The effects of food stamps on food consumption a review of the literature

              A 9tJI: F73/+
Current Perspectives on Food Stamp Program Participation
States
Department of
Agriculture
Food and
Nutrition
Service
Office of
Analysis and
Evaluation
(I)
he Effects of
Food Stamps on
Food Consumption
A Review of the
Literature
iss^
r *PIPP!
Current Perspectives on Food Stamp Program Participation
Titles in this series:
Food Stamp Program Participation Rates
(November 1988) Pat Doyle and Harold Beebout
Food Stamp Program Participation Rates Among the Poverty
Population, 1980-1987
(November 1988) Carole Trippe and Harold Beebout
Determinants of Participation in the Food Stamp Program:
A Review of the Literature
(November 1989) Susan Allin and Harold Beebout
Estimating Rates of Participation in the Food Stamp Program:
A Review of the Literature
(November 1989) Carole Trippe
Food Stamp Program Participation Rates:
August 1985
(April 1990) Pat Doyle
The Effects of Food Stamps on Food Consumption:
A Review of the Literature
(October 1990) Thomas M. Fraker
(I
United States Food and 3101 Park Center Drive
Department of Nutrition Second Floor
Agriculture Service Alexandria, VA 22302
The Effects of Food Stamps on
Food Consumption:
A Review of the Literature
Thomas M. Fraker
A product of
Mathematics Policy Research, Inc.
600 Maryland Avenue, S.W.
Suite 550
Washington, DC 20024
October 1990
m
ACKNOWLEDGMENTS
This report has benefitted from contributions by many people. The idea for the report
originated with Steven Carlson of the Food and Nutrition Service (FNS), who recognized a need
for the consolidation of existing research findings on the effects of food stamps on food
consumption. Harold Beebout and Jim Ohls of Mathematica Policy Research (MPR) reviewed
the outline for the report and suggested several major structural changes that are reflected in the
final report. Harold Beebout also reviewed and commented on an early draft of the report.
Robert Moffitt of Brown University reviewed Appendix A, which presents the economic theory
of the effects of food stamps on food consumption. Any errors that remain in either the body
of the report or in Appendix A are the responsibility of the author(s) rather than the reviewers.
Three additional individuals made especially notable contributions to the report:
Gary Bickel of FNS reviewed an early draft of the report and provided
numerous detailed suggestions for improvements in its structure and text;
he is also the coauthor of Appendix A.
Tom Good of MPR edited both an early draft of the report and a much-revised
later draft.
Liz Quinn of MPR drafted several chapter summaries, contributed to the
editing of the report, and oversaw the production of the draft and final
versions of the report.
MPR Project Number:
FNS Contract Number:
FNS Project Officer:
7925-040
53-3198-0-22
Alana Landey
This analysis was performed under a competitively awarded contract in the amount of $1,812,081.
iV
CONTENTS
Chapter Page
ACKNOWLEDGMENTS iii
I. INTRODUCTION 1
A. OBJECTIVE AND SUMMARY OF TfflS REPORT 2
B. THE STRUCTURE OF TfflS REPORT 4
H. MEASUREMENT ISSUES IN ESTIMATING THE EFFECTS
OF FOOD STAMPS ON FOOD CONSUMPTION 5
A MEASURING FOOD CONSUMPTION 6
1. Measures of Food Consumption 6
2. Existing Survey Data on Food Consumption 12
3. Issues Associated with Measuring and Analyzing
Food Consumption 16
B. MEASURING FSP ELIGIBILITY AND PARTICIPATION 23
1. Errors in Measuring FSP Participation and Benefits 23
2. Errors in Modeling Food Stamp Eligibility 24
3. Data Requirements for Modeling FSP Participation 27
Dl THE CONSUMPTION PATTERNS OF FOOD STAMP
RECIPIENTS AND LOW-INCOME NONRECTPIENTS 29
A HOUSEHOLD EXPENDITURE PATTERNS 29
1. Expenditure Shares 29
2. The Money Value of Food Used 30
3. Nutrients per Dollar's Worth of Food 34
4. Home Food Use by Food Group 35
5. Frequency of Food Shopping 37
6. Perceived Food Adequacy 37
B. THE NUTRIENT AVAILABILITY OF HOUSEHOLDS
AND THE NUTRIENT INTAKE OF INDIVIDUALS 40
1. Nutrient Availability 41
2. Nutrient Intake 43
CONTENTS (continued)
Chapter £2fi£
HI. (continued)
C INDIVIDUAL FOOD INTAKE BY FOOD GROUP 46
D. SUMMARY 50
IV. THE EFFECTS OF FOOD STAMPS ON FOOD EXPENDITURES 53
A. A FRAMEWORK FOR ESTIMATING THE EFFECTS
OF FOOD STAMPS ON FOOD EXPENDITURES 55
1. Research Strategies 55
2. Specification of an Empirical Model of Food Expenditures 57
B. HOW EFFECTIVE ARE FOOD STAMPS AT INCREASING
FOOD EXPENDITURES? 6°
1. Estimates of the MPCf Out of Food Stamps 61
2. Critique of the Estimates 65
C. ARE COUPONS MORE EFFECTIVE THAN CASH
BENEFITS AT INCREASING FOOD EXPENDITURES? 68
1. Findings from Non-Cashout Studies 69
2. Findings from Food Stamp Cashout Studies 72
D. SUMMARY 75
V. THE EFFECTS OF FOOD STAMPS ON THE QUALITY
OF DIETS 79
A THE RELATIONSHIP BETWEEN DIETARY QUALITY
AND NUTRITIONAL STATUS 80
1. Nutritional Status, Dietary Quality, and Analyses
of the FSP 81
B. THE EFFECTS OF FOOD STAMPS ON NUTRIENT
AVAILABILITY 86
1. The Structure of Models of Nutrient Availability 87
2. Estimates of the Effects of Food Stamps on
Nutrient Availability 91
VI
CONTENTS (continued)
Chapter Page
V. (continued)
C THE EFFECTS OF FOOD STAMPS ON NUTRIENT
INTAKE 94
1. The Data Sets and Models Used in Studies of
Nutrient Intake 95
2. Estimates of the Effects of Food Stamps on
Nutrient Intake 100
3. Estimates of the Effects of Cash Food Assistance
on Nutrient Intake 104
D. SUMMARY 105
REFERENCES 109
APPENDDC A: THE ECONOMIC THEORY OF THE EFFECTS OF FOOD STAMPS
ON FOOD CONSUMPTION
APPENDDC B: MEASURES OF HOUSEHOLD SIZE AND COMPOSITION
APPENDDC C: INFORMATION ON THE SOURCE OF THE MPCf ESTIMATES
IN CHAPTER IV
vu
i- it J Jli ipaas
i////
TABLES
Table Page
DX1 HOUSEHOLD NUTRIENT AVAILABILITY AS A PERCENTAGE
OF THE RDA FOR PERSONS EATING IN THE HOUSEHOLDS 42
m.2 NUTRIENT INTAKE AS A PERCENTAGE OF THE RDA:
MEAN PER INDIVIDUAL, ONE DAY OF INTAKE DATA 44
HI.3 NUTRIENT INTAKE AS A PERCENTAGE OF THE RDA:
MEAN PER INDIVIDUAL, FOUR NONCONSECinTVE DAYS
OF INTAKE DATA 45
IV.l ESTIMATES OF THE MARGINAL PROPENSITY TO CONSUME
FOOD (MPCf) AT HOME, FROM SELECTED STUDIES 62
V.l SELECTED STUDIES OF THE EFFECTS OF FOOD STAMPS
ON NUTRIENT AVAILABILITY 89
V.2 THE EFFECTS OF INCOME AND FOOD STAMPS ON NUTRIENT
AVAILABILITY, MEASURED AS A PERCENTAGE OF THE
ADULT MALE RDA: A COMPARISON OF ESTIMATES FROM
THREE STUDIES 93
V.3 SELECTED STUDIES OF THE EFFECTS OF FOOD STAMPS
ON NUTRIENT INTAKE 96
V.4 THE PERCENTAGE CHANGE IN NUTRIENT INTAKE ASSOCIATED
WITH PARTICIPATION IN THE FSP: A COMPARISON OF
ESTIMATES FROM FOUR STUDIES 101
ix
mm P
m
FIGURES
Figure Page
ELI HOUSEHOLD EXPENDITURE SHARES BY MAJOR EXPENDITURE
CATEGORY , 31
ffl.2 MONEY VALUE OF FOOD USED IN A WEEK BY HOUSEHOLDS 32
IH3 MONEY VALUE OF FOOD USED PER PERSON IN A WEEK
BY HOUSEHOLDS 33
m.4 SHARE OF HOME FOOD EXPENDITURES BY FOOD GROUP 36
IIL5 FREQUENCY OF MAJOR FOOD SHOPPING BY HOUSEHOLDS 38
ffl.6 SELF-EVALUATION OF HOUSEHOLD FOOD ADEQUACY 39
m.7 FOOD INTAKE BY FOOD GROUP: MEAN PER INDIVIDUAL PER
DAY, ONE DAY OF DATA 47
m.8 FOOD INTAKE BY WOMEN AGES 19 TO 50 BY FOOD GROUP:
MEAN PER INDIVIDUAL PER DAY, FOUR NONCONSECUTIVE
DAYS OF DATA 48
HL9 FOOD INTAKE BY CHILDREN AGES 1 TO 5 BY FOOD GROUP:
MEAN PER INDIVIDUAL PER DAY, FOUR NONCONSECUTIVE
DAYS OF DATA 49
L INTRODUCTION
Studies on the determinants of household food expenditures have a long history, dating
to the time of the Prussian statistician Ernst Engel (1857). Engel used several 19th-century data
sets to analyze the relationship between food expenditures and income, and used his analytical
findings to formulate Engel's Law: the proportion of income spent on food falls as income rises.
This law has been confirmed in study after study over the past 130 years.
Research on the effects of food stamps on food consumption has a much shorter history,
in that food stamps did not come into existence in the United States until the 1930s.1 Herman
Southworth's pioneering theoretical analysis of the effects of food stamps on household food
expenditures was published in 1945, but the first empirical studies on this topic were not
conducted until the early 1970s. Interest in the effectiveness of food stamps at increasing food
expenditures and the quality of diets was generated at that time by growing concern about the
existence of hunger in the United States and by the rapid growth of the Food Stamp Program
(FSP). The program's growth during the early-to-mid 1970s can be traced to the adoption of two
sets of amendments to the Food Stamp Act of 1964: the amendments of 1970, which mandated
nationally uniform food stamp eligibility standards and allotment schedules, and the amendments
of 1973, which required that all U.S. counties begin operating the FSP by mid-1974. During this
same period, two nationally representative household survey data sets that provide information
on household income, food stamp benefits, and food expenditures became available to
1During the Great Depression of the 1930s, food stamps were provided to needy households
not only in an attempt to alleviate hunger but also to reduce surplus agricultural commodities that
had been accumulated by the federal government This early Food Stamp Program was
terminated in 1943, after the country's war effort eliminated agricultural surpluses. After a lapse
of nearly twenty years, food stamps were reintroduced as a pilot program during the Kennedy
Administration. By the late 1970s, the program had evolved into what is essentially the current
Food Stamp Program.
researchers: the first five years (1968-72) of the University of Michigan's Panel Study of Income
Dynamics and the Bureau of Labor Statistics' 1973-74 Consumer Expenditure Diary Survey. The
combination of a pressing public policy problem (hunger among low-income households), a rapidly
growing program designed to alleviate that problem, and the availability of data sets capable of
supporting research on the problem and the programmatic response precipitated a number of
empirical studies of the FSP in the mid-1970s.
The release of data from the low-income supplement to the USDA's 1977-78 Nationwide
Food Consumption Survey, and from a special follow-up survey in 1979-80, was followed by a
steady flow of empirical research throughout the 1980s on the effects of food stamps on food
consumption (including measures of the nutritional quality of food used by households and eaten
by individuals, as well as measures of the money value of food used). This research was
stimulated by the fact that the FSP was (and continues to be) one of the country's largest social
welfare programs, providing benefits to approximately 20 million persons per month over most
of the decade. The on-going policy debate about the merits of coupons versus cash food
assistance provides additional stimulus for continued research on the effectiveness of food stamps
at increasing food consumption and the quality of diets. As this debate continues, the imminent
release of data from the 1987-88 Nationwide Food Consumption Survey is likely to generate
renewed interest in research on the food consumption effects of food stamps in the 1990s.
