ISSN (Print) - 0012-9976 | ISSN (Online) - 2349-8846

A+| A| A-

The Sensitivity of Food Security in India to Alternate Estimation Methods

There are a number of assumptions required to estimate actual calories consumed from expenditure data regarding calories contained in processed foods and meals eaten outside the household. Ideally, estimates of calorie consumption and the assessment of food security would be similar under a wide variety of assumptions, and the range of estimates would be quite small. However, this paper, which estimates the calories consumed per Indian household using the 61st round of the National Sample Survey, finds that the assessment of food security varies significantly under slightly different assumptions. Given the significant amount of measurement error in estimates of calories consumed, it is important to analyse not only household consumption surveys, but also aggregate food availability studies and health surveys collecting anthropometric measures that accompany undernourishment, such as stunting.

SPECIAL ARTICLE

The Sensitivity of Food Security in India to Alternate Estimation Methods

Sharad Tandon, Rip Landes

There are a number of assumptions required to estimate actual calories consumed from expenditure data regarding calories contained in processed foods and meals eaten outside the household. Ideally, estimates of calorie consumption and the assessment of food security would be similar under a wide variety of assumptions, and the range of estimates would be quite small. However, this paper, which estimates the calories consumed per Indian household using the 61st round of the National Sample Survey, finds that the assessment of food security varies significantly under slightly different assumptions. Given the significant amount of measurement error in estimates of calories consumed, it is important to analyse not only household consumption surveys, but also aggregate food availability studies and health surveys collecting anthropometric measures that accompany undernourishment, such as stunting.

Sharad Tandon (standon@ers.usda.gov) and Rip Landes are with the United States Department of Agriculture, Economic Research Service.

1 Introduction

D
espite rapid growth in income in India over the past two decades, current estimates of the number of food-insecure population derived from the aggregate production and trade statistics suggest that the country accounts for nearly 40% of the world’s food-insecure population (e g, Shapouri et al 2009). This estimate is corroborated by a high prevalence of a number of anthropometric indicators which accompany food insecurity, such as stunting (e g, Deaton and Dreze 2009). Even though there is overwhelming evidence that a very large share of the population does not have an adequate access to food, quantifying the extent of the problem remains problematic in both developed and developing countries (e g, Barrett 2010; Nord et al 2007).

Policymakers often focus on calories consumed and the sources of those calories – e g, carbohydrates, protein, etc. These figures are difficult to estimate precisely, where estimates can be derived from either aggregate production and trade statistics, or detailed household-level consumption data. Each approach describes a slightly different but important aspect of food security. Aggregate production and trade statistics better describe food availability and then use assumptions about the income distribution to estimate the number of food insecure households (e g, Shapouri et al 2009). However, household-level data is generally agnostic about the source of the consumption, but better allows policymakers to identify which individual households are food-insecure and also allows for slightly more precise estimation of calories consumed since the consumption data is provided for much more detailed food items (e g, Deaton and Subramanian 1996).

Given the importance of identifying the raw numbers and identity of the food-insecure, this study focuses on estimating the calorie consumption of individual households in India using the 61st round of the National Sample Survey (NSS). There have been a number of estimates using similar data (e g, Deaton and Dreze 2009; Deaton and Subramanian 1996; NSSO 2007). Estimates of calorie consumption are simply the estimates of the sum of calories contained in purchases of non-processed and processed food items, adjusted for calories contained in meals given to nonhousehold members and calories consumed in meals outside the household. But in constructing these estimates, there are a number of strong assumptions made to attach a calorie figure to each component, which in turn, lead to measurement error in calories consumed per household.

In this sample, over 30 days each household consumed 13.9 meals outside the household on average, gave on average 0.67

may 28, 2011 vol xlvi no 22

EPW
Economic & Political Weekly

meals to non-household members, and spent approximately Rs 132 on processed foods which do not have precise calorie i nformation available. These factors can significantly affect estimates of calories consumed per household given a number of households consume many more meals outside the household than the average of the entire sample, and given that the average amount spent on processed foods was approximately 8% of total expenditure on food items.

Based on the possible importance of these factors to total calorie consumption, this study analyses the size of the error in measuring calorie consumption. We first estimate the household calorie consumption using the approach first proposed by Deaton and Subramanian (DS) (1996). However, we alter the assumptions used to reach these estimates in a number of ways to analyse the resulting range of calorie consumption:

Calculation of Non-processed Calories Purchased: The DS approach calculates this figure by multiplying the quantity of each food item purchased, multiplied by the average number of calories contained per kilogram reported in the nutritional study Nutritive Value of Indian Foods (NVIF) (Gopalan et al 1989), published by the Ministry of Health and Family Welfare. However, we also calculate this figure by using the average number of calories contained per kilogram used by the Food and Agriculture Organisation (FAO), which results in a different figure.

Processed Calories Purchased: The DS approach assumes a household will spend more per calorie of processed foods than non-processed foods, and imputes the processed calories purchased from expenditure on non-processed foods. However, we also estimate processed calories purchased when assuming that processed calories are both more and less expensive relative to non-processed calories than assumed in DS.

