yet the caloric tax remains.
Finally, I present evidence that migrants do reduce their purchases of origin-state foods when
their origin-state preferences are particularly unsuited to the local price vector. This moderation
suggests that migrants are aware of the trade-offs between indulging in their preferences and their
caloric intake. I return to the bundle similarity regressions, equation 2 in section 5.1, but further interact the migrant dummy variables with the indicator for a negative value of [ln K (bshareot , Pvt ) −
ln K (bsharevt , Pvt )] that I used in section 5.2. This regression is reported in column 10 of table V.
The positive coefficient on migranti × Id6=s,o=s remains; migrant households still consume bundles
that more closely resemble their origin-state bundle compared to non-migrant households in their
village. However, this effect is tempered if the migrant’s origin-state bundle provides relatively
few calories compared to the local bundle. (E.g. I find a significantly negative coefficient on the
triple interaction migranti × Id6=s,o=s × 1[ln K (bshareot , Pvt ) < ln K (bsharevt , Pvt )] but the sum of
this coefficient and the coefficient on migranti × Id6=s,o=s is still positive and significantly different
from zero at the 1 percent level.)
Therefore, migrants seem to be aware that their origin-state preferences can have negative consequences for nutrition since they moderate their consumption choices in contexts where these
preferences are most disadvantageous. However, the adaptation is incomplete. Even in contexts
where their origin-state preferences are calorically disadvantageous, migrant households still consume bundles that more closely resemble the bundles consumed in their origin state, consistent
with my previous finding that these migrant households consume fewer calories than locals.
25
6.2
A technology story
A second competing explanation is that migrants do not posses the technologies to make de-
licious meals using the locally-cheap foods. These technologies encompass cooking and foodpreparation equipment as well as recipes and preparation techniques that allow households to
turn raw foods into enjoyable meals. For example, a family in Rajasthan may be expert at transforming wheat into delicious roti (a flat bread), but may lack the training or equipment to make a
decent jhal-muri (a rice ball popular in West Bengal). If the family migrated to West Bengal, they
may continue to consume wheat as they enjoy well prepared meals over badly prepared meals
rather than wheat over rice.
Once more the evidence from various subpopulations in the data contradicts a story in which
technology is the sole explanation for my findings. If technology was responsible, the caloric
tax should disappear for long-term migrants. These households have spent many years away,
providing a sufficient time frame over which to purchase new equipment as well as learn new
recipes and techniques. Similarly, there should be no tax for migrants who are moving into an
existing household where appropriate kitchen equipment is already present and wives can learn
recipes and preparation techniques from other family members such as mother-in-laws. As shown
in the previous subsection, the caloric tax persists for both these subpopulations.
Finally, women are typically in charge of food preparation in Indian households. Therefore,
if a lack of recipes and preparation techniques were the cause of the caloric tax, the tax should
be smaller when only the husband is a migrant compared to when only the wife is a migrant. As
shown in Panel 4 of Table IX, I find no evidence in support of this conjecture. In fact, the caloric tax
is larger when only the husband is a migrant (1.93 percent compared to 1.08 percent), a finding I
previously attributed to migrants bringing their origin-state preferences with them and husbands
having greater bargaining power in household decision making.
7
Conclusion and policy implications
This paper sets out to answer two simple questions: i) to what extent does culture influence
caloric intake?; and ii) does culture constrain caloric intake even for households on the edge of
malnutrition? I quantify the caloric costs of culture by exploiting the fact that migrants and nonmigrants face the same relative prices yet posses very different preferences. Drawing on detailed
household survey data from India, I find that inter-state migrants consume fewer calories per Rupee of food expenditure compared to their non-migrant neighbors. This “caloric tax” on migrants
corresponds to 1.6 percent of caloric intake, and is evident even for households on the edge of malnutrition. I then provide two pieces of evidence in support of an explanation for this finding based
on culture: that migrants make nutritionally-suboptimal food choices due to strong preferences
for the favored foods of their origin states. First, I document that migrants bring their originstate food preferences with them when they migrate. Second, I show that the heterogeneity in the
size of the migrant caloric tax is related to the suitability and intensity of these origin-state food
26
preferences. The most adversely affected migrants (households in which both husband and wife
migrated to a village where their origin-state preferences are unsuited to the local price vector)
would consume 7 percent more calories if they possessed the same preferences as their neighbors.
These results have important implications for tackling hunger and malnutrition. This paper
establishes that even households on the edge of malnutrition place considerable value in accommodating their cultural food preferences. These cultural causes of hunger need to be understood
when designing programs to alleviate malnutrition. Three type of program are particularly relevant: programs that provide food aid or price subsidies to consumers; programs that reduce tariffs
or use other trade policies to increase food imports; and programs that aim to develop bio-fortified
or high-yield crop varieties. In all three cases, programs will be more effective if the targeted foods
are those which are favored by households on the edge of malnutrition.
As a concrete example, white maize is greatly preferred to yellow maize in much of Africa.39
However, much food aid to Africa comes in the form of imported yellow maize, and yellow maize
is also more nutritious due to its higher levels of vitamin A. Programs that provide cheap yellow
corn to hungry communities, or try to reduce vitamin-A deficiency through wider availability of
yellow maize are less effective in contexts where there are cultural preferences for white maize.
Food vouchers that allow consumers to choose their favored foods or bio-fortification of traditional foods may prove more successful in these situations.40 Similarly, the introduction of highyield varieties (HYV) of rice, wheat and yellow maize spurred “the green revolution” in much
of the developing world. However, this revolution bypassed sub-Saharan Africa. Alongside a
range of other factors, adoption of these three HYV crops was held back by local preferences for
sub-Saharan staples such as sweet potato, cassava, sorghum, teff and white maize.41
39 See (McCann, 2005) for the historical origins of this preference for white maize. Muzhingi et al. (2008) and
De Groote and Kimenju (2008) provide empirical evidence for this preference ordering.