A. OBJECTIVE AND SUMMARY OF THIS REPORT
The many studies of the effects of food stamps on food consumption that have been
conducted during the past two decades have been based on underlying data sets, analytic
techniques, and food consumption outcome measures that vary widely. Such variation, as well
as the sheer volume of the research results, makes it difficult for the potential user of this
research to grasp either the consensus findings or the range of findings on the effects of food
stamps on food consumption. The objective of this report is to rectify this situation by
systematically summarizing in one document the findings from 17 studies of the effects of food
stamps on the money value of food used by households, 8 studies of the effects of the FSP on
the availability of nutrients in the household from the home food supply, and 8 studies of the
effects of the FSP on the intake of nutrients by individuals.
On the basis of this review, we can report that the provision of an additional dollar's worth
of food stamps (i.e., food coupons) to a recipient household is estimated to stimulate the
consumption of additional food from the home food supply with a money value of roughly 20 to
45 cents. This effect may be compared with estimates of the food-consumption response to a
dollar of cash income that range from 5 to 10 cents. Whether the effect of cash food assistance
on food consumption would be more similar to the effect of coupons or the effect of ordinary
cash income is a major question that is unanswered by this literature.
This review also notes that the existing estimates of the effects of food stamps on the
quantity of nutrients that are available to recipient households from their home food supplies are
consistently large and positive. The estimates of the effects on nutrient availability are roughly
two to seven times greater for a dollar's worth of food coupons than for a dollar of cash income.
The research findings on the effects of food stamps on the intake of nutrients by
individuals are far less definitive than the findings on nutrient availability and the money value
of food used by households. Across studies and nutrients, only a small proportion of estimates
of the effects of food stamps on nutrient intake differ from zero at conventional levels of
statistical significance. The estimated effects are both positive and negative in sign, but, as noted,
few of those estimates are statistically significant. Moreover, estimates of the effects of cash
income on nutrient intake also tend to be statistically insignificant and inconsistent in sign. Thus,
this body of literature provides little support for the hypothesis that food coupons and cash
income have positive effects on the intake of nutrients.
B. THE STRUCTURE OF THIS REPORT
This report consists of four substantive chapters. The first two present background
material that provides a context for interpreting the findings from the studies on food
consumption that are reviewed in the final two chapters. Chapter II identifies the data sets that
have served as the basis for most studies of the effects of food stamps on food consumption. It
describes the various measures of food consumption that have been constructed with those data,
as well as the limitations of the data for studies of the FSP. Chapter III provides a comparative
overview of the food and nonfood expenditure patterns and quality of the diets of households and
individuals that receive food stamps and of those that do not, based on published descriptive
studies.
Chapter IV summarizes and critiques 17 empirical studies of the effects of food stamps on
the money value of food used at home. Central to the chapter is a table that presents estimates
from each of the reviewed studies of the effects of an additional dollar of food stamps and of
ordinary cash income on food expenditures.
The literature on the effects of food stamps on the availability of nutrients in the
household and on the intake of nutrients by individuals is based on more heterogeneous statistical
models than is the expenditure literature, and is thus more difficult to summarize succinctly.
Chapter V classifies eight existing models of nutrient availability and eight models of nutrient
intake into several different categories, describes the qualitative estimates of food stamp and cash
income effects generated by the models, and summarizes the quantitative estimates from three
studies of nutrient availability and four studies of nutrient intake.
EL MEASUREMENT ISSUES IN ESTIMATING THE EFFECTS
OF FOOD STAMPS ON FOOD CONSUMPTION
The theoretical basis for much of the research on the effects of food coupons and cash
benefits on household food consumption is provided by Southworth (1945).2 Using the basic
tools of microeconomic theory that were originally expounded by Hicks (1939), Southworth
derived a model which predicts the effects of marginal changes in the value of food stamps, cash
food assistance, and ordinary cash income on household food consumption. In analyzing the
effectiveness of the Food Stamp Program at achieving its food expenditure and nutritional
objectives, researchers often test hypotheses generated by Southworth's theoretical model by
using data from household surveys whose methodology and purpose vary widely. The surveys
range from geographically limited data collection efforts designed for specific program evaluations,
to nationally representative surveys of household food use and individual food intake, to general-purpose,
nationally representative surveys that gather a wide range of information from
respondent households, including their usual expenditures on food. This chapter describes the
major household surveys that provide food consumption data, examines the measures of food
consumption that are used, and assesses their appropriateness for analyzing the effects of food
stamps. It also examines the implications of the survey designs for the reliability of statistical
estimates of the effects of food stamps on food consumption. Finally, the chapter assesses the
capacity of the data from these surveys to support modeling food stamp eligibility and
participation by low-income households. That capacity can influence the size and even the sign
of estimates of the effects of food stamps on household food consumption.
2Sce Appendix A for an analysis of Southworth's theory, including an explanation of
Southworth's methodology, a discussion of the hypotheses generated by his theory, and an
examination of empirical research findings on household food consumption behavior which, in
general, fail to support those hypotheses.
A. MEASURING FOOD CONSUMPTION
The different measures of food consumption that are used in household surveys and in
analyses of data from those surveys can be bewildering. This section defines the most commonly
used measures, explains how they are related to each other, and indicates the research
applications for which each measure is most appropriate. It also identifies the primary household
surveys that provide data on those measures and form the basis for most analyses of the effects
of the FSP on food consumption. The section concludes with a discussion of selected issues
associated with measuring and analyzing survey data on food consumption.
1. Measures of Food Consumption
Measures of food consumption fall into three categories: measures of expenditures on
food by the household, measures of food used by the household from its home food supply, and
measures of food actually eaten by members of the household. We examine each of these
categories in turn.
Food Expenditures. The most straightforward approach for measuring household food
consumption is to ask households to recall or to keep a record of their purchases of food over
a given period of time. For example, respondents to the 1985-86 Continuing Survey of Food
Intakes by Individuals (CSFII) and the 1987-88 Nationwide Food Consumption Survey (NFCS)
were asked to recall the average expenditure of cash and food coupons on food per week or per
month by their households over the previous several months, both for foods purchased for home
use and for meals and snacks eaten away from home. Respondents to the University of
Michigan's continuing Panel Survey of Income Dynamics (PSID) are asked to recall their cash
expenditures for both at-home and away-from-home food over the previous month. Respondents
to the on-going diary component of the Consumer Expenditure Survey (CEX) are provided with
structured ledgers in which they record on a dairy basis their household food purchases and away-
from-homc meals over a two-week period. With all of these methodologies, respondents must
distinguish between food purchased in supermarkets, specialty or convenience stores, carry-outs,
and the like for home use, and purchases of meals and snacks that are eaten away from home.
A measure of household food expenditures has several deficiencies when used in analyses
of the effects of food stamps on food consumption:
1. For some households, a high level of spending on food represents the
purchase of more expensive foods rather than foods capable of providing
more ample or better diets. Overall, however, the dollar value of food
purchased is a good proxy for the physical quantity or nutritional quality
of foods purchased.
2. An expenditure measure of food consumption omits home-produced
foods and foods received as gifts, charity, or payment-in-Ljd and, thus,
may understate actual food consumption.
3. An expenditure measure of food consumption may, by including food
that is provided to boarders or guests, fed to pets, or lost through
spoilage or other waste, overstate the actual physical consumption of
food by household members.
4. Hie expenditure recall methodology (as opposed to the diary
methodology) is vulnerable to the omission of purchases made with food
stamp benefits. Despite instructions to include such purchases, food
stamp recipients tend to include only cash purchases in their reported
average expenditure on food. The PSID addresses this problem by
asking food stamp recipient respondents to report the amount of their
cash food expenditures, over and above food-stamp purchases. Because
the diary methodology requires that the respondent record the quantity
and cost of eaca food item purchased, it is less vulnerable than the recall
methodology to the omission of foods purchased with food stamps.
5. The diary methodology is sensitive to monthly cycles in household food
shopping.3 A large proportion of food stamp recipients conducts its
major food shopping on a monthly basis. For some such households that
are participating in a diary survey of food consumption, the major food
shopping occurs within the reporting period; for others it does not
While such variation has no effect on the sample mean of the diary
measure of food purchases, it increases the standard error of the mean,
^The monthly food shopping patterns of low-income households are described in Chapter HI,
below.
thus making it more difficult to obtain statistically significant estimates of
the effects of food stamps on food purchases.
A fundamental strategic question in measuring and analyzing household food consumption
is whether the two separate measures of food expenditures for home use and for meals and
snacks away from home should be combined into one measure of total food expenditure, or
whether the two components should be analyzed separately. An argument for treating them
separately is that the cost of restaurant meals includes the value added for preparation and
serving, as well as the cost of food ingredients, and is thus not strictly comparable with the cost
of foods purchased for home use. However, with the proliferation of "fast-food" restaurants
which sell foods that may be eaten away from home or brought into the home, and of the many
highly pre-processed and "ready-to-eat" foods now available in food markets, this distinction has
become less meaningful.
Most measures of the nutrient content of foods consumed have been computed from
survey data on household food use and individual food intakes, as described below. However,
diary data on the quantities and types of foods purchased by households can also be converted
into measures of the nutrients provided by those foods. One of the studies of nutrient availability
(Scearce and Jensen, 1979), which is reviewed in Chapter VI, is based on diary data on household
food purchases.
Food Use. A methodology for collecting data on food consumption that provides a
comprehensive and detailed measure of a household's home food use is employed by the
Nationwide Food Consumption Survey. This methodology generates data on all foods used at
home by the household, whether purchased, home-produced, or received as a gift or payment-in-kind.
The NFCS household food consumption data are thus more inclusive than expenditure
measures of home food consumption. Under the NFCS methodology, the survey respondent
keeps informal records of all foods used by the household from the home food supply-both foods
8
eaten at home and those carried from home (e.g., bag lunches)-over a one-week period. The
source of the food is noted, as are the quantities used of each item and the costs of all purchased
items.
Both dollar scales and nutrient scales can be used to measure a household's use of food
from its home food supply. Thus, these data will support both economic analyses of food
consumption behavior that focus on the dollar value of foods used from the home supply, and
dietary analyses that focus on the nutrient content of the foods used. The money value of food
used by a household is computed by multiplying the unit cost of each type of food by the number
of units used by the household and summing over all of the different types of food.4 The
availability of nutrients in the food used by a household is computed on a nutrient-by-nutrient
basis by multiplying the amount of a nutrient per pound of each type of food by the number of
pounds used by the household and summing over all of the different types of food.5 Most
analyses of the effects of food stamps on food consumption have relied on one or the other of
these two measures of food used at home. In interpreting the findings from those studies (as
reported in Chapters IV and V), readers should note that "moneyvalue" and "nutrient availability"
are alternative measures of the same food consumption behavior-a household's use of food from
its home food supply.
Analysts often use measures of per-capita nutrient availability that have been adjusted to
compensate for meals eaten away from home by household members to assess the nutritional
4For a food item that was used but not purchased by a household, the price used to compute
its money value is the average price paid for the same item by the households of the other
respondents to the survey.
sOne of several university or USDA nutrient databases can be used to convert data ou food
quantities to data on nutrient availability. These databases provide information on the nutrient
content of roughly 4,000 foods and food combinations in the form in which they enter the
household, with adjustments for cooking losses and inedible components of foods. Most of the
nutrient values are supported by laboratory analyses, but some are imputed on the basis of data
for similar foods. Hepburn (1982) provides a description of the USDA's nutrient data base.
adequacy of the food used from the household food supply. They compute the measures of
nutrient availability by adjusting the measure of household size downward by an amount that
depends on the number of meals eaten away from home, as described in Appendix B. The
smaller adjusted measure of size reflects the fact that the household members may not be fully
dependent on the home food supply.
Because survey measures of household food use are based on an item-by-item accounting,
rather than on an aggregate recall, they are believed to be relatively accurate. Unlike the diary
measures of food purchases, which are also based on an item-by-item accounting, measures of
food use are subject to relatively little variation from week to week within a month, because
households exhibit greater stability in their food use than in their food purchases.
In addition to measuring home food use, the NFCS uses the recall method to measure
usual purchases of food at home and food away from home. According to data from the 1977-78
NFCS core sample, the mean of the money value of food used at home is 9 percent larger than
the mean expenditure on food used at home. This difference is to be expected, given that the
measure of food use is more comprehensive than the measure of food purchases.
Food Intake. Food intake data are collected at the individual level, in contrast to food use
data, which are collected at the household level Two different survey methodologies are used
to measure the food intakes of individuals: a 24-hour recall of all foods eaten and a daily diary
of foods eaten. Under either methodology, respondents report the types and quantities of foods
that they actually ate during the survey's reference period. The NFCS combines these
methodologies, using the recall method to obtain intake data for the first of three consecutive
days and the diary method to obtain data for the other two days.