Calories Consumed in Meals Eaten Outside the Household and Given to Non-household Members: We calculate the average calories consumed in each of these types of meals according to the methodology in DS by analysing how the purchase of calories, both processed and non-processed, varies with the number of these meals. Under the assumption that calories from these meals are perfectly interchangeable with the other types of calories purchased, the added calories purchased per meal given to non-household members will give an estimate of calories given away rather than consumed. Similarly, the fewer calories purchased per meal eaten outside the household will give an estimate of calories consumed that are not captured by the household’s calorie purchases. However, given the uncertainty in these estimates, we also estimate calorie consumption using a slightly inflated and deflated calorie value per meal consumed outside the household.

Altering the assumptions described above, we find a large range of possible estimates of the number of food-insecure people when using slightly different assumptions. Out of the 1,16,409 households in the sample, estimates of the number of households consuming less than 2,100 calories per household member ranged between 65,746 and 85,453, a difference of approximately

Economic & Political Weekly

EPW
may 28, 2011 vol xlvi no 22

17% of the total sample. Based on these sample households, estimates for the entire country suggest a difference of 17.5 crore food-insecure people between the highest and lowest estimate. Furthermore, the variation between the low and high estimate is qualitatively identical when using the definition of food insecurity used by the FAO, which is daily consumption below 1,800 calories per household member. These large ranges suggest that there is a lot of uncertainty over the actual number of food-insecure households in India. However, regardless of the large range, even the most conservative estimates of the number of food-insecure households is nearly two-thirds of the entire population.

Estimates of both average calorie consumption and the number of food-insecure households were particularly sensitive to two assumptions. First, estimates of food insecurity particularly d epend on which source one uses to calculate the number of calories per unit of non-processed food. Although seemingly a straightforward exercise, different sources result in large differences in calories consumption. Estimates derived when using the calorie data used by FAO result in a significant underestimate of calories consumed relative to the source used by the Indian government. Second, the consumption of packaged foods makes the results very dependent on oversimplified assumptions about how to estimate calories consumed based on the value of those packaged foods. Without more detailed information on the types of processed foods being eaten, it is difficult for estimates of calories consumed and estimates of food security to be more precise.

We draw two primary conclusions from this analysis. First, given India’s large population, this uncertainty over the assessment of food security also has large implications about the number of food-insecure people in the world. Second, given the large amount of uncertainty regarding the actual number of calories consumed, it is important to analyse not only consumption surveys, but also aggregate food availability studies and heath surveys collecting anthropometric measures that accompany undernourishment, such as stunting. The combination of the three types of studies is likely to result in the best assessment of food security and the nutritional well-being of household members. Furthermore, corroborating both trends in calorie consumption and correlations between calorie consumption and other household characteristics with a number of indicators would minimise the chance of detecting spurious relationships.

2 Data Summary and Baseline Estimation of Calories Consumed

This analysis utilises the 61st round of the NSS, which reports d etailed consumption data for 1,24,624 households between 2004 and 2005. In order to convert purchases of food quantities to calories purchased, we utilise the nutritional information r eported in NVIF. The text provides average calories contained per quantity of hundreds of distinct Indian foods, which allows us to find the aggregate number of calories purchased per household.

When aggregating the household level consumption data, we provide a few robustness checks to make sure the household consumption patterns calculated are consistent with NSS calculations. Namely, the NSS already calculates the monthly per capita expenditure (MPCE) on food items, which should simply