40 In the case of white maize, vitamin-A bio-fortification currently involves the addition of carotenes which turn the
maize yellow-orange, leading to reduced uptake since consumers associate it with yellow maize (Nestel et al., April
2006).
41 See Paarlberg (2010) for a more complete discussion of the reasons for the failure of Africa’s green revolution.
27
References
Algan, Yann, and Pierre Cahuc. 2010. “Inherited Trust and Growth.” American Economic Review,
100(5): 2060–92.
Almond, Douglas, and Janet Currie. 2010. “Human Capital Development Before Age Five.”
NBER Working Papers 15827.
Atkin, David. Forthcoming. “Trade, Tastes and Nutrition in India.” The American Economic Review.
Barrett, Christopher B., and Daniel G. Maxwell. 2005. Food Aid After Fifty Years: Recasting Its Role.
London:Routledge.
Birch, Leann L. 1999. “Development of Food Preferences.” Annual Review of Nutrition, 19(1): 41–62.
Bronnenberg, Bart J., Jean-Pierre H. Dube, and Mathew Gentzkow. 2012. “The Evolution
of Brand Preferences: Evidence from Consumer Migration.” The American Economic Review,
102(6): 2472–2508.
Deaton, Angus. 1988. “Quality, Quantity, and Spatial Variation of Price.” The American Economic
Review, 78(3): 418–430.
Deaton, Angus, and Jean Drèze. 2009. “Nutrition in India: Facts and Interpretations.” Economic
& Political Weekly, 44(7): 43.
De Groote, Hugo, and Simon Chege Kimenju. 2008. “Comparing consumer preferences for color
and nutritional quality in maize: Application of a semi-double-bound logistic model on urban
consumers in Kenya.” Food Policy, 33(4): 362–370.
Fernandez, Raquel. 2010. “Does Culture Matter?” National Bureau of Economic Research NBER
Working Papers 16277.
Fernandez, Raquel, Alessandra Fogli, and Claudia Olivetti. 2004. “Mothers and Sons: Preference
Formation and Female Labor Force Dynamics.” The Quarterly Journal of Economics, 119(4): 1249–
1299.
Fernandez, Raquel, and Alessandra Fogli. 2009. “Culture: An Empirical Investigation of Beliefs,
Work, and Fertility.” American Economic Journal: Macroeconomics, 1(1): 146–77.
Fogel, Robert W. 1994. “Economic Growth, Population Theory, and Physiology: The Bearing
of Long-Term Processes on the Making of Economic Policy.” The American Economic Review,
84(3): pp. 369–395.
Giuliano, Paola. 2007. “Living arrangements in western europe: Does cultural origin matter?”
Journal of the European Economic Association, 5(5): 927–952.
28
Guiso, Luigi, Paola Sapienza, and Luigi Zingales. 2004. “The Role of Social Capital in Financial
Development.” The American Economic Review, 94(3): 526–556.
Guiso, Luigi, Paola Sapienza, and Luigi Zingales. 2006. “Does Culture Affect Economic Outcomes?” Journal of Economic Perspectives, 20(2): 23–48.
Harris, Marvin. 1985. Good to Eat: Riddles of Food and Culture. Simon and Schuster.
Kelder, Steven H. 1994. “Longitudinal Tracking of Adolescent Smoking, Physical Activity, and
Food Choice Behaviors.” American Journal of Public Health, 84(7): 1121–1126.
Logan, Trevon, and Paul Rhode. 2010. “Moveable Feasts: A New Approach to Endogenizing
Tastes.” Working Paper, Ohio State University.
McCann, James. 2005. Maize and Grace: Africa’s Encounter with a New World crop, 1500-2000. Harvard University Press.
Mennella, Julie, Coren Jagnow, and Gary Beauchamp. 2001. “Prenatal and Postnatal Flavor
Learning by Human Infants.” Pediatrics, 107(6).
Muzhingi, Tawanda, Augustine S. Langyintuo, Lucie C. Malaba, and Marianne Banziger. 2008.
“Consumer acceptability of yellow maize products in Zimbabwe.” Food Policy, 33(4): 352 – 361.
Nestel, Penelope, Howarth E. Bouis, J. V. Meenakshi, and Wolfgang Pfeiffer. April 2006. “Biofortification of Staple Food Crops.” The Journal of Nutrition, 136(4): 1064–1067.
Nunn, Nathan, and Nancy Qian. 2011. “The Impact of Potatoes on Old World Population and
Urbanization.” The Quarterly Journal of Economics.
Paarlberg, Robert. 2010. Food Politics: What Everyone Needs to Know. Oxford University Press, USA.
Srinivas, M.N. 1980. India: Social Structure. Delhi:Transaction Publishers.
Staehle, Hans. 1934. “The Reaction of Consumers to Changes in Prices and Income: A Quantitative Study in Immigrants’ Behavior.” Econometrica, 2(1): 59–72.
Subramanian, Shankar, and Angus Deaton. 1996. “The Demand for Food and Calories.” Journal
of Political Economy, 104(1): 133.
UN Millenium Project Task Force on Hunger. 2005. Halving Hunger: it can be done. Earthscan.
Woodham-Smith, Cecil. 1964. The Great Hunger: Ireland 1845-1849. Signet.