In principle, because individual intake data usually include an indication of where each
eating occasion occurred, respondents' at-home and away-from-home food consumption can be
distinguished. However, most dietary assessments based on intake data pertain to the total
intakes of individuals. The individual intake data are limited for undertaking economic analyses
10
of food consumption, since the costs of foods are not captured in these measures. However, the
intake data lend themselves well to dietary assessments, using separate nutrient-based scales for
measuring the nutrient content of food intakes. For example, an individual's intake of calcium
can be computed by using a nutrient database to determine the amount of calcium that is
provided by each food item eaten by the individual and then summing over all of the reported
food items.6 It is also possible to measure nutrient intake at the household level by summing the
computed intakes of all members of the household.
The sum of nutrient intake over all members of a household (i.e., household nutrient
intake) differs from a measure of household nutrient availability in two ways. First, nutrient
availability is computed only on the basis of food used from the home food supply, whereas
nutrient intake can be computed on the basis of both food at home and food away from home.
Second, even when nutrient intake is computed on the basis of food obtained from the home
food supply, the combined nutrient intake of household members will in principle be smaller than
the household's nutrient availability because some food that is used by a household is not eaten
by household members; it is served to guests or boarders, lost, wasted, or fed to pets. In addition,
the 1977-78 NFCS data for household-level nutrient availability and the combined nutrient intakes
of household members from home supplies indicate that the individual data tend to understate
food consumption relative to the household-level data, even after all known differences are
accounted for (Batcher, 1983). Such understatement suggests that some degree of systematic
error also may be present in one or both types of the NFCS food consumption data.
By placing restrictions on the use of food coupons, the FSP is designed to stimulate
primarily purchases of food for use at home rather than food away from home. Purchased food
6The key distinction between nutrient databases that are used to evaluate individual food
intake and those that are used to evaluate household food use is that the former provide nutrient
information on foods in the forms in which they art- jaten rather than in the forms in which they
enter the household.
11
m
is by far the largest component of a'i food used at home; thus, measures of nutrient availability
based on food used from the home food supply are well-focused measures for assessing the
effectiveness of the FSP at achieving its dietary objectives. On the other hand, because measures
of total nutrient intake are based on food eaten away from home as well as at home, they may
not be as effective at addressing the behavior on which the FSP is designed to have a direct
influence.7 It may be that measures of nutrient availability from the home food supply thus
provide more sensitive indicators of the potential dietary effects of the FSP than do measures of
total nutrient intake by individuals.
2. Existing Survey Data on Food Consumption
Three household surveys have provided the data for most empirical studies of food
consumption: the Nationwide Food Consumption Survey provides data on food use and food
expenditures by households, as well as data on food intake by individuals; the Consumer
Expenditure Survey and the Panel Study of Income Dynamics provide data only on household
food expenditures. This section provides an overview of these three surveys and identifies the
methodologies that they use to measure food consumption.
TheNFCS. The Nationwide Food Consumption Survey is the most widely used source of
data for analyzing the effects of food stamps on food consumption. The USDA has conducted
seven national surveys of household food consumption since the 1930s, the most recent of which
was the 1987-88 NFCS. All of those surveys collected data on food consumption by the
household as a unit, and, in addition, the three latest surveys (1965-66, 1977-78, and 1987-88)
collected data on food intake by individual members of the household.
7On the basis of 1977-78 NFCS data, HNIS (July 1982) reports that food away from home
accounts for 13 percent of the total expenditure on food by low-income households. This
relatively small percentage suggests that food away from home is unlikely to dramatically dampen
the effect of the FSP on total nutrient intake relative to its effect on the intake of nutrients from
foods used from the home food supply.
12
To facilitate analyses of USDA food assistance programs, the two most recent editions of
the NFCS have included special supplemental surveys of low-income households. These
supplements were much like the core surveys, except that the samples were restricted to
households that satisfied approximations to the income-eligibility screens for the FSP.8 Two
nationally representative supplemental surveys of low-income households were conducted in
conjunction with the 1977-78 NFCS-4,400 low-income households were interviewed in 1977-78,
prior to the elimination of the food stamp purchase requirement (EPR), and 2,900 low-income
households were interviewed in 1979-80, just subsequent to the EPR.9 More existing estimates
of food stamp effects on food consumption are based upon the 1977-78 low-income supplement
than any other data source. The sample of completed interviews for the low-income supplement
to the 1987-88 NFCS will contain approximately 2,400 households. Public-use files containing
data for that sample are scheduled to be released in 1991.
The NFCS uses the recall methodology to measure a household's usual purchases of food
at home and food away from home. In addition, the survey obtains data from each participating
household on all foods used from the home food supply over a one-week period. Finally, the
NFCS obtains three consecutive days of food-intake data for each member of a participating
household. To avoid seasonal biases, the NFCS distributes interviews with sample households
evenly over a one-year period.
Over the past two decades, the trend in NFCS sample response rates has been strongly
downward. The response rate for the household component of the 1965-66 survey was 85
households that participated in the two low-income supplemental surveys that were
conducted in conjunction with the 1977-78 NFCS satisfied an approximation to the FSP eligibility
screen on liquid asset holdings in addition to an approximation to the FSP income screens. The
low-income sample for the 1987-88 NFCS was not subjected to an asset screen.
9These two supplemental surveys are formally referred to as the "Low-Income Supplement
to the 1977-78 NFCS" and the "USDA Survey of Food Consumption in Low-Income Households,
1979-80."
13
percent, the response rate for the 1977-78 low-income supplement was 69 percent, and a
preliminary estimate of the response rate for the 1987-88 low-income supplement is SO to 55
percent The sharp drop in the NFCS response rate appears to be associated with changes in
family structures and family meal preparation and eating patterns, and with increases in labor-market
activity by women. These fundamental social and economic changes have reduced the
likelihood that, for a given sample household, a survey worker will be able to locate and complete
an interview with an adult who is knowledgeable about the food consumption of the entire
household. If nonresponse is not a random occurrence, but is associated instead with household
characteristics, then a low response rate introduces the possibility that the component of the
sample for which interviews were completed successfully is not representative of the survey's
target population.
The CEX. Between 1888 and 1973, the Bureau ofLabor Statistics conducted eight surveys
of expenditures by U.S. households. In 1979, BLS began to collect household expenditure data
on an on-going basis via the continuing Consumer Expenditure Survey. That survey consists of
two separate components-an Interview Survey, which collects data on major purchases and on
smaller periodic expenses (such as utility bills) over a three-month reference period, and a Diary
Survey, which collects data on small, frequently purchased items (such as food) over a two-week
reference period. Sample units for the Interview Survey are interviewed quarterly for five
successive quarters, generating approximately 4,800 completed interviews per quarter. The
sample for the Diary Survey is drawn annually, and interviews with sample units are distributed
over all weeks of the year. Each sample unit is interviewed only once, and the completed annual
sample size is approximately 4,800.
Respondents to the CEX Interview Survey are asked to recall their usual expenditures on
food at home and on food away from home during the preceding three months, whereas
respondents to the CEX Diary Survey keep daily logs of their food purchases for two consecutive
14
weeks. As noted previously, the recall methodology is subject to the omission of food purchases
made with food stamps. Because the diary methodology is perceived to measure food purchases
more accurately, most CEX-based studies of the effects of food stamps on food consumption have
used data from the Diary Survey.
The PSID. The Panel Study of Income Dynamics is an on-going longitudinal survey of
approximately 5,000 U.S. households from all income strata. It is conducted annually by the
Institute for Social Research at the University of Michigan under contract to DHHS. Historically,
approximately 10 percent of the sample of households have reported receiving food stamps in the
month preceding the interview. The PSID is not as popular a source of data for research on food
consumption as the NFCS or the CEX; however, two of the earliest analyses of the effects of
food stamps on food expenditures are based on data from the PSID (Benus, Kmenta, and
Shapiro, 1976; and Hymans and Shapiro, 1976), as is one of the most recent of such studies
(Senauer and Young, 1986).
The PSID uses the recall methodology for measuring household expenditures on food at
home and food away from home. Nonrecipients of food stamps are asked to recall their average
weekly or monthly expenditures on food over the preceding year. Survey respondents who
received food stamps in the previous month are asked to recall their average weekly or monthly
purchases of food away from home, as well as their cash purchases of food at home. The survey
assumes that food stamp recipients spend the full amount of their monthly benefits on food at
home. Thus, the PSID's measure of total expenditures on food at home is obtained by adding
to the food stamp benefit amount the reported amount of cash purchases of food at home. This
methodology eliminates the possibility that purchases made with food stamps could be omitted
from the measure of expenditures on food at home; however, it overstates actual food
expenditure to the extent that food stamps are lost or hoarded by recipient households or are
traded for cash or nonfood items. As explained by Senauer and Young (1986), establishing a
15
floor on measured expenditures on food at home at the amount of the food stamp benefit
presents some statistical problems for analyses of food expenditures; however, analytic techniques
exist for dealing with those problems.
3. Issues Associated with Measuring and Analysing Food Consumption
This section examines several issues associated with measuring food consumption and
analyzing survey data on food consumption. Issues that are specific to the empirical study of food
consumption are addressed first, followed by an examination of a general issue associated with
analyzing data from complex sample surveys.
The Timeliness of Data. Because the Nationwide Food Consumption Survey uses three
different methodologies for measuring food consumption-a recall of usual household
expenditures on food, a recall of food used from the home food supply, and a combination
recall/diary of food intake by individuals-and because it obtains data from a large sample of low-income
households, the NFCS is the most frequently used source of data for analyzing the effects
of food stamps on food consumption. However, the survey's decennial schedule and the relatively
long lag (as much as two years or more) between the completion of the one-year data collection
process and the release of public-use data files mean that most NFCS-based research is conducted
with data that are three to ten years old. If a major change in FSP regulations occurs soon after
the completion of the survey, as was the case with the elimination of the purchase requirement
just one year after the completion of the 1977-78 NFCS, then program analysts may face the
unwelcome prospect of conducting research for the better part of a decade on the basis of data
that have only limited relevance to the current FSP.
The timeliness of data is far less of a problem with the CEX and the PSID because they
are on-going surveys in which sample units are interviewed at least once per year. Unfortunately,
neither of those surveys collects information on household food use or individual food intake.
16
Indeed, the complexity of collecting and processing such data is a major barrier to releasing NFCS
public-use files early and to fielding a survey like the NFCS more frequently.
The USDA has responded to the untimeliness of NFCS data by fielding NFCS
supplemental surveys on an "as-needed" basis and by fielding the Continuing Survey of Food
Intakes by Individuals in selected off-NFCS years. Neither of these solutions has proved to be
fully satisfactory. The most notable examples of NFCS supplemental surveys are the USDA
Survey of Food Consumption in Low-Income Households, 1979-80, and a 1983 survey of 2,400
low-income households in Puerto Rico. Both of these surveys were conducted shortly after the
implementation of major changes in the FSP: the EPR in 1979 and the replacement of the FSP
in Puerto Rico with the cash-based Nutrition Assistance Program in 1982. For reasons that are
not well-documented, researchers have been wary of the quality of the 1979-80 data, preferring
to use pre-EPR data from the 1977-78 low-income supplement to the NFCS. The limitations of
the 1984 Puerto Rico data pertain to its narrow geographic scope.
The limitations of the 1985 and 1986 editions of the CSFII pertain to its restricted sample-women
ages 19 to SO years and their children ages 1 to 5 years-and its focus on food intake by
individuals rather than on food use by households. As noted previously, the FSP is designed to
have its most direct impact on the use of food at home; its impact on the intake of all foods by
individuals is less direct, and may be diluted by the fact that some proportion of food eaten is
usually not derived from the home food supply and cannot be purchased with food stamps. Along
with other factors noted later, this diluted impact on individuals' intakes may explain the fact that,
as documented in Chapters IV and V of this report, researchers have consistently found
significant positive effects of the FSP on the use of food by households (whether measured in
dollar values or by nutrient content), but they rarely have found significant effects of the FSP on
food intakes by individuals.
17
According to current plans, new editions of the CSFQ will be fielded annually from 1989
through 1992. The samples in those editions of the survey will be defined more broadly than
those in the 1985 and 1986 editions. Separate samples of 1,500 households from all income strata
and 750 low-income households will be selected for each of the four survey years regardless of
their demographic characteristics. The substantive focus >f tfee survey will continue to be on food
intakes by individuals.
Underreporting of Food Expenditures by hSP Participants. The CEX Interview Survey,
the NFCS, and the CSFII obtain data on usual household expenditures on food at home through
similarly short sequences of questions that include a prompt for respondents to include purchases
made with food stamps in their reported food expenditures. Mathematica Policy Research used
a similar sequence of questions to obtain data on food expenditures from participants in the
SSI/Elderly Food Stamp Cashout Demonstration, which was conducted in 1980 and 1981 (Butler,
Ohls, and Posner, 1985). Tabulations of data from early interviews revealed markedly low
reported expenditures on food by coupon recipients relative to recipients of cash food assistance.