SPECIAL ARTICLE

be the sum of the values of all the food items a household con non-household members will give an estimate of calories con
sumes in positive quantities. Calculating this sum and checking sumed in those meals. Similarly, the fewer calories purchased
if it does in fact equal to the MPCE, there are 7,277 out of the per meal eaten outside the household will give an estimate of
1,24,624 households where this equality does not hold. In all calories consumed in those types of meals.
empirical analysis, these households have been excluded. Based on these estimates, the sample suggests that there are
Furthermore, we exclude households with implausibly high and 403.4 calories contained in each meal given to non-household
low calorie purchases to try to avoid including households with member and each meal consumed outside the household.5 Thus,
obvious recording errors in consumption.1 In all, these adjust to find an estimate of total calories consumed, we simply multiply
ments exclude 8,215 of the 1,24,624 households, leaving a total the number of meals given to non-household members by 403.4
of 1,16,409.2 calories, and subtract this value from total calories purchased.
However, this is not a random sample of Indian households. Similarly, we multiply the number of meals consumed outside the
Rather, the sample is geographically dispersed, stratified into household by 403.4 calories and add it to total calories purchased.
r ural and urban portions, and further stratified based on meas- It is important to note that there is a great amount of measure
ures of income. Portions of rural villages and urban towns are ment error introduced by this procedure. First, it is possible that
randomly selected to sample households, and within these different types of meals received are likely to be associated with
r egions, households of particular income groups and a particular different calorie amounts consumed from the average of all meals
sector (i e, rural or urban) are randomly sampled. Thus, the sam given to non-household members. Second, the estimate used
ple is a representative of the entire country when properly before has a relatively large confidence interval, and we cannot
weighting the households in the survey based on their prevalence statistically differentiate between any calorie adjustment between
in the entire country. Given the oversampling of affluent house 218.0 calories and 588.9 calories at the 5% significance level.
holds, in all the following empirical results, we present both sum- Regardless, summary statistics of this adjusted estimate of
mary statistics for the entire sample and population estimates calorie consumption are presented in Table 1. Columns (1) and
derived by weighting households by the inverse of the probability (2) present summary statistics for all households in the sample
of being selected into the sample.3 and columns (3) and (4) present estimates of the entire popula-
However, as discussed in the introduction, here are a number tion. All figures are less than the threshold of 2,100 calories typi
of types of processed foods that are difficult to match to precise cally used to measure food security. Furthermore, the estimated
nutritional information. For example, the data set reports con average calories consumed for the entire country is lower than
sumption of items such as “salted refreshment”, “cake/pastry”, that estimated from all sample households, which reflects the
“other processed food”, etc. Additionally, since some of these sampling procedure – richer households are oversampled relative
vague food items come in a variety of different forms, it is diffi to their prevalence in the entire population.
cult to report quantities purchased and the data set only reports Table 1: Summary Statistics of Adjusted Calorie Totals
the value of those purchases in rupees.4 Following Deaton and Variable All Sample Households Total Country Estimates Ave Count of Ave Count of
Subramanian (1996), we calculate the non-processed calories Food-insecure Food-insecure
purchased per rupee spent for each household and assume that each calorie from processed food costs 50% more than non (1) (2) (3) (4) Calories consumed per household member 2,026.8 73,137 1,948.6 70,57,06,578
processed calories. Number of households 1,16,409 -
Additionally, individuals consume meals outside the home and provide meals to others from their purchased food items. These (1) Estimates calculated using the 61st round of the NSS. Population estimates are constructed using NSS multipliers to calculate population weights. (2) The first row calculates the number of non-processed calories purchased per household
issues further obfuscate the actual number of calories consumed member; the second row calculates the sum of non-processed calories and processed calories purchased, with an estimate of calories consumed in meals given to non-household members
by household members in many instances. For example, if poorer subtracted from this total, and an estimate of calories consumed in meals eaten outside of the household added to this total.
households are more likely to eat meals at their place of employ
ment, then looking only at food items purchased is likely to un 3 Estimating Calories Consumed Using an Alternative
derstate their caloric intake. On the other hand, if richer house- Source for Nutritional Information
holds are more likely to give a higher number of meals to employ- As discussed in Section 2, an estimate of total calories consumed
ees and guests, then their food item purchases are likely to over is produced using assumptions about the link between quantity
state their caloric intake. However, the NSS data set provides of food consumed and calories, the amount of calories consumed
d etailed information on the number of meals received by house in processed foods, and assumptions about meals consumed
hold members outside of the household as well as the number of outside the household and meals served to non-household mem
meals given to non-household members. bers. However, when we slightly alter these assumptions, we can
Following Deaton and Subramanian (1996), we use regression drastically affect estimates of the number of food-insecure
techniques to analyse how the purchase of calories, both proc households.
essed and non-processed, vary with the number of meals con- In addition to using the concordance between NVIF, we also
sumed outside the household and meals given to non-household link calories and quantities of food consumed used by the FAO
members. Under the assumption that calories from these meals for estimating aggregate calorie availability for south Asian
are perfectly interchangeable with the other types of calories countries.6 Each source differs slightly in calorie information,
purchased, the added calories purchased per meal given to where 69 non-processed food items have an NVIF calorie value
94 may 28, 2011 vol xlvi no 22 Economic & Political Weekly
EPW

that is higher than the FAO calorie value, and 37 non-processed food items have an FAO calorie value being higher than the NVIF value. Thus, whether one source results in a higher estimated calorie consumption depends on which of these food items households tended to spend more of their income.

Analysing the relationship between the two measures, we estimate the following specification: Non – processedCaloriesir = IJ + ȕFAON on – ProcessedCaloriesir

r

+ ĮControlsir + İir where caloriesir denotes the total number of calories purchased in household i and district r when calories are calculated using NVIF; FAO Calories denotes the same measure except calories are calculated using a concordance used by the FAO; IJ denotes district fixed effects and Controls denotes a number of household characteristics that are used as control variables to help absorb unobserved variation in calories purchased.7 Estimates of ȕ will d escribe how calories consumed will differ between the two sources. Estimates of calorie consumption using the nutritional information in NVIF will be higher if ȕ>1, and conversely, will be lower if ȕ<1.

Estimates of the specification are presented in Table 2. Column

(1) estimates a sparse specification with no fixed effects or control variables, column (2) presents the estimate when adding

Table 2: Correlation between NVIF Calories and FAO Calories

Dependent Variable: Total Calories Purchased

(1) (2) (3)

FAO calories 1.058*** (.0038) 1.064*** (.0033) 1.063*** (.0068)

Fixed effects N Y Y

Control variables N N Y

Observations 1,16,409 1,16,409 1,16,409

*** Denotes significance at the 1% level. Standard errors clustered by district are reported in parentheses.

Table 3: Comparing Summary Statistics of Calories Using Alternate Sources

Variable All Sample Households Total Country Estimates Ave Count of Ave Count of Food-insecure Food-insecure

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

Calories consumed using NVIF 2,026.8 73,137 1,948.6 70,57,06,578

Calories consumed using FAO 1,916.6 81,557 1,826.0 78,89,82,256

Number of households 1,16,409

  • (1) Estimates calculated using the 61st round of NSS. Population estimates are constructed using NSS multipliers to calculate population weights.
  • (2) The first row calculates the estimate of calories consumed presented in Table 1, which accounts for calories from non-processed foods purchased using NVIF nutritional totals, processed foods purchased, calories eaten in meals away from home, and calories given to non-household members. The second row is exactly the same except for using FAO sources to calculate the total non-processed calories purchased.
  • district fixed effects and column (3) adds a number of household characteristics as control variables. Since ȕ is greater than one in all specifications, the estimates suggest that calorie estimates using FAO sources predict calorie purchases that are significantly less than estimates using the more detailed NVIF. Column (3) suggests that increasing the FAO calories purchased by one unit is associated with an increase in NVIF calories by 1.063, which is a difference of 6.5%.8