29
.0004
0
.0002
Density
.0006
.0008
Figure I: Caloric intake under actual and hypothetical preferences for rice and wheat
1000
Urban Rural
RDA RDA
3000
4000
5000
Caloric intake per person per day (1983 and 1987/88)
Urban Areas (Actual)
Rural Areas (Hypothetical)
Urban Areas (Hypothetical)
0
Percent of Switchers
5
10
15
Rural Areas (Actual)
−.6
−.4
−.2
Difference in log caloric intake between real and hypothetical preferences
0
Note: The upper panel displays the counterfactual distribution of caloric intake if households switched the quantities
purchased of rice and wheat where advantageous, spending any cost savings on the cheaper of the two foods. The
distribution includes all households in the 22,148 villages where I observe both rice and wheat purchases in the two
survey rounds. The lower panel displays a histogram of the caloric gains available to the switchers in the counterfactual
exercise.
30
2
1
0
−1
−2
−3
−10
−5
0
5
10
−1
0
1
2
Log difference between rice and wheat consumption
(village−level caloric intake)
−2
Log difference between rice and wheat prices
(village−level median price per calorie)
Log difference between rice and wheat prices
(village−level median price per calorie)
Figure II: Regional consumption patterns for rice and wheat
−5
0
5
Log difference between rice and wheat consumption
(village−level caloric intake)
Rajasthan
Gujarat
Bihar
Andhra Pradesh
Note: The upper panel plots the relative price of rice and wheat calories against the relative consumption of rice and
wheat calories for the 22,148 villages where I observe both rice and wheat purchases in the two survey rounds. The
lower panel highlights the observations for four particular states.
31
32
0.112
0.000
0.000
Only spouse is
migrant
0.413
0.807
1.000
1.000
0.487
0.452
Proportion of full
sample
Consuming below
2000 calories per
capita per day
0.433
0.436
0.419
Consuming below
2400/2100 calories
(rural/urban)
0.566
0.591
0.591
10 to 20 years
0.293
0.320
0.312
0.132
0.000
0.000
Only head is migrant
0.778
0.847
0.876
Both migrants
(different state)
0.019
0.020
0.000
Proportion of
migrants - rural
households
0.459
0.576
0.703
Below median per Below 25th percentile
capita expenditure
of per capita
expenditure
0.307
0.126
0.369
0.164
0.432
0.201
20 or more years
0.413
0.470
0.503
Both migrants
(same state)
0.324
0.173
0.000
0.061
0.061
0.049
Proportion of sample Proportion of sample
-rural
- migrant households
Note: Table shows proportion of the sample households in various categories. All proportions use the household weights provided by the NSS.
Migrant breakdown by per capita expenditure and nutrition
Consuming below
1850 calories per
capita per day
Full Sample
0.338
Wife moved for marriage sample A
0.338
Wife moved for marriage sample B
0.323
Migrant breakdown by years since migration (most recent migrant in household)
0 to 4 years
5 to 9 years
Full Sample
0.152
0.141
Wife moved for marriage sample A
0.072
0.137
Wife moved for marriage sample B
0.060
0.124
Full Sample
Wife moved for marriage sample A
Wife moved for marriage sample B
Migrant breakdown by within-household migrant structure
No spouse
Full sample
Wife moved for marriage sample A
Wife moved for marriage sample B
Number of
households
(unweighted)
240,081
93,823
84,628
Table I: Sample descriptive statistics
Table II: Differences between migrant and non-migrant households
Log caloric intake per person per day
Log household per capita expenditure
(Rupees, 30 days)
Log food expenditure
(Rupees, 30 days)
Log household size
Proportion males 0-4
Proportion females 0-4
Proportion males 5-9
Proportion females 5-9
Proportion males 10-14
Proportion females 10-14
Proportion males 15-55
Proportion females 15-55
Proportion males over 55
Proportion females over 55
Rural self-employed in non-agriculture
Urban self-employed
Rural agricultural labor
Urban wage earner
Rural other labor
Urban casual labor
Rural self-employed in agriculture
Rural other
Urban other
Log price paid (no controls)
(Rupees per 1000 calories)
Log price paid (controls)
(Rupees per 1000 calories)
(1)
Mean (full sample)
(2)
Migrant difference
(full sample)
7.6286
0.0133***
(3)
(4)
Migrant difference Migrant difference
(wife moved for
(wife moved for
marriage sample A) marriage sample B)
0.0034
-0.0007
(0.3712)
(0.0049)
(0.0080)
(0.0089)
4.8272
0.0624***
0.0423***
0.0332***
(0.5816)
(0.0069)
(0.0108)
(0.0121)
4.4205
0.0460***
0.0272***
0.0183*
(0.5045)
(0.0059)
(0.0094)
(0.0105)
1.7522
-0.0314***
0.0001
0.0066
(0.4848)
(0.0068)
(0.0104)
(0.0119)
0.0677
0.0025
0.0034
0.0033
(0.1043)
(0.0015)
(0.0026)
(0.0029)
0.0634
0.0021
0.0044*
0.0044*
(0.1020)
(0.0014)
(0.0024)
(0.0027)
0.0722
-0.0007
-0.0023
-0.0034
(0.1055)
(0.0014)
(0.0024)
(0.0027)
0.0656
-0.0001
-0.0002
0.0004
(0.1004)
(0.0015)
(0.0025)
(0.0028)
0.0666
-0.0007
-0.0030
-0.0017
(0.1052)
(0.0014)
(0.0024)
(0.0028)
0.0575
-0.0006
-0.0027
-0.0019
(0.0962)
(0.0013)
(0.0022)
(0.0024)
0.2701
0.0113***
0.0023
0.0001