This finding led MPR to append to the sequence of questions on food expenditures a probe
which asked respondents whether their estimates of usual expenditures on food at home had
included purchases made with food stamps. In response to the probe, approximately 25 percent
of coupon recipients said that their estimates had not included purchases made with food stamps.
Such omissions could lead to significantly lower sample mean values of food expenditures by food
stamp recipients and to negatively biased estimates of the effects of food stamps on food
expenditures.
Fortunately, both the CEX and the NFCS provide alternative measures of food costs,
based on a recall of individual food items purchased (in the CEX Diary Survey) or a recall of
individual food items used (in the NFCS). Thus, researchers who use data from those sources
are not restricted to using the problematic measures of usual household food expenditures.
18
However, the CSFU provides no such alternative measure of food cost As explained previously,
the PSID addresses the omission of purchases made with food stamps from reported usual food
expenditures by assuming that recipient households use all of their stamps to buy food in the
month in which they receive the benefits, and by asking them to report only additional food
purchases made with cash.
Reference Periods for Expenditure and Income Data. The reliability of estimates of the
effects of food stamps that are generated by econometric models of food consumption is partly
a function of the degree to which the reference periods for the data on food consumption,
income, and food stamp benefits coincide. In this regard, the NFCS receives high marks relative
to the CEX and the PSID. The NFCS obtains income data for the calendar month that
immediately precedes the survey month. It also obtains data on the amount of food stamp
beneGts as of the most recent receipt of benefits, which for current recipients is either the survey
month or the preceding month. Moreover, the NFCS obtains household food-use data for the
week prior to the interview, individual intake data for three days including the day before, the
day of, and the day after the household interview, and data on usual food expenditures for the
three months preceding the household interview. Thus, the degree to which the reference
periods for NFCS income, food stamp, and food consumption data coincide is about the maximum
that is feasible with existing survey technology.
In the CEX Diary Survey, the degree to which the reference periods for the value of food
stamps received (the past month) and food purchases (the past two weeks) coincide is high, but
they diverge sharply from the reference period for household income, which is the previous 12
months. The situation is much the same for the PSID, which obtains data on the amount of food
stamps received and on food expenditures during the calendar month prior to the month of the
survey, but obtains household income data for the calendar year prior to the year of the survey.
If current income is a better predictor of current food consumption than is income received over
19
the course of the previous year, then food consumption models estimated on the basis of CEX
and PSED data may not produce valid estimates of the most relevant income-consumption
relationship. This in turn may cause the estimates of the effects of food stamps on food
consumption that are generated by those models to be biased.
Intra-individual Variation in Dietary Intake. In assessments of the adequacy of dietary
intake by individuals, the behavior of interest is the average, or "usual," dairy intake that would
persist over time. The actual dairy intake of food by individuals varies substantially, with intake
generally varying more within each person over time (intra-individual variation) than it does
among persons (inter-individual variation).10 The presence of intra-individual variation causes
the variance of average dairy intake in a sample of individuals to exceed the variance of usual
dairy intake in the population from which the sample was drawn. This discrepancy tends to be
largest when only one day of intake data is available for each sample member. The NFCS seeks
to reduce the overestimation of the population variance of usual daily intake by collecting three
days of intake data from each survey respondent
The positive bias in the sample variance of conventional survey measures of dietary intake
as an estimate of the population variance of usual dietary intake has important implications for
the validity of a number of dietary assessment techniques, as explained by the National Research
Council (1986). In the context of this review, the most important of those implications is that
the standard errors of estimates of the effects of food stamps on dietary intake are positively
biased when the estimates are based on a small number of days of intake data. That bias could
lead to the incorrect rejection of the hypothesis that the diets of food stamp recipients are of
higher nutritional quality than those of eligible nonrecipients. The fundamental problem is that
measures of average dairy intakes computed on the basis of only a few days of data incorporate
10The National Research Council (1986, Chapter 4) and Rittenbaugh et aL (1988, Chapter
III) review the literature on intra-individual and inter-individual variation in dietary intake.
20
substantial intra-individual variation, which amounts to random "noise" in the measurement of
usual intake, making it difficult to obtain statistically significant estimates of the effects of food
stamp*. This may partially explain why few studies have found statistically significant effects of
food stamps on dietary intakes, along with the fact that, as noted above, the direct effect of food
stamps on at-home consumption may tend to be "diluted" in measures of total food intake that
include away-from-home consumption. Findings from that body of research are reviewed in
Chapter V of this report
Complex Sample Designs. In large sample surveys such as the NFCS, the CEX, and the
PSID, the probabilities with which sample units are selected into the sample typically vary
somewhat. For example, low-income households that reside in high-poverty areas are selected
into the NFCS low-income sample with a higher probability than are low-income households that
reside in low-poverty areas. Sample units whose probabilities of selection are lower represent
more units in the target population of a survey than do sample units whose probabilities of
selection are higher. Those differences are reflected in the value of the sample weight for each
sample unit
When analyzing sample-weighted data, most researchers appropriately use the sample
weights to compute descriptive statistics such as sample means. Far fewer researchers use the
sample weights in multivariate analyses. As explained by DuMouchel and Duncan (1983), the
omission of the sample weights in a multivariate analysis may be appropriate if the outcome
variable is unrelated to the strata that form the basis for the sample selection probabilities, or if
the model fully controls for the effects of those strata. If neither of those conditions is satisfied,
then the sample weights should be used.
Most existing estimates of the effects of food stamps on food consumption are based on
data from complex surveys in which the probability of selection into the samples varies across the
sample units. Nevertheless, very few of those estimates have been generated on the basis of
21
sample-weighted data. In many studies, the decision to eschew using sample weights appears to
have been made without taking into account whether the conditions for omitting the weights from
a multivariate analysis were satisfied. Devaney and Fraker (1989) show that NFCS-based
multivariate estimates of the effects of food stamps on the money value of food used at home are
very sensitive to whether or not the sample weights are used in the estimation process.
Sample design effects are a second issue associated with analyzing data from complex
sample surveys. Standard multivariate regression procedures typically compute standard errors
for regression coefficients on the assumption that the samples were selected through simple
random sampling. However, because it is expensive, simple random sampling is rarely undertaken
in nationally representative surveys; clustered sampling is much more common. Standard errors
that are computed on the basis of clustered samples, under the assumption of simple random
sampling, tend to be underestimates. These underestimates can lead to larger t-statistics for
regression coefficients and, consequently, to the finding that estimates of program or other effects
are statistically significant when in fact those estimates are not The divergence between standard
errors computed on the assumption of simple random sampling and the true standard errors
computed on the assumption of clustered sampling reflects sample design effects.
Most of the empirical studies of the effects of food stamps on food consumption that are
reviewed in Chapters IV and V of this report are based on complex household surveys that have
clustered sample designs. Special regression packages that yield correct standard errors when
applied to data from clustered samples have existed for more than ten years (Shah, Holt, and
Folsom, 1977) and are widely available; however, there is no indication that these packages were
used to generate any of the empirical results that are reported in the studies reviewed in this
report
22
B. MEASURING FSP ELIGIBILITY AND PARTICIPATION
Errors in measuring food stamp participation and in modeling food stamp eligibility are
additional sources of potential bias in survey-based estimates of the effects of food stamps on
food consumption. Furthermore, the ability of researchers to eliminate yet another source of bias
in their estimates of the effects of food stamps-sample selection bias-is contingent upon
developing and estimating models of the decision to participate in the FSP that have good
explanatory power. The success of such modeling depends on the quality of the measures of
program participation and eligibility, as well as on the availability of variables that measure or are
correlated with the costs and benefits of participation in the FSP. This section explores the
availability and quality of these data elements in the data sets that have formed the basis for most
existing estimates of the effects of food stamps on food consumption.
1. Errors in Measuring FSP Participation and Benefits
A recent report issued by the U.S. Department of Commerce (1987) indicates that FSP
participation tends to be systematically underreported in household survey data. For example,
that report provides evidence that one-third of food stamp recipients interviewed by the Current
Population Survey fail to report receiving food stamps. Of course, the same households fail to
report the dollar value of their food stamp benefits. The existing evidence suggests that the
underreporting of food stamp participation is a common feature of household surveys. Thus,
there is reason to believe that FSP participation is underreported in the household surveys that
have provided the data for most of the existing estimates of the etlccts of food stamps on food
consumption.
As explained in Chapters IV and V, most estimates of the effects of food stamps on
household food expenditures or on the dollar value of food used are generated with regression
models in which the household food stamp benefit is a key explanatory variable. The models that
23
are used to generate estimates of the effects of food stamps on the nutrients that are available
in the food used by a household or on the nutrients that are provided by the food eaten by an
individual are more heterogeneous, but virtually all of them include among the explanatory
variables either an indicator of participation in the FSP or the dollar amount of the food stamp
benefit If participation or the benefit amount is underreported in the databases that are used
to estimate these models, then the models suffer from an "errors in variables" problem. Kmenta
(1986) shows that measurement error in an explanatory variable yields estimates of the regression
coefficient on that variable that are biased toward zero. Therefore, errors in measuring FSP
participation or the dollar value of the food stamp benefit would be expected to yield estimates
of the effects of food stamps on food consumption that are smaller than the true effects.
2. Errors in Modeling Food Stamp Eligibility
A household's eligibility to participate in the FSP is determined by its gross income, its net
income after certain deductions, its liquid asset holdings, and nonfinancial factors, such as
regulations that specify the individuals who are considered to comprise the household for the
purpose of determining its eligibility and benefit amount11 It is not possible to observe a
household's FSP eligibility status directly from survey data. However, it is usually possible to
model a survey respondent's eligibility status, which entails using the information obtained by the
survey to approximate what the outcome of a formal determination of eligibility would be. The
amount of information that general household surveys and surveys of food consumption obtain
11The following are allowable deductions from gross income for determining a household's
eligibility: a standard deduction that is invariant across all households, a deduction of 20 percent
of earned income, and deductions for qualified expenditures on shelter, dependent care, and (for
households with elderly or disabled persons only) medical care. Under the food stamp net
income screen, monthly gross income, net of allowable deductions, must be less than the federal
poverty guidelines. Households that do not contain elderly or disabled members must also have
gross incomes below 130 percent of the poverty guidelines. In addition, households must satisfy
a screen on liquid assets, which is set at $3,000 for households that consist of two or more
individuals (of whom at least one is elderly), and at $2,000 for all other households.
24
on the factors that determine food stamp eligibility differs greatly; thus, the degree of error in
modeling eligibility varies greatly across survey data sets.
Selected waves of the PSID provide data on most of the factors that are considered in a
formal determination of a household's eligibility to receive food stamps. However, those data are
provided on an annual basis, whereas a formal determination of food stamp eligibility is made on
the basis of monthly income and expenses. Researchers have used PSID data to model FSP
eligibility (Coe, 1983), but modeling eligibility with annual data can lead to misclassifying
households that have experienced recent changes in income, expenses, and household
composition. Given the usual patterns of change in household income, the most frequent error
associated with modeling eligibility with annual data is misclassifying currently eligible households
as ineligible. Of more importance is the absence of data on liquid asset balances in the PSID and
the consequent necessity of imputing those balances on the basis of reported asset income. That
process tends to generate underestimates of asset balances and to lead to classifying some asset-ineligible
households as eligible to receive food stamps.
The CEX Diary Survey, like the PSID, provides data on annual income, including income
from assets, but it does not provide data on asset balances. The data that it provides on
deductible expenses are more limited than those provided by the PSID; consequently, researchers
have avoided using the CEX data on deductible expenses to model net income eligibility for food
stamps. Instead, they approximate net income on the basis of simple rule-of-thumb assumptions
about the relationship between deductions and gross income. For example, West (1984) assumes
that deductions equal 23 percent of gross income.12
12Other examples of rule-of-thumb assumptions that have been used to estimate the
deductible expenses of respondents to the CEX Diary Survey are provided by Salathe (1980) and
Chavas and Yeung (1982).
25
Unlike the PSID and the CEX, the NFCS low-income supplemental surveys are targeted
toward households that might be eligible to receive food stamps. The full NFCS survey
instrument is administered only to those sample households that, on the basis of data provided
during a short screening interview, are estimated to be eligible to receive food stamps. In 1977-
78, the screening instrument obtained data on income and deductible expenses during the
previous month and on liquid asset balances. Households were screened into the low-income
sample if their gross and net incomes and their liquid asset balances were less than the FSP
eligibility limits. In 1987-88, the screening instrument obtained data only on income during the
previous month. Households were screened into the low-income sample if their reported income
was less than the food stamp gross income limit The NFCS screening procedures for both 1977-
78 and 1987-88 represent rough approximations to the food stamp eligibility criteria. The absence
of screens on liquid asset balances and net income in the 1987-88 survey suggests that the low-income
sample for that survey may include more FSP-ineligibles than does the low-income sample
for the 1977-78 survey.