    Given this significant difference between the two sources, we analyse the summary statistics of calorie consumption and the count of food-insecure households. As suggested by the estimates in Table 2, estimates of calorie consumption in Table 3 is lower when using the FAO source to calculate calories than when using NVIF. As a consequence of this difference, the estimates of the

    Economic & Political Weekly

    EPW
    may 28, 2011 vol xlvi no 22

    food-insecure households are larger when using the FAO source. Column (2) of Table 3 suggests that the number of food-insecure households is 8,420 higher, which is slightly over 7% of the total sample, when using FAO sources. In column (4), when estimating the difference in food security for the entire country, there is an additional 8.3 crore people who are food insecure.

    4 Estimating Calories Consumed Using Alternative Meal Assumptions

    As discussed in Section 2, estimation of the calories consumed per meal introduces a significant amount of measurement error into the estimates of calories consumed per household. Here we alter our assumptions about the amount of calories consume in each meal and analyse the effect these changes have on the estimates of calories consumed and the count of food-insecure households.

    First, we account for the measurement error introduced by e stimating the calories used to prepare meals for guests. Based on the estimates of calories contained in meals given to nonhousehold members and meals eaten outside the household, we cannot statistically reject the hypothesis of the true number of calories per meal given to guests to be anywhere in the 95% confidence interval. Thus, rather than simply insert 403.4 calories per meal given to guests for all meals not accounted for in raw calorie purchases, we also look at the estimates of calorie consumption when using the bounds of the 95% confidence interval, which are 218 calories and 588.9 calories.9

    The summary statistics are presented in Table 4. As expected, we find in column (1) that the average calories consumed increased when using the upper bound and decreased when using the lower bound. In column (2), this translated into small adjustments in the number of food-insecure households, but the difference between using the point estimate and using one of the bounds of the confidence interval is not large relative to the differences discussed above when using the nutritional information used by FAO rather than the nutritional information reported in NVIF. However, in column (4), the difference in the number of food-insecure people between the estimates derived using the upper and lower bound is approximately 2.3 crore people.

    Table 4: Calorie Consumption Using Alternative Meal Assumptions

    Variable All Sample Households Entire Country Estimates
    Ave Count of Ave Count of
    Food-insecure Food-insecure
    (1) (2) (3) (4)
    Calories consumed with original
    meal assumption 2,026.8 73,137 1,948.6 70,57,06,578
    Calories consumed using lower
    bound for meal estimate 2,003.7 74,338 1,930.1 71,67,42,564
    Calories consumed using upper
    bound for meal estimate 2,050.0 71,766 1,967.2 69,36,07,698
    Number of households 1,16,409 -
  • (1) Estimates calculated using the 61st round of the NSS. Population estimates are constructed using NSS multipliers to calculate population weights.
  • (2) The first row calculates the estimate of calories consumed presented in Table 1, which accounts for calories from non-processed foods purchased using NVIF nutritional totals, processed foods purchased, calories eaten in meals away from home, and calories given to nonhousehold members. The second row is exactly the same except uses the lower bound of the 95% confidence interval of the estimate of calories per meal received outside the household and given to others, which is 218 calories. The third row is similar, except it uses the upper bound of the 95% confidence interval of the estimate of calories per meal received outside the household and given to others, which is 588.9 calories.
  • SPECIAL ARTICLE

    5 Estimating Calories Consumed Using Alternative Processed Food Assumptions

    We have to make an assumption about the number of calories each household consumes for each rupee spent on processed foods which are not easily assigned a calorie value. In Section 2, we simply assumed that the price of each calorie purchased through processed foods was marked up by 50% over the price of each calorie purchased of non-processed food items. However, here we relax this assumption and analyse the effect the assumption has on food security as allow this relationship to vary. Specifically, we analyse the effects of assuming that calories purchased through processed foods were the exact same price as calories purchased in non-processed foods, and we also analyse the scenario in which the process food calories were four times as expensive.

    As expected, we find in column (1) of Table 5 that the average number of calories consumed is significantly higher when we a ssume that processed foods are cheaper, where the difference between the cheapest and most expensive scenarios corresponds to approximately a 6% decrease in average calories consumed per household member. In column (2), this translated into a large range of food-insecure households between scenarios. Specifically, when comparing the cheapest and most expensive scenarios, the number of food-insecure households rises by 9,019, which is 7.8% of the entire sample. In column (4), the difference in the count of food-insecure people in the entire country is nearly 7.2 crore people.