(0.1596)
(0.0023)
(0.0030)
(0.0033)
0.2625
-0.0089***
-0.0016
-0.0023
(0.1339)
(0.0017)
(0.0026)
(0.0029)
0.0375
-0.0013
-0.0007
0.0001
(0.0875)
(0.0011)
(0.0017)
(0.0020)
0.0370
-0.0036***
0.0005
0.0010
(0.0917)
(0.0010)
(0.0016)
(0.0018)
0.0972
0.0036
0.0041
0.0016
(0.2963)
(0.0038)
(0.0068)
(0.0080)
0.0853
0.0212***
0.0158***
0.0082
(0.2793)
(0.0043)
(0.0060)
(0.0059)
0.2151
-0.0110**
-0.0160*
-0.0142
(0.4109)
(0.0044)
(0.0085)
(0.0100)
0.0513
-0.0054*
-0.0022
0.0012
(0.2207)
(0.0031)
(0.0042)
(0.0036)
0.0553
0.0069**
-0.0006
-0.0009
(0.2286)
(0.0028)
(0.0046)
(0.0054)
0.0143
-0.0037**
-0.0010
0.0002
(0.1188)
(0.0018)
(0.0023)
(0.0021)
0.3464
-0.0075
-0.0016
0.0004
(0.4758)
(0.0049)
(0.0097)
(0.0115)
0.0643
0.0080**
0.0141***
0.0131**
(0.2453)
(0.0031)
(0.0049)
(0.0055)
0.0707
-0.0120***
-0.0126***
-0.0096
(0.2564)
(0.0033)
(0.0046)
(0.0049)
0.645825
0.00341***
-0.00033
-0.0015
(1.0874)
(0.00115)
(0.00204)
(0.0023)
0.6447275
0.00102
-0.00184
-0.0026
(1.0874)
(0.00112)
(0.00202)
(0.0023)
Note: Column 1 shows the mean of each household-level variable in the row title. Column 2 shows the coefficient on a migrant
dummy when the variable is regressed on a village-round fixed effect and a migrant-status dummy. Columns 3 and 4 show the
coefficient for the same regression but restricting attention to households in the two “wife moved for marriage samples” described in the
text. The last two rows show the coefficient on a migrant dummy from a regression of log unit values for every product purchased
on product-village-round fixed effects and a migrant-status dummy. The “controls” row includes the same vector of controls for log
per-capita food expenditure and household characteristics used in all later regressions. Regressions are weighted using household
weights except the last two rows that use budget shares interacted with household weights. Regressions clustered at the level of the
fixed effects. * signifies significance at the 10 percent level, ** at the 5 percent level and *** at the 1 percent level.
33
34
Wife Moved For
Marriage Sample A
-0.0121***
(0.00422)
93,113
0.836
Yes
No
No
No
Yes
Yes
Basic — food
expenditure controls
-0.0159***
(0.00279)
235,203
0.826
Yes
No
No
No
Yes
Yes
(4)
(5)
(6)
Log caloric intake (per person per day)
Wife Moved For
Total expenditure Real food expenditure State-specific food
controls
expenditure controls
Marriage Sample B
controls
-0.0109**
-0.0136***
-0.0117***
-0.0167***
(0.00475)
(0.00360)
(0.00288)
(0.00283)
84,099
235,180
235,199
235,203
0.837
0.688
0.790
0.818
Yes
No
No
No
No
Yes
No
No
No
No
Yes
No
No
No
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
(3)
Food expenditure
instrumented
-0.0118***
(0.00269)
234,961
0.711
Yes
No
No
No
Yes
Yes
(7)
(8)
Log caloric intake per
Rupee of food
expenditure
-0.0145***
(0.00286)
235,180
0.815
No
Yes
No
No
Yes
Yes
** at the 5 percent level and *** at the 1 percent level.
All regressions are weighted using household weights and the standard errors are clustered at the village-round level. * signifies significance at the 10 percent level,
column 8 which use alternate expenditure controls). Column 7 instruments the polynomial in total food expenditure with a polynomial in total non-food expenditure.
round-specific controls for household size, demographics and type as well as a third-order polynomial in log per-capita food expenditure (bar columns 4 to 6 and
wives who moved interstate (migrants) to those who moved intrastate (non-migrants). All specifications include village-round fixed effects and flexible survey-
to only spouses). Columns 2 and 3 restrict attention to households in which the wife of the household head moved village at the time of marriage and compares
expenditure. Independent variable migranti is a dummy for whether the household head or their spouse is an interstate migrant (columns 2 and 3 restrict attention
Note: Dependent variable for columns 1-7 is log caloric intake per person per day. Dependent variable for column 8 is log caloric intake per Rupee of food
Observations
R-squared
Food Expenditure Controls
Total Expenditure Controls
Real Food Expenditure Controls
State-Specific Food Expenditure Controls
Demographics/Household Type Controls
Village-Round FE
migrant i
(2)
(1)
Table III: Comparing the Caloric Intake of Migrants and Non-Migrants
35
-0.0132***
(0.00457)
105,042
0.839
Yes
Yes
Yes
(11)
Consuming < 2000
calories per day
-0.0145***
(0.00468)
92,911
0.770
Yes
Yes
Yes
Children under 5 and Children under 16 in
below median food
household
expend.
-0.0141*
-0.0169***
(0.00808)
(0.00319)
52,650
178,628
0.840
0.826
Yes
Yes
Yes
Yes
Yes
Yes
(10)
(9)
Children under 5 in
household
Consuming < 1850
calories per day
-0.0120**
(0.00610)
71,376
0.785
Yes
Yes
Yes
Basic — food
expenditure controls
-0.0159***
(0.00279)
235,203
0.826
Yes
Yes
Yes
(3)
(12)
(13)
Log caloric intake (per person per day)
Children under 16
Migrated 5 or more
and below median
years ago
food expend.