In analyses of the effects of the FSP on food consumption, an analysis sample that consists
of FSP eligibles serves two purposes. First, the homogeneity of a sample of FSP-eligibles reduces
the risk of obtaining biased estimates of the effects of food stamps if the model of food
consumption is not as well-specified as one would like. For example, if an analysis sample
included some high-income households, then failing to specify the correct functional form of the
relationship between income and food consumption might generate highly biased estimates of the
effects of food stamps. We would expect that the bias would be smaller with a more
homogeneous sample. Second, a sample of eligibles will support estimating a model of
participation in the FSP. As explained in the following section, the estimation of a participation
model is a critical component of an econometric procedure that generates estimates of the effects
of food stamps on food consumption that are free of sample selection bias.
26
3. Data Requirements for Modeling FSP Participation
Multivariate regression models are used to obtain estimates of the effects of food stamps
on food consumption while controlling for observed differences between food stamp participants
and eligible nonparticipants that may also influence food consumption, such as income and
household size. In the past decade, researchers have become aware that most survey databases
do not provide data on all of the important respects in which participants may differ from eligible
nonparticipants (e.g., a knowledge of nutritional requirements). If those unobserved differences
influence food consumption, then they may bias regression estimates of the effects of the FSP.
This bias is referred to as "sample selection bias."
The econometric solution to the problem of sample selection bias is to estimate a model
ofFSP participation with a sample of eligible households and then to compare the actual program
participation of the sample cases with the model's predictions of their probabilities of
participation. Actual participation is an outcome of the influence of both observed and
unobserved factors, whereas *he predicted probability of participation is a function of observed
factors only, thus, the difference between the two reflects (and is a measure of) the influence of
the unobserved factors.
In his pathbreaking articles on selection bias, Heckman develops a methodology for
incorporating the information on unobserved factors from the participation analysis into a
synthetic variable that can then be included in the food consumption equation (see Heckman,
1978 and 1979; and Heckman and Robb, 1985). By controlling for the influence of those
unobserved factors on food consumption, the synthetic variable may eliminate sample selection
bias from the regression estimate of the effect of food stamps on food consumption.13
13Formally, when applied under appropriate conditions, Heckman's methodology is a
consistent estimator of program effects (i.e., it is biased for small samples, but the bias disappears
as the sample size increases).
27
A number of researchers have used Heckman's procedure to control for selection bias in
their estimates of the effects of food stamps on food consumption. They include Chen (1983),
Aiken et al. (1985), Devaney, Haines, and Moffitt (1989), and Fraker, Long, and Post (1990).
However, to ensure that the procedure is fully effective at eliminating selection bias, the program
participation model must include some significant predictors of participation, and at least one of
those predictors must be a variable that is not also a significant predictor of food consumption.
Examples of such variables are the following measures of the cost of participating in the FSP:
(1) the mode in which food stamps are issued in a household's home county (e.g., over-the-counter
or by mail); (2) the time and monetary cost of traveling to the local food stamp office
for over-the-counter issuances; and (3) the psychological costs of participating in the FSP (i.e.,
stigma). These and similar variables are not generally available in survey databases that pro ie
data on food consumption. In their absence, it may be technically feasible to implement
Heckman's procedure, but one cannot be confident that it appreciably reduces the problem of
sample selection bias.
28
m. THE CONSUMPTION PATTERNS OF FOOD STAMP RECIPIENTS
AND LOW-INCOME NONRECIPIENTS
This chapter reviews findings from descriptive studies of the expenditure shares and food
consumption patterns of food stamp recipients and low-income nonrecipients.14 Some of the
recipient-nonrecipient differences that are presented herein are attributable to differences in
income, household size, and other characteristics, rather than to the effects of food stamps.
Subsequent chapters review findings from studies that have attempted to disentangle the effects
of the food stamps on consumption from the effects of household and individual characteristics.
A HOUSEHOLD EXPENDITURE PATTERNS
1. Expenditure Shares
Using data from the interview component of the 1982-83 Consumer Expenditure
Survey,15 Boldin and Burghardt (1989) find that expenditures on all food items (food used at
home, as well as food purchased and used away from home) account for 28.7 percent and 22.5
percent of the total expenditures of, respectively, food stamp recipient households and low-income
nonrecipient households. They do not indicate whether that difference is statistically
significant; however, they do note that the actual difference in food expenditure shares between
these two groups may be larger than is indicated by those percentages, because it is likely that
MA11 comparisons between food stamp recipients and nonrecipients in this chapter are made
between recipient households (or individuals in those households) and low-income nonrecipient
households (or individuals in those households).
15As the principal source of data on U.S. households' expenditures on all consumer goods and
services, the CEX Interview Survey has provided the basis for most recent analyses of the total
expenditure patterns of food stamp households. The other component of the CEX, the Diary
Survey, provided data for several early studies of food consumption patterns of food stamp
households, although the Nationwide Food Consumption Survey is now the most widely used
source of data on food consumption. See Chapter II for further description of the CEX and the
NFCS.
29
some food stamp recipients omitted food purchases made with food stamps from the expenditure
amounts that they reported in the CEX.
For 27 of 36 expenditure categories, encompassing both food and nonfood items, Brown
(1988) reports that the mean expenditure shares of food stamp recipients differ from those of
low-income nonrecipients at the .01 level of statistical significance. With some aggregation across
expenditure categories, Figure ELI summarizes the results of Brown's analysis of data from the
interview component of the 1984-85 CEX. Most notable among his results is the finding that
food stamp households have significantly larger expenditure shares for food used at home and for
total food than do low-income households that do not receive food stamps; however, that
relationship is the converse for food bought and consumed away from home.
2. The Money Value of Food Used
Households that participate in the FSP allocate a larger percentage of their total
expenditures to the purchase of food than do low-income nonparticipating households, but the
money value of all food used by recipients is less than that of nonrecipients. Based on its analysis
of data from the 1979-80 low-income supplement to the 1977-78 NFCS, the Human Nutrition
Information Service (July 1982) reports that the average participating household uses food worth
$52.97 per week, whereas the average nonparticipating household uses food worth $59.96 per
week (see Figure DX2). Food purchased and used away from home accounts for $5 of the
difference, while food used at home accounts for $2 of the difference.
When adjustment is made for the larger average size of nonparticipating households, the
average money value of food used by food stamp recipients and nonrecipients converges. Figure
JJL3 displays the finding by HNIS (Jury 1982) that the money value of food used at home per
household member is slightly higher for food stamp recipient households than for nonrecipient
households. However, recipient households spend only about half as much per member on food
30
FIGURE 111.1
HOUSEHOLD EXPENDITURE SHARES BY MAJOR EXPENDITURE CATEGORY
(Source: 1984-85 Consumer Expenditure Survey, Interview component)
Food away from homo 1.6X
Olnof OKponoM
Apparel 5.2*
Transportation
Rtcrwation SX
Medfeol cant 3.7X
FSP Participant8
Apparel 4.7%
Tfonsportooon
Food away from homo 4.6X
Food at homo
Other cxpcnoM
RoenMtfon 3.9%
Modfeal ear* 6.5%
FSP Nonparticipants
31
70-
60- ■
50- ■
40- •
tsS
oo
O
30- -
20- ■
10- -
FIGURE III.2
MONEY VALUE OF FOOD USED IN A WEEK BY HOUSEHOLDS
(Source: USOA Survey of Food Consumption in Low-Income Households. 1979-80)
$59.96
$52.97
$4.71
$48.26
FSP Participants
$9.62
$50.34
LEGEND
Food bought &c used away from home
Food used at home
FSP Nonparticipants
\z
20
18- ■
16-
14- ■
12- -
a
8
I 10
8- ■
6- ■
4- •
2- ■
FIGURE 111.3
MONEY VALUE OF FOOD USED PER PERSON IN A WEEK BY HOUSEHOLDS
(Source: USDA Survay of Food Consumption in Low-Income Households, 1979-80)
$16.61
$17.20
$148
$15.13
FSP Participants
$2.76
$14.44
FSP Nonparticipants
3^—
LEGEND
Food bought Ac used away from home
Food used at home
bought and used away from home at do nonrccipicnt households. These pa.tially offsetting
differences mean that the gap in the total money value of food used per person between the two
groups is small.
Devaney and Kislcer (1988) use a more sophisticated adjusted measure of the average value
of food used at home per household member. They measure household size in "equivalent
nutrition units" (ENUs), which is the number of adult-male-equivalent persons eating meals from
the home food supply. This measure of household size controls for the number of persons in the
household and their age-and-sex-based differences in nutritional requirements, for the proportion
of meals eaten away from home by household members, and for meals served to guests.16 Using
the same data set as HNIS, Devaney and Kislcer find that the average money value of food used
at home per ENU is 11 percent higher for food stamp households than for low-income
households that do not receive food stamps. This figure contrasts with the 5 percent difference
obtained by HNIS on the basis of its simpler per-person measure of home food use.
3. ^utrjents per DoUaj's Worth of food,
Among all households, those with larger money values of food used at home per person
(measured in ENUs) obtain fewer nutrients for each dollar's worth of food used than do
households with smaller money values of food used (Peterlrin and Hama, 1983; and Morgan et
al, 1985b). This implies that households with limited food budgets tend to use foods that are
relatively high in nutrients and low in cost
Among low-income households, food stamp recipients have a higher average money value
of food used at home per person than nonrecipients, as documented in the previous section.
Nevertheless, the nutrient efficiency of the home food dollar is not generally lower for recipients
16Appendix B describes the computation of household size in ENUs and compares that
measure of size with several alternative measures.
34
than for nonrecipients. On the basis of a simple comparison of mean values between food stamp
recipients and nonrecipients, Peterkin and Hama (1983) report that recipients obtain more
nutrients per dollar's worth of food used at home for nine nutrients and less for only two. Using
regression analysis to control for the effects of a number of socio-economic factors, Morgan et
al. (1985b) find that food stamp recipients, relative to nonrecipients, have a higher availability per
dollar's worth of food used at home of food energy, protein, calcium, iron, and magnesium, but
a lower availability of vitamin A. The recipient-nonrecipient difference is statistically significant
only for calcium.
Thus the existing evidence indicates (albeit with limited statistical reliability) that food
stamp recipients have a higher average money value of food used at home per person than low-income
nonrecipients and they also receive more nutrients for each dollar's worth of food used
at home.
4. Home Food Use bv Food Group
As summarized in Figure III.4, HNIS (Jury 1982) finds that food stamp recipients and
nonrecipients have very similar patterns of home-food use when food groups are defined at a high,
level of aggregation (i.e., seven groups). The most notable difference is that food stamp
recipients allocate a larger percentage of the average home-food dollar to meat, poultry, and fish
than do nonrecipients. Conversely, nonrecipients spend a somewhat larger percentage of their
average home-food dollar on grain products and on fruits and vegetables than do food stamp
recipients. It is not known whether these differences are statistically significant
A study based on data from the Low-Income Supplement to the 1977-78 NFCS provides
additional insight into the recipient-nonrecipient difference in the share of home-food
expenditures allocated to meat, poultry, and fish Morgan et al. (1985a) report that most of this
difference is due to greater expenditure shares by recipients on fish, poultry, and lower-cost
35
FIGURE 111.-4-
SHARE OF HOME FOOD EXPENDITURES BY FOOD GROUP
(Source: USDA Survey of Food Consumption in Low-Income Households, 1979-80)
Meat, poultry, fish
Eggs, legumes, nuts 5.2%
Fruits, vegetables
Other
Fats, sugars 5.7%
Grain products
Milk products
FSP Participants
Meat, poultry, fish
Eggs, legumes, nuts 5.6X
Fruits, vegetables
Groin products
Fats, sugars 5.7%
Milk products
FSP Nonparticipants
36
meats. Recipients have slightly lower expenditure shares on higher-cost meats than do eligible
nonrecipients.
5. Frequency of Food iMJM
One dramatic difference in expenditure behavior between food stamp recipients and low-income
nonrecipients pertains to the frequency of their major food shopping. As shown in Figure
IH.5, HNIS (July 1982) reports that recipient households are far more likely than nonrecipients
to conduct their major food shopping on a monthly basis, presumably timed to coincide with their
monthly food stamp allotment Most nonrecipients conduct their major food shopping on a
weekly basis. Data from an ongoing demonstration project in Reading, Pennsylvania, in which
an "Electronic Benefit Transfer" (EBT) system is being used to issue food stamp benefits (plastic
cards in place of coupons), show that recipients spend an average of 19 percent of their monthly
benefit on the day of issuance, 70 percent within the first week, and 89 percent within two
weeks.17
The apparent sensitivity of the frequency of major food shopping to food stamp receipt
suggests that the quantity and/or quality of food used by food stamp households may also follow
a monthly cycle. Despite the fact that it may enhance our understanding of why econometric
studies show that food stamps have a much larger effect on food use than does cash income,
research on the existence and nature of this cycle has been scarce.