    Given the large amount of uncertainty introduced by oversimplified assumptions about the nutritional information of processed foods, it could be valuable to combine the NSS

    Table 5: Calorie Consumption Using Alternative Processed Food Assumptions

    Variable All Sample Households Entire Country Estimates Ave Count of Ave Count of Food-insecure Food-insecure

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

    Calories consumed with original processed food assumption 2026.8 73,137 1948.6 70,57,06,578

    Calories consumed using cheap processed calories 2106.2 67,201 2016.2 65,73,55,404

    Calories consumed using expensive processed calories 1987.2 76,220 1914.9 72,92,25,470

    Number of households 1,16,409

  • (1) Estimates calculated using the 61st Round of the NSS. Population estimates are constructed using NSS multipliers to calculate population weights.
  • (2) The first row calculates the estimate of calories consumed presented in Table 1, which accounts for calories from non-processed foods purchased using NVIF nutritional totals, processed foods purchased, calories eaten in meals away from home, and calories given to non-household members. The second row is exactly the same except assumes that processed calories cost the same as non-processed calories. The third row is similar, except it assumes that processed calories are four times as expensive as non-processed calories.
  • consumption data with other data sources better describing the consumption habits of Indian consumers. For example, Euromonitor International reports both the total expenditure on all packaged foods sold by grocery retailers and the total by subcategories, such as “salted and savory snacks”, that correspond to NSS codes of processed foods. Furthermore, the data source reports the share of each of these subcategories captured by particular brands, which could better indicate the calorie information of the items that account for the majority of processed food consumption. Although it would still be difficult to estimate actual calorie consumption, these sorts of estimation procedures have a more logical basis than simple assumptions about how much more expensive calories from processed foods are than the non-processed.

    6 Total Range of Food Security Assessment

    Sections 3 through 5 have demonstrated that slightly altering the assumptions used by Deaton and Subramanian (1996) lead to possibly large changes to the food security assessment of India. In order to present a possible range for the assessment of food security, we present assessments using the assumptions discussed above that lead to the largest and smallest average calories consumed per household member. Specifically, the high calorie estimate of calories consumed uses the nutritional information presented in NVIF, uses the upper bound of the e stimate of calories consumed in meals outside the home and given to non-household members (588.9 calories), and assumes that processed calories are the same price as non-processed calories. On the other hand, the low calorie estimate uses the nutritional information used by FAO, uses the lower bound of the confidence interval to estimate the calories consumed in meals outside the home and given to non-household members and assumes that processed calories are four times the price of non-processed calories.

    Summary statistics for these variables are presented in Table 6a. As expected, there is a very large range of average calories consumed for each household. In column (1), there is a difference of

    273.5 calories between the average high and low calorie estimate. In column (2), this translates into a difference in the count of food-insecure households of 19,707, which is approximately 17% of the total sample. Estimates in column (4) suggest that the difference in food-insecure people between the two scenarios is nearly 17.5 crore, which was approximately 14.3% of the foodinsecure people estimated in the United States Department of Agriculture – Food and Nutrition’s (USDA) Food Security Assessment (Meade et al 2006).

    Furthermore, Table 6b (p 97) demonstrates that the severity of food-insecurity is drastically different between the two scenarios. When using the low calorie estimate, there is nearly twice as high an incidence of severe food-insecurity (consumption less than 1,500 calories) in column (1), and the share of the households that

    Table 6a: Calorie Consumption Using Alternative Processed Food Assumptions

    Variable All Sample Households Entire Country Estimates
    Ave Count of Ave Count of
    Food-insecure Food-insecure
    (1) (2) (3) (4)
    Calories using Deaton-
    Subramanian (1996) assumptions 2,026.8 73,137 1,948.6 70,57,06,578
    High calorie estimate 2,129.4 65,746 2,034.8 64,41,26,869
    Low calorie estimate 1,855.9 85,453 1,775.7 81,87,60,913
    Number of households 1,16,409 -
  • (1) Estimates calculated using the 61st round of the NSS. Population estimates are constructed using NSS multipliers to calculate population weights.
  • (2) The first row calculates the estimate of calories consumed presented in Table 1, which follows Deaton and Subramanian (1996) to account for calories from non-processed foods purchased using NVIF nutritional totals, processed foods purchased, calories eaten in meals away from home, and calories given to non-household members. The second row is the same except it uses the upper bound of the estimate of calories consumed in meals eaten outside the household and given to non-household members, and it assumes that processed calories cost the same as non-processed calories. The third row uses an estimate derived from the FAO nutritional source for south Asian countries, uses the lower estimate for calories consumed in meals eaten outside the home and given to non-household members, and assumes that processed foods cost four times as much as non-processed foods.
  • may 28, 2011 vol xlvi no 22

    EPW
    Economic & Political Weekly

    Table 6b: Changes to the Depth of Food Security Using Alternative Processed Food Assumptions

    Variable Share of All Sample Households in Food Security Range

    Cal<1500 1500<Cal<1800 1800<Cal<2100 2100<Cal<2400

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

    High calorie estimate .121 .201 .243 .197

    Low calorie estimate .256 .255 .223 .141

    Share of Total Population in Food Security Range

    Variable Cal<1500 1500<Cal<1800 1800<Cal<2100 2100<Cal<2400

    High calorie estimate .134 .232 .260 .185

    Low calorie estimate .300 .285 .214 .116

  • (1) Estimates calculated using the 61st round of the NSS. Population estimates are constructed using NSS multipliers to calculate population weights.
  • (2) The high calorie estimate uses the upper bound of the estimate of calories consumed in meals eaten outside the household and given to non-household members, and it assumes that processed calories cost the same as non-processed calories. The low calorie estimate uses an estimate derived from the FAO nutritional source for south Asian countries, uses the lower estimate for calories consumed in meals eaten outside the home and given to non-household members and assumes that processed foods cost four times as much as non-processed foods.
  • are marginally food-insecure and marginally food-secure in c olumns (3) and (4) are significantly lower.