-0.0140**
-0.0114***
(0.00584)
(0.00298)
79,162
231,359
0.824
0.827
Yes
Yes
Yes
Yes
Yes
Yes
(4)
(5)
Log caloric intake (per person per day)
Consuming < 2400
Below median per
(rural) 2100 (urban)
capita expenditure
-0.0158***
-0.0107*
(0.00376)
(0.00550)
132,472
90,098
0.772
0.830
Yes
Yes
Yes
Yes
Yes
Yes
(15)
Below median per
capita food expend.
-0.0147***
(0.00553)
89,745
0.820
Yes
Yes
Yes
(7)
-0.00650**
(0.00323)
228,538
0.827
Yes
Yes
Yes
0.00183
(0.00412)
223,615
0.828
Yes
Yes
Yes
Migrated 10 or more Migrated 20 or more
years ago
years ago
(14)
Bottom quartile per
capita expenditure
-0.0141
(0.0105)
41,014
0.869
Yes
Yes
Yes
(6)
-0.0171***
(0.00352)
130,175
0.844
Yes
Yes
Yes
Literate
(16)
Bottom quartile per
capita food expend.
-0.00981
(0.0105)
41,592
0.861
Yes
Yes
Yes
(8)
at the 5 percent level and *** at the 1 percent level.
regressions are weighted using household weights and the standard errors are clustered at the village-round level. * signifies significance at the 10 percent level, **
well as a third-order polynomial in log per-capita food expenditure. The column headings denote the various subpopulatons on which the regressions are run. All
an interstate migrant. All specifications include village-round fixed effects and flexible survey-round-specific controls for household size, demographics and type as
Note: Dependent variable is log caloric intake per person per day. Independent variable migranti is a dummy for whether the household head or their spouse is
Observations
R-squared
Food Expenditure Controls
Demographics/Household Type Controls
Village-Round FE
migrant i
Observations
R-squared
Food Expenditure Controls
Demographics/Household Type Controls
Village-Round FE
migrant i
(2)
(1)
Table IV: Comparing the Caloric Intake of Migrants and Non-Migrants: Subpopulations
36
7,291,293
0.835
No
Yes
Yes
Yes
7,291,293
0.860
Yes
No
Yes
Yes
7,180,153
0.822
Yes
No
Yes
Yes
7,291,293
0.836
Yes
No
Yes
Yes
(0.00450)
7,261,099
0.838
Yes
No
Yes
Yes
-0.0166***
(0.00458)
(6)
(7)
(8)
(9)
(10)
Correlation ρ is of the budget shares of household i with the average budget shares of state s (169 foods)
State-specific food
Caloric share
Reference basket
Reference basket
Interactions with lnK
expenditure controls
correlations
by income quartile
includes migrants
difference
-0.0149***
-0.0114***
-0.0124***
-0.0132***
-0.0134***
(0.00225)
(0.00261)
(0.00228)
(0.00221)
(0.00256)
0.00282***
0.00796***
0.00524***
0.00399***
0.00406***
(0.000395)
(0.000446)
(0.000417)
(0.000391)
(0.000393)
0.0202***
0.0264***
0.0215***
0.0211***
0.0265***
(0.00210)
(0.00249)
(0.00208)
(0.00206)
(0.00248)
-0.000849
Real food expenditure
controls
-0.0147***
(0.00217)
0.00294***
(0.000383)
0.0203***
(0.00203)
7,291,169
0.842
No
Yes
Yes
Yes
(5)
significance at the 10 percent level, ** at the 5 percent level and *** at the 1 percent level.
fewer calories than the local bundle. All regressions are weighted using household weights and the standard errors are clustered at the comparison state-village-round level. * signifies
who moved at the time of marriage and column 10 includes additional interactions with an indicator variable that takes the value of 1 if the migrant’s origin-state bundle provide
third-order polynomial in log per-capita food expenditure (bar columns 4 to 6 which use alternate expenditure controls). Columns 2 and 3 restrict attention to households with wives
origin-state o. All specifications include comparison state-village-round fixed effects and flexible survey-round-specific controls for household size, demographics and type as well as a
for whether the household head or their spouse is an interstate migrant) interacted with indicators for the relationship with the comparison state s of a household living in state d from
the reference basket and column 9 matches households to a state-specific reference basket that is calculated separately by income-quartile. Independent variables are migranti (a dummy
budget shares for non-migrant households in state s (31 observations per household). Column 7 uses caloric shares in lieu of budget shares, column 8 includes migrant households in
Note: Dependent variable in columns 1 to 6 and 10 is the correlation between household is vector of food budget shares and a state-s reference basket comprised of the vector of mean
migrant i ×I d ≠s,o=s
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
Observations
R-squared
Food Expenditure Controls
Alternative Expnediture Controls
Demographics/Household Type
Comparison state s -Village-Round FE
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
migrant i ×I d=s
migrant i ×I d ≠s,o=s
migrant i ×I d ≠s,o≠s
migrant i ×I d=s
Observations
R-squared
Food Expenditure Controls
Alternative Expnediture Controls
Demographics/Household Type
Comparison state s -Village-Round FE
migrant i ×I d ≠s,o=s
migrant i ×I d ≠s,o≠s
migrant i ×I d=s
(1)
(2)
(3)
(4)
Correlation ρ is of the budget shares of household i with the average budget shares of state s (169 foods)
Basic — food
Wife Moved For
Wife Moved For
Total expenditure
expenditure controls
Marriage Sample A
Marriage Sample B
controls
-0.0137***
-0.00995**
-0.00711
-0.0143***
(0.00221)
(0.00388)
(0.00438)
(0.00224)
0.00393***
-0.000525
-0.000155
0.00354***
(0.000392)
(0.000620)
(0.000684)
(0.000397)
0.0213***
0.0100***
0.00239
0.0209***
(0.00206)
(0.00363)
(0.00424)
(0.00210)
7,291,293
2,886,503
2,607,069
7,360,609
0.838
0.875
0.878
0.831
Yes
Yes
Yes
No
No
No
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Table V: Comparing Bundles of Migrants and Non-Migrants
37
-0.00778**
(0.00350)
0.00570***
(0.000615)
0.0184***
(0.00349)
3,256,302
0.866
Yes
Yes
Yes
Children under 5 in
household
(9)
(11)
(12)
(13)
(14)
(15)
Correlation ρ is of the budget shares of household i with the average budget shares of state s (169 foods)
Children under 5 and Children under 16 in Children under 16 and Migrated 5 or more
Migrated 10 or more Migrated 20 or more
below median food
household
below median food
years ago
years ago
years ago
expend.