6. Perceived Food Adequacy
Gear majorities of both food stamp recipient households and nonrecipient households
report having adequate supplies of food. However, as shown in Figure HL6, HNIS (July 1982)
17These findings will be reported in a forthcoming FNS report entitled "Household Shopping
Patterns in the Food Stamp Electronic-Benefit-Transfer Demonstration."
37
FIGURE 111.5
FREQUENCY OF MAJOR FOOD SHOPPING BY HOUSEHOLDS
(Source: USDA Survey of Food Consumption In Low-Income Households, 1979-80)
Every other week
More than weekly
FSP Participants
More than weekly
Monthly
Every other week
FSP Nonparticipants
38
FIGURE III.6
SELF-EVALUATION OF HOUSEHOLD FOOD ADEQUACY
(Source: USDA Survey of Food Consumption In Low-Income Households, 1979-80)
Enough, not kind wonted
Enough, kind wanted
Often not enough 4%
Sometimes not enough
FSP Participants
Enough, kind wanted
Enough, not kind wanted
Often not enough 2X
Sometimes not enough
FSP Nonparticiponts
39
finds that 24 percent of food stamp recipient households, compared with only 8 percent of
nonrecipient households, report that they sometimes or often have inadequate supplies of food.
Basiotis (1987) uses data from the 1977-78 NFCS to investigate whether the expenditures
on food and the use of food energy by low-income households that report having inadequate food
supplies sometimes or often are more responsive to changes in income (i.e., are more income
elastic) than is the case for other low-income households. His estimates of the income elasticities
of food expenditure and food energy usage are significantly larger for households that report
inadequate food supplies than for other households. The larger elasticities are consistent with
more aggressive efforts to economize on food usage in response to reductions in income. This
correlation between objective measures of food economizing behavior in response to income
reductions and survey respondents' perceptions of the adequacy of their home food supplies
substantiate the validity of self-reported measures of food adequacy by low-income households.
B. THE NUTRIENT AVAILABILITY OF HOUSEHOLDS AND THE NUTRIENT
INTAKE OF INDIVIDUALS
As described in Chapter n, the NFCS provides data on the nutrient availability of
households that are based on the quantity of each food item used by a household from its home
food supply over a one-week period. An existing USDA nutrient database is used to convert the
survey data on the quantity of each food item into data on the nutrients provided by that item.
The availability of a specific nutrient is the sum of the units of that nutrient provided by all foods
used from the home food supply during the reporting period.
As also described in Chapter II, the NFCS data on individual nutrient intake are computed
on the basis of the reported types and quantities of foods eaten either at home or away from
home by the individual members of a household. The sum of total nutrient intakes over all
members of a household may differ from the availability of nutrients in the home food supply for
four reasons: (1) some food that is used by a household is lost or wasted rather than eaten; (2)
40
nutrients provided by foods purchased and eaten away from home are included in the NFCS
measure of total nutrient intake but not in the measure of at-home nutrient availability, although
the latter measure is often adjusted for meals eaten away for home to obtain a proxy measure
of total nutrient availability; (3) food served to guests or boarders or fed to pets may account for
some of the food used from the home food supply but not actually consumed by household
members; and (4) either or both of the NFCS food-consumption data sets may contain some
degree of measurement error (e.g., a tendency to underreport individual intake). Even after
making rule-of-thumb adjustments for the first three of these reasons, the nutrient availability of
the household tends to exceed the sum of the nutrient intake by all household members (Batcher,
1983). The residual difference in the two nutrient measures is attributable to imprecision in the
adjustments and to measurement error.
1. Nutrient Availability
Controlling for guest meals, meals away from home, and the age-sex composition of
household members, HNIS (July 1982) computes the availability of nutrients in the household
as a percentage of the combined household members' recommended dietary allowances (RDAs).
The results of its analysis of data from the USDA Survey of Food Consumption in Low-Income
Households, 1979-80, are reproduced in Table III.l, which shows that the average availability of
each of twelve selected nutrients exceeds the RDA for both food stamp recipient households and
nonrecipient households, and for all of the nutrients their availability relative to the RDA is
higher for food stamp recipients than for nonrecipients. It should be noted, however, that even
though the availability of a nutrient relative to the RDA may be high on average within a
population group, the availability of the nutrient may be less than is adequate to meet the dietary
requirements of some proportion of households in the group. Furthermore, even within a
household for which the availability of a nutrient is, in principle, adequate, the average intake of
41
TABLE 111.1
HOUSEHOLD NUTRIENT AVAILABILITY AS A PERCENTAGE OF THE RDA
FOR PERSONS EATING IN THE HOUSEHOLDS
(Source: USDA Survey of Food Consumption in Low-Income Households, 1979-80)
FSP Participants FSP Nonparticiponts Difference
Nutrient (A) (B) (A-B)
Food energy 139% 121% +18%
Protein 232 203 +29
Calcium 119 111 +8
Iron 151 137 +14
Magnesium 134 123 +11
Phosphorus 202 183 +19
Vitamin A 213 178 +35
Thiamin 194 185 +29
Riboflavin 204 180 +24
Vitamin B6 132 114 +18
Vitamin B12 235 191 +44
Vitamin C 290 264 +26
NOTE: The table shows mean nutrient availability per equivalent nutrition unit as a percentage of
the RDA. As explained in Appendix B, household size in ENUs is a measure of size that
adjusts for the age and sex composition of household members, the number of meals per
week that they eat from the household food supply, and meals served to guests.
42
the nutrient by household members may be inadequate due to waste or other food loss; or even
when the average intake by household members is adequate, specific individuals within the
household may have an inadequate intake of the nutrient due to the pattern of food allocation
within the household.
2. Nutrient Intake
Because measures of nutrient availability include nutrients provided by food that has been
wasted or lost, we expect that (after adjustments are made for guest meals, meals away from
home, and age-sex composition) they will indicate the possibility of more nutrients in the diets
of the low-income population than do measures of nutrient intake. Table ni.2 presents the
findings of HNIS (September 1982) on the intake of nutrients by individuals of all ages and sexes
in low-income households, based on one day of data from the USDA Survey of Food
Consumption in Low-Income Households, 1979-80. The table shows that the mean intake of four
and five of the twelve selected nutrients for, respectively, food stamp nonrecipients and food
stamp recipients is less than the RDA. For none of the nutrients is the intake by food stamp
recipients substantially less than that by nonrecipients, and for three of the nutrients it exceeds
the intake by nonrecipients by more than 10 percentage points (relative to the RDA).
On the basis of four days of data from the 1986 Continuing Survey of Food Intakes by
Individuals, HNIS (1989) reports that the average intake of ten of twelve selected nutrients by
women ages 19 to SO in food stamp households is slightly lower than that by women in low-income
households that do not receive food stamps. Those findings are summarized in Table
III.3, along with findings for children ages 1 to 5. The results for children are quite different
from those for women. HNIS finds that young children in food stamp households have a higher
average intake of nine of the twelve selected nutrients than do young children in low-income
43
TABLE 111.2
NUTRIENT INTAKE AS A PERCENTAGE OF THE RDA:
MEAN PER INDIVIDUAL, ONE DAY OF INTAKE DATA
(Source: USDA Survey of Food Consumption in Low-Income Households, 1979-80)
FSP Participants FSP Nonparticipants Difference
Nutrient (A) (B) (A-B)
Food energy 85% 83% +2%
Protein 172 168 +4
Calcium 87 90 -3
Iron 96 100 -4
Magnesium 85 88 -3
Phosphorus 130 132 -2
Vitamin A 132 118 +14
Thiamin 130 113 +17
Riboflavin 141 132 +9
Vitamin B6 79 72 +7
Vitamin B12 142 143 -1
Vitamin C 144 133 +11
44
TABLE III.3
NUTRIENT INTAKE AS A PERCENTAGE OF THE RDA:
MEAN PER INDIVIDUAL. FOUR NONCONSECUTIVE DAYS OF INTAKE DATA
(Source: NFCS-Continuing Survey of Food Intakes by Individuals. 1986)
Worn en Ages 19 o 50 Children Ages 1 to 5
FSP FSP Non- FSP FSP Non-
Participants participants Difference Participants participants Difference
Nutrient (A) (B) (A-B) (C) (D) (C-D)
Food energy 68% 71% -3% 103% 97% +6%
Protein 126 130 -4 234 214 +20
Calcium 67 72 -5 105 105 0
Iron 53 55 -2 88 80 +8
Magnesium 59 62 -3 119 110 +9
Phosphorus 109 113 -4 135 127 +8
Vitamin A 99 109 -10 188 204 -16
Thiamin 100 100 0 162 146 + 16
Riboflavin 101 106 -5 202 195 +7
Vitamin B6 52 55 -3 128 119 +9
Vitamin B12 149 143 +6 21 1 210 + 1
Vitamin C 109 112 -3 182 183 -1
w
households that do not receive food stamps. Most of those differences exceed 5 percentage
points.
Tables D3.2 and TTT.3 provide only a partial picture of the differences between food stamp
recipients and nonrecipients in nutrient intake. However, they suggest that small aggregate
differences may mask important distinctions across demographic groups. For example, among
young children, FSP recipients have substantially larger average intakes of most nutrients than
do nonrecipients; among women ages 19 to SO, the intake of most nutrients by food stamp
recipients and nonrecipients differs very little.
C INDIVIDUAL FOOD INTAKE BY FOOD GROUP
On the basis of data from the USDA Survey of Food Consumption in Low-Income
Households, 1979-80, HNIS (September 1982) reports that the average intake of six of seven
selected food groups by individuals in food stamp households is smaller than that by individuals
in low-income households that do not receive food stamps. These findings, which are based on
one day of intake data, are presented graphically in Figure IIL7. Intake by recipients relative to
intake by nonrecipients ranges from -25 percent for eggs, legumes, and nuts to +9 percent for
grain products.18
Figures III.8 and III.9 summarize findings by HNIS (1989) about the patterns of the food
intake of women ages 19 to SO and children ages 1 to 5 by food group. The results are based on
four days of intake data from the 1986 CSFII. For women, the findings on differences in food
intake by food group between food stamp recipients and nonrecipients are broadly consistent with
l8It should be noted that Figure DX7 shows grams of food intake bv individuals, while Figure
ni.4 shows shares of home food expenditure, bv households. Thus, the data in the two figures
are not directly comparable; for example, FSP nonparticipating households spend a larger share
of their home food expenditures on grain products than do FSP participating households, but the
average intake of grain products by individuals in nonparticipating households is less than the
average intake of grain products by individuals in participating households.
46
(Source:
FIGURE 111.7
FOOD INTAKE BY FOOD GROUP:
MEAN PER INDIVIDUAL PER DAY, ONE DAY OF DATA
USDA Survey of Food Consumption in Low-Income Households. 1979-80)
Meat, poultry, fish
Eggs, legumes, nuts
Fruits, vegetables
Grain products
Milk products
Fats, sugars
Beverages
LEGEND
FSP Participants
FSP Nonparticipants
100 200 300 400
Grams
500 600 700
V-7
FIGURE III.8
FOOD INTAKE BY WOMEN AGES 19 TO 50 BY FOOD GROUP:
MEAN PER INDIVIDUAL PER DAY, FOUR NONCONSECUTIVE DAYS OF DATA
(Source: NFCS-Continuing Survey of Food Intakes by Individuals, 1986)
oo
Meat, poultry, fish
Eggs, legumes, nuts
Fruits, vsgstables
Grain products
Milk products
Fats, sugars
Beverages
LEGEND
FSP Participants
FSP Nonparticipants
w
FIGURE III.9
FOOD INTAKE BY CHILDREN AGES 1 TO 5 BY FOOD GROUP:
MEAN PER INDIVIDUAL PER DAY. FOUR NONCONSECUTIVE DAYS OF DATA
(Source: NFCS-Continuing Survey of Food Intakes by Individuals. 1986)
vc-
Meat, poultry, fish
Eggs, legumes, nute
Fruits, vegetables
Grain products
Milk products
Fats, sugars
Beverages
LEGEND
2 FSP Participants
FSP Nonparticipants
600
c/g
the previously discussed findings on differences in nutrient intake among all individuals in low-income
households. For three of the seven food groups (meat, poultry, and fish; eggs, legumes,
and nuts; and grain products), the average intake by women in participating households is virtually
the same as that by women in nonparticipating households. The average intake of foods from
the other four groups by women in participating households is between 4 percent and 14 percent
less than that by women in nonparticipating households. Young children in participating
households have a substantially greater average intake of foods from three of the seven food
groups (meat, poultry, and fish; eggs, legumes, and nuts; and grain products), a substantially
smaller average intake of fruits and vegetables, and a similar average intake of foods from the
remaining three groups, compared with young children in nonparticipating households.