    7 Measurement Error and Estimation Using Calories Consumed

    In addition to obfuscating the actual number of food-insecure people in the sample, the measurement error introduced when estimating calories consumed has important implications for any estimation using calories consumed as a dependent or independent variable. Analysis of the household characteristics that help determine calories consumed can be particularly interesting and helpful to policymakers trying to decrease health problems in poorer households, and has been studied extensively in many different settings (e g, Strauss 1986, Behrman and Deolalikar 1988, Deaton and Dreze 2009, etc).

    First, in our construction of calories consumed per household member, we used household size, the amount of money used to purchase food and a number of other household-level variables. We cannot consistently estimate the relationship between calories consumed as the dependent variable and any of these variables, or even variables constructed using them, given the possibility of division bias (Borjas 1980). Measurement error in each of these variables used in constructing the dependent and independent variables can cause a spurious correlation even if the true correlation between the variables is precisely zero. This precludes a simple estimation of the relationship between calories consumed per household member and a number of interesting household characteristics.

    Furthermore, it is still difficult to consistently estimate correlations between calories consumed and any household characteristic that was not used in constructing the estimate of calories consumed. For example, the measurement error in calories consumed is increasing in the amount spent on processed foods. However, the amount a household spends on processed foods is correlated with income and a number of household decision variables that might be of interest to researchers and policymakers analysing health outcomes.

    Thus, for example, it would be particularly difficult to interpret a simple positive correlation between income and calories consumed. Even without measurement error in calories consumed, an omitted variable bias makes it difficult to know whether income was driving the correlation or something correlated with it. But

    Economic & Political Weekly

    EPW
    may 28, 2011 vol xlvi no 22

    we would know that households with larger incomes tended to consume more calories and we would have a rough estimate of the sensitivity. However, with the measurement error described in Sections 5-7 being correlated with income, we cannot even be confident that true calorie consumption is actually higher in those households, let alone the sensitivity of calorie consumption to income, since the correlation could be driven by the correlation between income and the measurement error.

    Additionally, these issues affect the interpretation of trends of calorie consumption over time. For example, estimations suggest that calorie consumption in India has been decreasing over time (e g, NSSO 2007). However, given the measurement error des cribed above, it is difficult to argue that true calorie consumption has actually been decreasing, as opposed to possible increases or decreases originating from changes in consumption of food categories that are adding to the measurement error, such as changes to consumption of processed foods and meals eaten away from the household. In particular, over this time period, the value of consumption of processed foods has been increasing and the share of processed foods in total food expenditures has also been increasing. Thus, it is difficult to argue that true calorie consumption has been decreasing, as opposed to systematic underestimates of calories consumed from processed foods.

    These issues suggest that researchers and policymakers should rely on a number of different indicators of calorie consumption and food security, such as health indicators associated with calorie consumption and studies of aggregate food availability, when analysing trends or correlations. Additionally, policymakers should also focus on the response of food security and calories consumed to policy interventions and income and production shocks that cause an exogenous variation in the independent variables of interest.

    8 Conclusions

    There are a number of assumptions required to estimate calories consumed per household member when using detailed consumption data. Assumptions regarding the amount of calories in each type of food, calories consumed in processed foods, meals eaten outside the household and meals given to non-household members all obfuscate the number of calories consumed by households. These assumptions result in a noisy estimate of calories consumed, which policymakers and researchers can use to help both assess and analyse the efficacy of programmes designed to improve food security.

    Estimating calories consumed in India using household-level data, the country with the largest number of food-insecure individuals in the world, we find that the estimated number of households that are classified as food-insecure varies significantly u nder slightly different assumptions. In particular, the estimates of calories consumed were quite sensitive to assumptions accounting for calories consumed from processed foods and also the particular link used to convert quantities of food consumed to calories. The difference between the highest and lowest estimate of the number of people consuming less than 2,100 calories in our sample is approximately 17.5 crore people. Although this suggests

    SPECIAL ARTICLE

    food security might be significantly better than the worst-case any change in the assessment of food security by a number scenario, even the most conservative estimates of food security of measures to minimise the chance of detecting spurious in India suggest that over 60% of the population is food-insecure. trends and correlations. Thus, it is important to analyse not This uncertainty over the assessment of food security translates only consumption surveys, but also aggregate food availability into a large amount of uncertainty in the number of food- insecure studies and heath surveys collecting anthropometric measures people in the world. that accompany undernourishment, such as stunting. The combi-

    Given the possibility that food security assessments have such nation of the three types of studies is likely to result in the best large bounds, the size of which depends on household behaviour, assessment of food security and the nutritional well-being of

    researchers and policymakers need to take care in corroborating household members.

    Notes

    1 Households have been excluded if they consume less than 500 calories purchased per member if they either gave more than one meal per day to a non-household member or did not receive at least the equivalent of one meal a day from other sources. Lastly, households with more than 10,000 calories purchased per household member per day have been excluded if they did not give more than 500 meals to non-household members.

    2 All empirical results and patterns are robust to inclusion of all 1,24,624 households.

    3 Probability weights are calculated using the multipliers provided by the NSS, which are the number of households in the entire population represented by the household in the sample. We first multiply the multiplier by the size of the household, which gives us the number of people that are represented by the household members, and then divide by the total population to get the weights used in the calculations. Results are qualitatively identical if households are treated as the level of observation, with weights equal to the multiplier divided by the total number of Indian households estimated in the 2001 Census.