expend.
-0.00551
-0.0120***
-0.00632
-0.00867***
-0.00719***
-0.00358
(0.00682)
(0.00251)
(0.00493)
(0.00236)
(0.00259)
(0.00333)
0.00644***
0.00339***
0.00381***
0.00357***
0.00297***
0.00114**
(0.00124)
(0.000437)
(0.000856)
(0.000411)
(0.000440)
(0.000547)
0.0118*
0.0186***
0.00890*
0.0212***
0.0214***
0.0203***
(0.00650)
(0.00234)
(0.00469)
(0.00220)
(0.00237)
(0.00310)
1,632,150
5,537,468
2,454,022
7,172,129
7,084,678
6,932,065
0.892
0.850
0.878
0.839
0.840
0.841
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
(10)
(3)
(4)
(5)
(6)
(7)
Correlation ρis of the budget shares of household i with the average budget shares of state s (169 foods)
Consuming < 1850
Consuming < 2000
Consuming < 2400
Below median per
Bottom quartile per
Below median per
calories per day
calories per day
(rural) 2100 (urban)
capita expenditure
capita expenditure
capita food expend.
-0.0116**
-0.0111***
-0.00875***
-0.00642
-0.00733
-0.00619
(0.00508)
(0.00393)
(0.00314)
(0.00457)
(0.00935)
(0.00459)
0.00663***
0.00683***
0.00622***
0.00582***
0.00612***
0.00427***
(0.000902)
(0.000707)
(0.000567)
(0.000831)
(0.00156)
(0.000806)
0.0204***
0.0222***
0.0214***
0.00672
0.00170
0.00940**
(0.00480)
(0.00369)
(0.00300)
(0.00457)
(0.00897)
(0.00442)
2,212,656
2,880,241
4,106,632
2,793,038
1,271,434
2,782,095
0.874
0.869
0.861
0.879
0.896
0.873
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
(2)
-0.0173***
(0.00273)
0.00280***
(0.000502)
0.0247***
(0.00255)
4,035,425
0.851
Yes
Yes
Yes
Literate
(16)
Bottom quartile per
capita food expend.
-0.00199
(0.00842)
0.00393***
(0.00133)
0.00348
(0.00797)
1,289,352
0.886
Yes
Yes
Yes
(8)
percent level, ** at the 5 percent level and *** at the 1 percent level.
regressions are weighted using household weights and the standard errors are clustered at the comparison state-village-round level. * signifies significance at the 10
well as a third-order polynomial in log per-capita food expenditure. The column headings denote the various subpopulatons on which the regressions are run. All
o. All specifications include comparison state-village-round fixed effects and flexible survey-round-specific controls for household size, demographics and type as
their spouse is an interstate migrant) interacted with indicators for the relationship with the comparison state s of a household living in state d from origin-state
shares for non-migrant households in state s (31 observations per household). Independent variables are migranti (a dummy for whether the household head or
Note: Dependent variable is the correlation between household is vector of food budget shares and a state-s reference basket comprised of the vector of mean budget
Observations
R-squared
Food Expenditure Controls
Demographics/Household Type
Comparison state s -Village-Round FE
migrant i ×I d ≠s,o=s
migrant i ×I d ≠s,o≠s
migrant i ×I d=s
Observations
R-squared
Food Expenditure Controls
Demographics/Household Type
Comparison state s -Village-Round FE
migrant i ×I d ≠s,o=s
migrant i ×I d ≠s,o≠s
migrant i ×I d=s
Basic — food
expenditure controls
-0.0137***
(0.00221)
0.00393***
(0.000392)
0.0213***
(0.00206)
7,291,293
0.838
Yes
Yes
Yes
(1)
Table VI: Comparing Bundles of Migrants and Non-Migrants: Subpopulations
38
233,993
0.711
Yes
No
Yes
Yes
234,206
0.689
No
Yes
Yes
Yes
(0.00729)
234,206
0.815
No
Yes
Yes
Yes
(0.00596)
235,075
0.826
Yes
No
Yes
Yes
(0.00521)
(8)
234,225
0.790
No
Yes
Yes
Yes
(0.00573)
234,206
0.689
Yes
No
Yes
Yes
(0.00542)
234,229
0.818
No
Yes
Yes
Yes
(0.00592)
Real food expenditure State-specific food
controls
expenditure controls
-0.00303
-0.00531
(0.00369)
(0.00335)
-0.0222***
-0.0295***
(7)
level.
weights and the standard errors are clustered at the village-round level. * signifies significance at the 10 percent level, ** at the 5 percent level and *** at the 1 percent
includes an additional interaction with the distance between a migrant’s destination region and their origin state. All regressions are weighted using household
alternate expenditure controls). Column 11 instruments the polynomial in total food expenditure with a polynomial in total non-food expenditure. Column 12
for household size, demographics and type as well as a third-order polynomial in log per-capita food expenditure (bar columns 6 to 8 and column 12 which use
interstate (migrants) to those who moved intrastate (non-migrants). All specifications include village-round fixed effects and flexible survey-round-specific controls
Columns 4 and 5 restrict attention to households in which the wife of the household head moved village at the time of marriage and compares wives who moved
state-specific reference basket that is calculated separately by income-quartile, columns 14 and 15 use state- or district-specific reference bundles for the local bundle.