D. SUMMARY
The following are selected findings on the differences between FSP recipients and low-income
nonrecipients from the studies discussed in detail in this chapter.
• Food Expenditures. Data from the interview component of the 1984-85
Consumer Expenditure Survey show that FSP participating households
spend a larger portion of their total expenditures on all food items than
do nonparticipating households; however, nonrecipients have larger
expenditure shares for food bought and consumed away from home.
The USDA Survey of Food Consumption by Low-Income Households,
1979-80, was the basis for a study which showed that, although the
average money value of food used at home per household member is
greater for food stamp recipients, nonrecipients spend about twice as
much per household member on food bought and used away from home,
thus causing the gap between the two groups in total money value of
food used per household member to be small Another study based on
the same data set showed that, when household size is measured on the
basis of "equivalent nutrition units" (a measure of household size that
controls for the age-and-sex-based differences in nutritional requirements
of household members, meals eaten away from home, and meals served
to guests), the average money value of food used at home per ENU is
11 percent higher for participating than for nonparticipating households.
50
• Nutrients per Dollar's Worth of Food. Relative to low-income
nonrecipients, food stamp recipients obtain more of most nutrients per
dollar's worth of food used at home. Thus, the food used at home by
recipients has a greater money value per person and provides more
nutrients per dollar than does the food used by nonrecipients.
• Home Food Use bv Food Group. A study based on the 1979-80 USDA
data also showed that, overall, home-food use patterns are similar for
participating and nonparticipating households when measured in terms
of the share of home food expenditures allocated to each of seven food
groups. However, recipients spend a larger percentage of the average
home-food dollar on meat, poultry, and fish than do nonrecipients, while
nonrecipients spend a larger percentage on grain products and on fruits
and vegetables.
• Frequency of Food Shopping. The same study showed that food stamp
recipients are far more likely than nonrecipients to conduct their major
food shopping on a monthly basis, while most nonrecipients shop for
food on a weekly basis.
• Perceived Food Adequacy. Another finding of that study was that,
although the majority of both food stamp recipients and nonrecipients
report having adequate supplies of food, more recipients than
nonrecipients report sometimes or often not having adequate supplies of
food (24 percent of recipients, compared with 8 percent of
nonrecipients).
• Nutrient Availability. According to the 1979-80 USDAdata, the average
availability of each of twelve selected nutrients exceeds the RDA for
both participating and nonparticipating households; for all of the
nutrients, availability relative to the RDA is higher for recipients than for
nonrecipients.
• Nutrient Intake. The 1979-80 USDA data also show that the mean
nutrient intake by individuals in low-income hc;v&holds is less than the
RDA for four of twelve selected nutrients for nonrecipients and for five
of the twelve nutrients for recipients.
Another finding from the 1979-80 USDA data is that the intake of three
of the twelve selected nutrients by recipients exceeds that by
nonrecipients by more than 10 percentage points (relative to the RDA),
and for none of the nutrients is its intake by recipients substantially less
than its intake by nonrecipients.
Data from the 1986 Continuing Survey of Food Intakes by Individuals
reveal differences in nutrient intake among demographic groups; for
example, among young children, food stamp recipients have higher
average intakes of most nutrients than do nonrecipients, whereas among
51
women ages 19 to SO, food stamp recipients have slightly lower intakes
of ten of the twelve nutrients than do nonrecipients.
Individual Food Intake bv Food Group. The 1979-80 USDA data show
that the average intake of foods from six of seven selected food groups
by individuals in households that participate in the FSP is smaller than
that by individuals in nonrecipient households.
Data from the 1986 Continuing Survey of Food Intakes by Individuals
show that young children in participating households have a greater
average intake of foods from three of the seven food groups, a roughly
equal average intake from three food groups, and a smaller average
intake from two food groups than young children in nonparticipating
households. Average food intake for women in participating households
is similar to that of women in nonparticipating households for three of
the food groups, and smaller than that of women in nonparticipating
households for four food groups.
52
IV. THE EFFECTS OF FOOD STAMPS ON FOOD EXPENDITURES
The economic theory of household consumption behavior predicts that food stamp benefits
will tend to increase both the food and nonfood expenditures of recipients. A particular
application of general consumer theory, developed by Southworth (1945), further predicts that
food stamp benefits may lead to greater spending on food than does an equal amount of cash
income. That is, due to the coupon form of the benefit, some recipient households may be
"constrained" to spend more on food than they would actually prefer given their level of total
resources. If a large proportion of participating households are "constrained," then the model
provides a basis for asserting that food stamps exert an overall effect on the food spending of
participants that is greater than that of an equivalent cash transfer. Conversely, if only a small
proportion of participants are "constrained," then the model predicts that the overall effect of
food stamps on food spending is only slightly greater than the overall effect of an equivalent cash
transfer. Appendix A provides a more detailed review of Southworth's model
Since the total desired food spending of constrained households is less than the amount
of their food stamp benefit, they will make few if any cash purchases of food for home use.
Empirically, only a small proportion of participating households (perhaps 10 to 15 percent) report
little or no cash food purchases and thus may be "constrained" as defined in the context of the
Southworth model.19 In this circumstance, the model's prediction is clear-cut: the overall effect
of food stamps on food expenditures will be very similar to the overall effect of regular income,
exceeding the latter only by a small margin at most.
19On the basis of data from the 1979 wave of the PSID, Senauer and Young (1986) estimate
that 14 percent of food stamp recipients make no cash purchases of food and, thus, in the
framework of Southworth's model, are constrained in their consumption behavior. See Appendix
A for additional empirical findings on this topic
53
Since the early 1970s, a large number of empirical studies have estimated the effects of
food stamps and of regular income on the food spending of participating households. The
Southworth model has been cited frequently in this literature. However, in nearly every case, the
empirical findings on the effects of food stamps have failed to confirm the model's central
prediction that the food-expenditure effects of food stamps and regular cash income are
approximately equal. Rather, these studies have consistently found a substantially greater
marginal effect on food spending from food stamps than from regular income.20
Consequently, it is fair to say that no currently existing theoretical model explains the much
greater observed effect of food stamps than of regular income on the food spending of
participating households. This anomalous situation might be explained within the framework of
the Southworth model if:
the consumption behavior of many more participating households is in
fact "constrained" than appears on the basis of monthly or annual data
(e.g., for certain periods of each month); or
the empirical findings are consistently misleading due to strong,
undetected self-selection bias that leads to spuriously high estimates for
food spending out of program benefits, or due to other errors in how the
empirical model is specified or how the analytic variables are measured.
Modern developments in consumption theory may provide several potentially fruitful avenues for
addressing this problem, since they incorporate additional dimensions of consumption behavior
(e.g., the household production-function approach); but such developments have not yet been
explored in depth in the food stamp literature.
*°In the 17 studies reviewed in the remainder of this chapter, the estimated marginal effect
on the food spending of participating households from a given increase in food stamps exceeds
the estimated effect of a comparable increase in regular income by approximately 2 to 10 times
(excluding the two most extreme values). The median value is a 3.8 times greater marginal effect
from food stamps than from cash income.
54
The previous chapter reviewed a number of descriptive findings on the food expenditures
of food stamp recipients and eligible nonrecipients. Those findings fail to show conclusively
whether food stamps are effective at increasing food expenditures, and they do not address
whether food stamps are more effective than cash assistance at increasing food expenditures.
This chapter reports on the application of econometric models for analyzing the effectiveness of
food stamps at increasing food expenditures, both absolutely and relative to cash assistance.
A. A FRAMEWORK FOR ESTIMATING THE EFFECTS OF FOOD STAMPS
ON FOOD EXPENDITURES
1. Research Strategies
Descriptive studies of the food consumption of food stamp recipients and low-income
nonrecipients, such as those reviewed in the previous chapter, generally do not provide a reliable
indication of the actual effect of food stamps on food consumption because they do not control
for nonprogrammatic differences between food stamp recipients and eligible nonrecipients.
Differences in income and other observable characteristics, as well as differences in unobservable
characteristics, such as an awareness of nutritional requirements, may influence food consumption
in ways that exaggerate or mask differences that are attributable to the form and amount of the
food stamp benefit Two alternative research strategies are available to control for the effects
of nonprogrammatic differences:
Qassical experimentation may entail assigning food stamp assistance
randomly to some eligible households but not to others, or assigning food
coupons randomly to some participating households and cash assistance
to others. In either case, the random assignment of benefits ensures that
no systematic differences exist between the treatment and control groups
other than the amount and/or form of the food assistance benefit
Multivariate statistical techniques, primarily regression analysis and
related econometric techniques, may be used to estimate the effects of
the amount or form of the food assistance benefit on food consumption
while controlling for differences in both the observable and unobservable
55
characteristics of sample households that may influence their
consumption behavior.
Of the two strategies, classical experimentation has the potential of generating the more
reliable results. However, a classical experiment cannot be implemented within the context of
the Food Stamp Program if it entails withholding assistance to some eligible applicants, and it
requires waiving certain program regulations if it entails changing the form of the benefit These
restrictions, combined with the high cost of implementing a social experiment, mean that classical
experimentation is rarely a realistic research strategy for food stamp research.
The use of multivariate statistical techniques is less expensive and less intrusive en program
participants than is classical experimentation. All but two previous studies and three on-going
studies have relied on this analytic strategy to estimate the food-consumption effects of food
stamps. Unfortunately, several problems are associated with using multivariate analysis which
may introduce considerable uncertainty or even bias into the estimates of the effects of food
stamps that are generated with these procedures. Among these problems are the following:
Model specification. The success of regression analysis and related
multivariate statistical procedures at generating reliable estimates of the
effects of food stamps on food consumption depends heavily on the
reasonableness and completeness of the underlying empirical model
Unfortunately, Southworth's theoretical model of household food
consumption provides little guidance on such basic issues as the choice
and specification of variables to be included in the empirical model, or
on more esoteric issues, such as how the household budget constraint
should be incorporated in the empirical model Given this lack of
guidance from the theory, most researchers have specified and estimated
simple linear models of food consumption. The naivete of these models
casts doubt on the reliability of the consequent estimates of the effects
of food stamp on food consumption.
Functional form. Even if a sound theoretical or other basis exists for
believing that a variable affects household food consumption, there may
be uncertainty about the functional form in which the variable should
enter the empirical model. For example, household income is included
among the explanatory variables in every multivariate model of food
consumption that we have reviewed; however, there is considerable
56
disagreement about whether Income should enter the food consumption
equation linearly, log-linearly, or quadratically. An equal diversity of
opinion exists about the appropriate way to control for the consumption
effects of the number, age, and sex of household members. Estimates of
the effects of food stamps on food consumption may be sensitive to these
and other functional-form decisions.
Selection bias. Participants in the FSP may differ from eligible
nonparticipants in ways that cannot readily be measured. For example,
participants may derive more satisfaction from the consumption of food
than do eligible nonparticipants or may feel a greater desire to improve
their families' diets. Standard multivariate regression has the capacity to
generate estimates of food stamp effects that are unbiased by observed
differences between participants and eligible nonparticipants; however,
those estimates are subject to "selection bias" arising from unobserved
differences. Procedures developed by Heckman (1978 and 1979),
Heckman and Robb (1985), and others have been used to control for
selection bias in food consumption analyses (e.g., Chen, 1983; and
Devaney and Fraker, 1989). However, implementing these procedures
can be expensive, and they have restrictive data requirements which often
are not satisfied.
General-purpose surveys of household labor-force and consumption behavior, such as the
Panel Study of Income Dynamics, the Consumer Expenditure Survey, and the Nationwide Food
Consumption Survey, have provided the basis for most existing estimates of the effects of food
stamps on food consumption. These data were not collected in an experimental context;
consequently, the researchers who use them must rely on multivariate statistical procedures to
estimate the effects of food stamps on food consumption. Those estimates may be subject to bias
and uncertainty from the above sources, as well as from other sources that are noted below.
2. Specification of an Empirical Model of Food Expenditures
We reviewed 17 studies in which multivariate regression analysis or related econometric
techniques were used to estimate the effects of food stamps on food expenditures by households.
No two of the empirical models underlying those studies are the same; however, most represent
some variant of the following basic model of household food expenditures:
57
(1) FOODCOST = a! + a2xBEN + a3xINC + a4xSIZE + XB + e,
where FOODCOST is the money value of food used at home, BEN is the net food stamp benefit
amount,21 INC is money income, X is a vector of control variables (e.g., the race, ethnicity, and
education of the principal meal preparer), a,... a4 are individual regression coefficients, B is a
vector of regression coefficients, and e is a random disturbance term.