    4 There are a number of items where only values of consumption are given, and not quantities. Also, there are other items which give quantities of consumption, but it might be too difficult to estimate from those quantities (e g, some particular types of processed foods). These items are baby food, other birds, egg products, other beverages, biscuits and confectionary, salted refreshment, prepared sweets, cooked meals, cake/pastry and other processed food.

    5 See Appendix for the derivation of this estimate and a more complete discussion.

    6 This concordance is available online at http://faostat.fao.org/site/368/DesktopDefault.aspx? PageID=368.

    7 Control variables include the amount of money spent on pan, tobacco, intoxicants, fuel, the MPCE for the household, the household size, an indicator for whether the household is rural and an indicator for whether the household reporting insufficient food sources during the year.

    8 The coefficients are very precisely estimated, which implies that even the lower bound of the 95% confidence interval still suggests a significant difference in estimates using different concordances. Out of all three estimates, column (1) has the smallest lower bound of the 95% confidence interval, which suggests that we can statistically reject the null hypothesis that the coefficient estimate is anything lower than 1.051 at the 95% significance level. 9 See Appendix for a discussion of this estimate.

    10 Same as fn 7.

    References

    Behrman, Jere and A Deolalikar (1988): “Seasonal Demands for Nutrient Intakes and Health Status in Rural South India” in David E Sahn (ed.), Causes and Implications of Seasonal Variability in Household Food Security (Baltimore, MD: Johns Hopkins University Press).

    Barrett, Christopher (2010): “Measuring Food Insecurity”, Science, 327(5967): 825-28.

    Borjas, George (1980): “The Relationship between Wages and Weekly Hours of Work: The Role of D ivision Bias”, Journal of Human Resources, 15(3): 409-23.

    Deaton, Angus and S Subramanian (1996): “The D emand for Food and Calories”, Journal of Political Economy, 104(1): 133-62.

    Deaton, Angus and J Dreze (2009): “Food and Nutrition in India: Facts and Interpretations”, Economic & Political Weekly, 44(7): 42-65.

    Food and Agriculture Organisation: Food Security Statistics, accessed 24 July 2010 (http://www. fao.org/economic/ess/food-security-statistics/ en/).

    Gopalan, C, B V Rama Sastri and S C Balasubramanian (1989): Nutritive Value of Indian Foods (Hyderabad: National Institute of Nutrition, Indian Council of Medical Research 3rd edition).

    Meade, Birgit, S Rosen and S Shapouri (2006): “Food Security Assessment, 2005”, Outlook GFA-17, US Department of Agriculture, Economic Research Service. Available at: www.ers.usda.gov/publications/.

    NSSO (2007): “Nutritional Intake in India 2004-2005”, National Sample Survey Organisation, Ministry of Statistics and Programme Implementation, Report No 513.

    Nord, Mark, M Andrews and S Carlson (2007): “Household Food Security in the United States, 2007”, ERR-66, US Dept of Agriculture, Economic Research Service, available at: www.ers.usda. gov/publications/.

    Shapouri, Shahla, S Rosen, B Meade and F Gale (2009): “Food Security Assessment, 2008-09”, Outlook GFA-20, US Department of Agriculture, Economic Research Service. Available at: www. ers.usda.gov/publications/.

    Strauss, John (1986): “Does Better Nutrition Raise Farm Productivity?”, Journal of Political Economy, 94(2): 297-320.

    Appendix: Estimation of Calories Contained in Meals Eaten Outside the Household and Meals Given to Non-household Members

    Following the methodology introduced by Deaton and Subramanian (1996), we estimate calories contained in various types of meals by estimating the following specification:

    Caloriesir=IJU+ȕHomeMealsir+ȕMealsGivenir

    + ȕMealsReceivedir+ĮControlVarsir+İLU

    where IJ denotes a district fixed effect to help

    r

    absorb unobserved characteristics shared by all households within a district; Caloriesir denotes the total amount of calories purchased, both processed and non-processed, by household i in district r; HomeMeals denotes the total number of meals eaten by household members at home; MealsGiven denotes the total number of meals the household served to non-household members; MealsReceived denotes the total amount of meals consumed by household members outside of the home, which is simply the sum of meals received from school, employers, other sources and meals received on payment; and ControlVars denotes a number of household characteristics that are used as control variables to help absorb unobserved variation in calories purchased.10 The coefficients from this specification can help us adjust the calories purchased to better reflect the number of calories actually consumed by household members. In particular,

    ,QGLD0LFUR 6PDOODQG0HGLXP(QWHUSULVHV5HSRUW

    FDUU\LQJ

    FRPSUHKHQVLYHLQIRDQGDQDO\VLVDWQDWLRQDODQGJOREDOOHYHOV

    -XVWUHOHDVHG

    DWWKHRPPRQZHDOWK060(RPSHWLWLYHQHVVRQIHUHQFH %HQJDOXUX0D\

    )RUGHWDLOVDQGFRSLHV

    3XEOLFDWLRQV0DQDJHU ,QVWLWXWHRI6PDOO(QWHUSULVHVDQG'HYHORSPHQW ,6('+RXVH,6('5RDGRFKLQ VHR#LVHGRQOLQHRUJLVHGLQVPDOO#JPDLOFRP