Column 9 uses caloric shares in lieu of budget shares for the reference basket (and calculates the cost to obtain 1000 calories), column 10 matches households to a
bundle from the migrant’s destination village both priced at destination-village prices. Columns 2 and 3 use alternative functions of these calorie differences.
with with an indicator variable that takes the value of 1 if a 1 Rupee reference bundle from the migrant’s origin-state provides fewer calories than a 1 Rupee reference
of food expenditure. Independent variables are migranti , a dummy for whether the household head or their spouse is an interstate migrant, and migranti interacted
Note: Dependent variable for all columns bar column 12 is log caloric intake per person per day. Dependent variable for column 12 is log caloric intake per Rupee
231,861
0.826
Yes
No
Yes
Yes
(0.00598)
(0.00351)
235,082
0.826
Yes
No
Yes
Yes
234,229
0.826
Yes
No
Yes
Yes
(0.00561)
× lndistance ov
Observations
R-squared
Food Expenditure Controls
Alternative Expnediture Controls
Demographics/Household Type Controls
Village-Round FE
(0.00577)
-0.00413
(0.00540)
83,965
0.837
Yes
No
Yes
Yes
(0.00999)
Total expenditure
controls
-0.00260
(0.00439)
-0.0285***
(6)
(12)
(13)
(14)
(15)
Log caloric intake (per person per day)
Log caloric intake per
Rupee of food
Log migrant travel Reference basket by Reference basket by
destination state
destination district
expenditure
distance interactions
-0.00369
0.0171
-0.00879**
-0.0107***
(0.00334)
(0.0200)
(0.00382)
(0.00342)
-0.0280***
-0.0267***
-0.0136***
-0.0110**
92,835
0.835
Yes
No
Yes
Yes
(0.00870)
(4)
(5)
Log caloric intake (per person per day)
Wife Moved For
Wife Moved For
Marriage Sample A Marriage Sample B
-0.00117
0.00410
(0.00580)
(0.00711)
-0.0220**
-0.0271***
migrant i
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
migrant i
migrant i
(10)
(11)
Log caloric intake (per person per day)
Reference basket
Food expenditure
Caloric share
by income quartile
instrumented
correlations
-0.0349***
-0.00746**
-0.00242
(0.00391)
(0.00340)
(0.00324)
0.0399***
-0.0246***
-0.0241***
(9)
234,229
0.826
Yes
No
Yes
Yes
(0.00954)
234,229
0.826
Yes
No
Yes
Yes
234,229
0.826
Yes
No
Yes
Yes
0.0629***
-0.0393***
(0.00585)
-0.0191***
(0.00471)
-0.0136***
(0.00485)
0.0227***
(0.00572)
lnK difference
interaction
(3)
× [lnK (bshareo ,Pv )-lnK (bsharev ,Pv )]
Observations
R-squared
Food Expenditure Controls
Alternative Expnediture Controls
Demographics/Household Type Controls
Village-Round FE
(0.00581)
Quartiles of lnK
difference
(2)
migrant i
× 1st quartile [lnK (.,.)-lnK (.,.)]
migrant i
× 2nd quartile [lnK (.,.)-lnK (.,.)]
migrant i
× 3rd quartile [lnK (.,.)-lnK (.,.)]
migrant i
× 4th quartile [lnK (.,.)-lnK (.,.)]
migrant i
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
migrant i
migrant i
Basic — food
expenditure controls
-0.00498
(0.00330)
-0.0283***
(1)
Table VII: Comparing the Caloric Intake of Migrants and Non-Migrants across Migration Routes
39
(0.00893)
104,699
0.839
Yes
Yes
Yes
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
Observations
R-squared
Food Expenditure Controls
Demographics/Household Type Controls
Village-Round FE
(0.0157)
52,540
0.840
Yes
Yes
Yes
(0.00655)
178,061
0.826
Yes
Yes
Yes
(0.0110)
89,948
0.830
Yes
Yes
Yes
(0.0116)
79,013
0.824
Yes
Yes
Yes
(0.00630)
230,627
0.827
Yes
Yes
Yes
(12)
(13)
Log caloric intake (per person per day)
Children under 16
Migrated 5 or more
and below median
years ago
food expend.
0.000205
-0.00100
(0.00915)
(0.00348)
-0.0217*
-0.0264***
(0.00784)
132,073
0.772
Yes
Yes
Yes
(4)
(5)
Log caloric intake (per person per day)
Consuming < 2400
Below median per
(rural) 2100 (urban)
capita expenditure
-0.00560
0.00347
(0.00460)
(0.00847)
-0.0249***
-0.0222**
(15)
(0.0111)
89,581
0.820
Yes
Yes
Yes
Below median per
capita food expend.
-0.000500
(0.00892)
-0.0220**
(7)
(0.00686)
227,978
0.827
Yes
Yes
Yes
0.00454
(0.00370)
-0.0276***
(0.00912)
223,332
0.828
Yes
Yes
Yes
0.0149***
(0.00436)
-0.0328***
Migrated 10 or more Migrated 20 or more
years ago
years ago
(14)
(0.0243)
40,958
0.869
Yes
Yes
Yes
Bottom quartile per
capita expenditure
0.0281
(0.0215)
-0.0547**
(6)
(0.00780)
129,493
0.844
Yes
Yes
Yes
-0.0108***
(0.00415)
-0.0223***
Literate
(16)
(0.0258)
41,532
0.861
Yes
Yes
Yes
Bottom quartile per
capita food expend.
0.0242
(0.0235)
-0.0449*
(8)
level.
weights and the standard errors are clustered at the village-round level. * signifies significance at the 10 percent level, ** at the 5 percent level and *** at the 1 percent
per-capita food expenditure. The column headings denote the various subpopulatons on which the regressions are run. All regressions are weighted using household
include village-round fixed effects and flexible survey-round-specific controls for household size, demographics and type as well as a third-order polynomial in log
origin-state provides fewer calories than a 1 Rupee reference bundle from the migrant’s destination village both priced at destination-village prices. All specifications
their spouse is an interstate migrant, and migranti interacted with with an indicator variable that takes the value of 1 if a 1 Rupee reference bundle from the migrant’s
Note: Dependent variable for all columns is log caloric intake per person per day. Independent variables are migranti , a dummy for whether the household head or
migrant i
0.00290
(0.00586)
-0.0364***
(11)
(0.0102)
92,592
0.770
Yes
Yes
Yes
Consuming < 2000
calories per day
-0.00412
(0.00558)
-0.0304***
(3)
Children under 5 and Children under 16 in
below median food
household
expend.
0.00296
-0.00588
(0.0120)
(0.00378)
-0.0266*
-0.0270***
(10)
(9)
Children under 5 in
household
(0.0129)
71,134
0.785
Yes
Yes
Yes
(0.00581)
234,229
0.826
Yes
Yes
Yes
Consuming < 1850
calories per day
-0.000922
(0.00717)
-0.0329**
Basic — food
expenditure controls
-0.00498
(0.00330)
-0.0283***
migrant i
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
Observations
R-squared
Food Expenditure Controls
Demographics/Household Type Controls
Village-Round FE
migrant i
migrant i
(2)
(1)
Table VIII: Comparing the Caloric Intake of Migrants and Non-Migrants across Migration Routes: Subpopulations
40
migrant i ×I d=s
migrant i ×I d=s
-0.00627*
(0.00322)
-0.000415
(0.000514)
0.00941***
(0.00322)
-0.00241
(0.00516)
-0.0173**
(0.00769)
-0.00182
(0.00478)
0.00471***
(0.000863)
0.0118**
(0.00471)
-0.00818
(0.00684)
-0.0253**
(0.0116)
-0.0193***
(0.00551)
-0.0173***
(0.00347)
0.0133***
(0.000687)
0.0461***
(0.00346)
-0.00958*
(0.00493)
-0.0628***
(0.0148)
-0.0236***
(0.00496)
Both migrants
7,291,293
234,227
235,203
Observations
7,291,293
234,227
235,203
Observations
0.838
0.826
0.826
R-squared
0.838
0.826
0.826
R-squared
fixed effects. * signifies significance at the 10 percent level, ** at the 5 percent level and *** at the 1 percent level.
order polynomial in log per-capita food expenditure. All regressions are weighted using household weights and the standard errors are clustered at the level of the
comparison state-village-round fixed effects. All panels include flexible survey-round-specific controls for household size, demographics and type as well as a third-
migrant, both head and spouse are migrants. Each panel comprises one regression. Panels 1-2 and 4-5 include village-round fixed effects, and Panels 3 and 6 include
different categories of within-household migrant structure: no spouse, only one of head or spouse is a migrant, only the spouse is a migrant, only the head is a
Note: Panels 1-6 repeat the regressions shown in Tables III, V and VII but interacting every instance of a migranti dummy variable with indicator variables for
Correlation ρ is of
the budget shares of migrant i ×I d ≠s
household i with the
migrant i ×I d ≠s,o=s
average budget
shares of state s
(6)
Log caloric intake migrant i
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
-0.0532***
(0.00541)
-0.00286***
(0.000971)
0.0102*
(0.00523)
0.00504
(0.00665)
-0.0487***
(0.0155)
migrant i
(5)
-0.0108***
(0.00372)
Only head is
migrant
Only spouse is
migrant
No spouse
-0.0114*
(0.00609)
-0.0174***
(0.00346)
0.0131***
(0.000686)
0.0460***
(0.00346)
-0.00943*
(0.00493)
-0.0628***
(0.0148)
-0.0234***
(0.00496)
Both migrants
-0.00511*
(0.00272)
0.000919**
(0.000448)
0.0100***
(0.00269)
migrant i
(4)
Log caloric intake
Correlation ρ is of
the budget shares of migrant i ×I d ≠s
household i with the
migrant i ×I d ≠s,o=s
average budget
shares of state s
(3)
Log caloric intake migrant i
× 1[lnK (bshareo ,Pv )<lnK (bsharev ,Pv )]
-0.00407
(0.00424)
-0.0187***
(0.00651)
-0.0130***
(0.00318)
Only head or
spouse is migrant
-0.0533***
(0.00541)
-0.00297***
(0.000970)
0.0101*
(0.00523)
0.00518
(0.00665)
-0.0488***
(0.0155)
migrant i
(2)
-0.0112*
(0.00608)
migrant i
(1)
Log caloric intake
No spouse
Table IX: Results broken-down by within-household migrant structure