    may 28, 2011 vol xlvi no 22

    EPW
    Economic Political Weekly

    ȕ2 provides an estimate of the number of the Table A1: Calories Purchased Per Meal Consumed
    total calories purchased that were consumed Dependent Variable: Total Household Calories Purchased
    by non-household members for each meal, and (1) (2) (3) (4)
    ȕ3 provides an estimate of the number of calo- Meals at home 662.3*** 664.5*** 663.8*** 611.8***
    ries consumed by household members in meals (3.15) (3.21) (7.22) (7.61)
    they eat outside of the household. Thus, to Meals given 409.1*** 405.9*** 403.4*** 325.3***
    a rrive at a better estimate of calories con (82.4) (81.4) (94.4) (76.6)
    sumed by household members, we would sim- Meals received -17.9 - - -
    ply subtract ȕ2*MealsGiven from calories pur (22.3)
    chased, and add ȕ3 *MealsReceived from calo- Meals received from school --344.6*** -191.0*** 6.93
    ries purchased. (27.4) (36.2) (35.1)
    Estimates of the specification are presented in Meals received from employer --367.2*** -123.7*** 157.5***
    Table A1. In column (1), we estimate a sparse (38.2) (45.4) (44.8)
    specification excluding the district fixed effects Meals received from payment --313.1*** 51.6 -67.1*
    and control variables. The coefficient for meals (28.2) (42.7) (37.2)
    consumed at home is precisely estimated with a Meals received from other sources -312.3*** 365.0*** 330.5***
    narrow confidence interval; the coefficient for (46.4) (50.3) (47.6)
    meals given is also precisely estimated, but has District fixed effects N N Y Y
    a much larger 95% confidence interval. On the Control variables N N N Y
    other hand, meals received is very imprecisely Obsevations 1,16,409 1,16,409 1,16,409 1,16,409
    estimated and has an implausibly small magni * Denotes significance at the 10% level; *** Denotes significance at the 1% level.
    tude for calories consumed at a meal. Standard errors clustered by district are reported in parentheses.
    Investigating whether this is true for all types variables on the total calories purchased. For expositive and has a similar m agnitude through
    of meals received outside the household, in col ample, given that these meals include meals at out specifica tions.
    umn (2) we break up the variable into the ceremonies such as weddings, it could be that Based on the se results, estim ates of the amount
    number of meals received at school, from em primarily richer households r eceive these meals of calories associated with meals given by
    ployers, from other sources, and meals received at ceremonies such as weddings, and thus, the households s eem to be stable and precisely esti
    for payment. The coefficients on meals eaten at coefficient is partially capturing the effects of inmated acros s specifications. However, meals re
    home and meals given to non-household mem come on calories purchased. Or perhaps, receivceived outsi de of the househ old are difficult to
    bers are essentially unchanged. However, there ing a meal from another household requires the interpret. T he coefficients of meals received
    seems to be a large asymmetry in the variation household to provide a meal to that guest in refrom school, employers and f or payment change
    in calories purchased in response to different turn on some other occasion. signs and ma gnitudes across specifications. The
    types of meals received. Meals from school, Columns (3) and (4) try to account for some unonly stable c oefficients are t hose on meals reto be
    employers and other sources are precisely esti observed heterogeneity by estimating specificeived from other sources, which seem
    mate and as expected, suggest that households cations including fixed effects and control varicapturing th e effect of omitte d variables since it
    decrease their total calorie purchases by ap ables. Again, the coefficients on meals eaten has a sign op posite to what o ne would expect.
    proximately 350 calories. at home and meals given to non-household Thus, when looking to ad just the calories
    However, still interpreting column (2), increas members are essentially unchanged. However, consumed by individual h ouseholds
    ing meals received from other sources actually the coefficients on the variables accounting for sponse to m eals eaten outs ide the home, we
    suggests households increase their calorie pur meals received outside of the household drop use the esti mated coefficient on meals given in
    chases by approximately 370 calories. This is the in magnitude dramatically, and they all change column (3) when adjusting h ousehold calories
    opposite sign of what one would expect when signs between specifications. The one exceppurchased t o a more reaso nable estimate of
    holding all other factors fixed, and the coefficient tion is the coefficient on meals received from calories con sumed for bot h meals received
    is likely picking up the effects of some omitted other sources. The coefficient consistently remains and given.

    in re-

    REFLECTIONS ON EMPIRE March 26, 2011 Resurrection and Normalisation of Empire – Rohit Chopra Taming the Imperial Impulse: Realising a Pragmatic Moral Vision – Abdullahi Ahmed An-Na’im Adam’s Mirror: The Frontier in the Imperial Imagination – Manan Ahmed Indian Empire (and the Case of Kashmir) – Suvir Kaul Imperial Democracies, Militarised Zones, Feminist Engagements – Chandra Talpade Mohanty Rethinking News Agencies, National Development and Information Imperialism – Oliver Boyd-Barrett Digital Imperialism through Online Social/Financial Networks – Radhika Gajjala, Anca Birzescu Pandemic, Empire and the Permanent State of Exception – Cindy Patton For copies write to: Circulation Manager, Economic and Political Weekly, 320-321, A to Z Industrial Estate, Ganpatrao Kadam Marg, Lower Parel, Mumbai 400 013. email: circulation@epw.in Economic Political Weekly may 28, 2011 vol xlvi no 22 99
    EPW

    To read the full text Login

    Get instant access

    New 3 Month Subscription
    to Digital Archives at

    ₹826for India

    $50for overseas users

    Comments

    (-) Hide

    EPW looks forward to your comments. Please note that comments are moderated as per our comments policy. They may take some time to appear. A comment, if suitable, may be selected for publication in the Letters pages of EPW.

    Back to Top