In this basic specification of the food expenditure model, the coefficients ^ and 83 are the
marginal propensities to consume food (MPCf) out of food stamps and income, respectively.
Regression estimates of these coefficients, in conjunction with their standard errors, can be used
to test hypotheses generated by economic theory about the effects of cash income and food
coupons on household food expenditures.
The studies that we reviewed exhibited several noteworthy differences in the specification
of the food expenditure model. In some of the studies, FOODCOST, BEN, and INC are
measured on either a per-person basis (e.g., Salathe, 1980b; and Smallwood and Blaylock, 1985)
or a per-adult-equivalent-person basis (e.g., Hymans and Shapiro, 1976; Brown, Johnson, and
Rizek, 1982; and West, 1984). In other studies, those variables are not adjusted as such for
household size and composition, although other means may be used to control for the effects of
those factors (e.g., Benus, Kmenta, and Shapiro, 1976; Neenan and Davis, 1977; and Chen, 1983).
SIZE may be a simple count of household members (West, Price, and Price, 1978; and Chen,
1983), or it may be the number of adult male equivalents in the household (Hymans and Shapiro,
1976; Basiotis et al., 1987; Senauer and Young, 1986; and Devaney and Fraker, 1989). To better
capture the expenditure effects of household members in different age categories, some studies
21In studies based on data collected prior to the elimination of the food stamp purchase
requirement, BEN is the food stamp bonus value-the difference between the face value of the
coupons actually received by the household and the amount that the household paid for those
coupons. In studies based on post-EPR data, BEN is simply the face value of food coupons
received by the household.
58
use variables that reflect a household's stage in its "life-cycle" (e.g., Allen and Gadson, 1983; and
Neenan and Davis, 1977) or that measure the proportion of household members in specified age
categories (e.g., Salathe, 1980; and Smalhvood and Blaylock, 1985). Benus, Kmenta, and Shapiro
(1976) and Chavas and Yeung (1982) use counts of the number of household members in each
of five age categories in lieu of SIZE.
The studies that we reviewed also differ according to the functional form in which income
and the food stamp benefit enter the food expenditure model. They are roughly equally divided
according to whether income enters the expenditure model linearly, as shown in Equation (1)
(e.g., Johnson, Burt, and Morgan, 1981; and Devaney and Fraker, 1989), log-linearly (e.g., West
and Price, 1976; and West, 1984), or quadratically (e.g., Allen and Gadson, 1983; and Basiotis et
al., 1987). The log-linear and quadratic specifications are intended to capture any tendency for
the MPQ out of income to be smaller among households with larger amounts of income. In a
majority of the studies, the food stamp benefit enters the model linearly (e.g., Brown, Johnson,
and Rizek, 1982; and Chavas and Yeung, 1982); however, it appears in quadratic form in the
models specified by Basiotis et al. (1983 and 1987). In the models specified by Neenan and Davis
(1977) and West (1984), the food stamp benefit appears linearly and is also interacted with
household income. The interaction term allows for the possibility that the MPCf out of food
stamps may vary with the amount of cash income. The model developed by Benus, Kmenta, and
Shapiro (1976) uses a Box/Cox transformation to capture the specific degree and form of
nonlinearity indicated by the data for each of the key variables-food expenditures, price, and
income. Hymans and Shapiro (1976) estimate both linear and double logarithmic models of food
consumption, and Senauer and Young (1986) also use the double logarithmic form in their
59
modeling.22 This form provides great flexibility, allowing the model to be nonlinear in all of is
parameters.
The differences across studies in the manner in which income, benefits, and household size
and composition enter the empirical food expenditure model contribute to the diversity in
reported estimates of the MPCf out of income and food stamps. The variation in reported MPCf
estimates is also due to (1) the different control variables (Le., the X vector in Equation (1)) that
are used across studies, (2) the different data sets that are used to estimate the models, (3) the
fact that the models are estimated with samples that represent different segments of the
population (e.g., food-stamp-eligible households, food stamp participants, and all households), and
(4) other factors, such as whether the estimation process uses the sample weights or deals with
the potential problem of sample selection bias. Two early studies based on the Michigan Panel
Survey of Income Dynamics differ further from all the subsequent studies in that they estimate
models on the basis of multiple years of data and estimate long-run equilibrium or steady-state
food expenditure responses rather than more immediate single-period effects (Hymans and
Shapiro, 1976; and Benus, Kmenta, and Shapiro, 1976).
B. HOW EFFECTIVE ARE FOOD STAMPS AT INCREASING FOOD
EXPENDITURES?
Each of the studies on the effects ol food stamps on food expenditures that we reviewed
provides estimates of the MPCf out of food stamps and income and their associated standard
errors. This section reviews the estimates of the effects of food stamps on food expenditures
from these studies. The next section compares those estimates with estimates of the effects of
^In a double logarithmic model, both the dependent and the independent variables appear
in logarithmic form. This is in contrast to a logarithmic model in which only one or more
independent variables appear in logarithmic form.
60
cash income on food expenditures and considers the implications of the differences for the
effectiveness of food coupons versus cash assistance at increasing food expenditures.
1. Estimates of the MPCf Out of Food Stamps
Many of the 17 studies that we reviewed provide more than one estimate of the MPCf out
of food stamps. Some of the studies provide multiple estimates because they use the same model
to generate estimates with two or more different samples drawn from the same data set (e.g.,
West, 1984); others estimate alternative models with the same sample (e.g., Brown, Johnson, and
Rizek, 1982), and still others estimate the same model with similarly defined samples drawn from
two different data sets (Chen, 1983; and Senauer and Young, 1986). Devaney and Fraker (1989)
obtain two different estimates of the MPCf by using weighted and unweighted data to estimate
a single model with the same sample. From among the many estimates provided by these studies,
we have choseu to display those that were generated with what we believe are the most
defensible procedures. They are shown in Table IV.l.23
The existing estimates of the MPCr out of food stamps that are reproduced in Table IV.l
vary greatly in size, ranging from .17 (Johnson, Burt, and Morgan, 1981; West, 1984; and Basiotis
et al., 1987) at the low end, to .64 (Hymans and Shapiro, 1976) and .86 (Benus, Kmenta, and
Shapiro, 1976) at the high end. The two highest estimates are clearly outliers, since the third
highest estimate is .47 (West, 1984), and three other estimates are in the range of .42 (Devaney
and Fraker, 1989) to .45 (Neenan and Davis, 1977; and West 1984).
There are several reasons why the Hymans-Shapiro and Benus-Kmenta-Shapiro estimates
differ substantially from those found in the other studies reviewed:
23Two entries in Table IV.l appear for the studies by Chen (1983) and by Senauer and
Young (1986) because each of these studies provides one set of estimates of the effects of income
and food stamps on food expenditures based on pre-EPR data and a second set of estimates
based on post-EPR data.
61
TABLE IV.1
ESTIMATES Of THE MARGINAL PROPENSITY TO CONSUME FOOO (MPCf) AT HOME. FROM SELECTED STUDIES
Study Data Set
Target Group;
Sample Size
HPCf 7553-
Stamos
Out of:
Money
Income
STUDIES BASED 0* PRE-EPR DATA
Benus, Kmenta, and
Shapiro (1976)
1968-72 Michigan PSID data All households;
n - 3,300
.86 .05
Hymens and Shapiro
(1976)
1968-72 Michigan PSID data All households;
1st half sample,
n - 1,659
2nd half sample,
n - 1,659
Full sample
linear
logarithml
linear
logarithml
logarithml
.35
c .29
.64
: .30
: .29
.14
.24
.17
.23
.23
Nest and Price
(1976)
1972-73 sanple of Washington
State households with child-ren
ages 8-12 years
All households;
n - 992
.37 .05
Neenan and Davis
(1977)
1976 sample of households In
Polk Co.. Florida
FSP participants;
n - 123
.45 .06
West, Price, and
Price (1978)
1972-73 sanple of Washington
State households with child-ren
ages 8-12 years
FSP ellglbles;
n - 331
.31 .03
Salathe (1980b) 1973-74 Consumer Expenditure
Diary Survey
FSP ellglbles;
n - 2,254
.36 .06
Johnson, Burt,
and Morgan (1981)
1977-78 LI supplement to the
NFCS
FSP ellglbles;
n - 3,800
.17 .06
Brown, Johnson,
and Rlzek (1982)
1977-78 LI supplement to the
NFCS
FSP participants;
n - 911
.45 .05
Chavas and Yeung
(1982)
1972-73 Consumer Expenditure
Diary Survey
FSP ellglbles In South;
n - 659
.37 .13
Allen and Gadson
(1983)
1977-78 LI supplement to the
NFCS
FSP ellglbles;
n - 3.850
.30 .08
Chen (1983) 1977-78 LI supplement to the
NFCS
FSP participants;
n - 1.809
.20 .09
West (1984) 1973-74 Consumer Expenditure
Diary Survey
FSP participants;
n - 587
FSP ellglbles;
n - 2,407
.17
.47
NA
NA
Saallwood and
Blay lock (1985)
1977-78 LI supplement to the
NFCS
FSP ellglbles;
n - 2.852
.23 .10
Senauer and Young
(1986)
1978 Michigan PSID data FSP participants;
n - 573
.33 .05
Baslotls, Johnson,
Morgan, and Chen
(1987)
1977-78 LI supplement to the
NFCS
FSP ellglbles;
n - 2,950
.17 .10
Devaney and Fraker
(1989)
1977-78 LI supplement to the
NFCS
FSP ellglbles;
n - 4.473
Weighted data
Unweighted data
.42
.21
.08
.07
62
TABLE IV.1 (continued)
HPC, Out of: "
Food Money
Study Data Set Stawps Income
Target Group;
Sanple Size
Chen (1983)
Senauer and Young
(1986)
Fraker, Long, and
Post (1990)
STUDIES BASED ON POST-EPR DATA
1979-80 LI supplement to the
NrCS
1979 Michigan PSID data
1985 Continuing Survey of
Food Intake by Individuals
FSP participants;
n - 1,630
FSP participants;
n - 574
FSP & WIC ellglbles;
n - 515
.23 .11
.26 .07
.29 .05
NOTE: Table C.l provides additional information on the estimates shown In this table.
63
As noted, these two studies are unique because they use multiple years
of data. The Hymans and Shapiro study uses average values of the first
five years of PSID data (1968-72), thus estimating long-run average or
steady-state MPCs. Benus, Kmenta, and Shapiro estimate a dynamic-adjustment
model with the same data, drawing on both their cross-sectional
and their longitudinal aspects.
The Benus-Kmenta-Shapiro estimate of the MPCf out of food assistance
benefits of .86 reflects the full long-run or steady-state responses of
households to changes in food stamp (and other food subsidy) benefits.
The study does not report an explicit value for the corresponding single-year
impact, which would be comparable to the MPCf estimates reported
in all the subsequent cross-sectional studies in the literature. However
it does note that the estimated steady-state MPCs reported are
approximately twice as large as the corresponding initial-impact MPCs.
Hymans and Shapiro estimate their linear and double logarithmic models
of food expenditures twice, with two randomly selected half-samples
drawn from the 1968-72 PSID, and also estimate the better-fitting double
logarithmic model with the full sample. With the linear model, the first
half sample yields an estimate of the MPCr for low-income urban
households of .35, whereas the second half sample yields the outlier
estimate of .64 for similar households. In contrast to the instability of
the MPCf estimates produced by the linear model, the double logarithmic
model generates estimates of the MPCr that are highly stable (.29 to .30)
across the two half samples and the full sample.
The income and food stamp benefit variables used in both of the early
studies based on the PSID differ substantially from those used in all later
studies. For example, the basic income variable excludes welfare and
nonwelfare transfers, but includes several imputed income elements not
feasible with other data sets. The variable for food subsidy benefits
includes, in addition to food stamp benefits, subsidized meals received at
school or work and other food assistance program benefits.
With the exception of the two outliers, the estimates of the MPCf out of food stamps are
roughly evenly distributed over the range of .17 to .47, indicating that a one-dollar increase in the
face value of the food stamp benefit of a typical recipient household would lead to additional
food expenditures of between $.17 and $.47. All these estimates differ significantly from zero at
levels of statistical precision that are customarily used in hypothesis-testing. Thus, these studies
unanimously confirm the expectation from economic theory that food stamps have a positive
effect on household food expenditures.
64
We are not aware of any pattern in the existing estimates which suggests that the actual
current value of the MPCf out of food stamps is more likely to be in one segment of the range
of estimates than in another. For example, on the basis of theoretical considerations, we expect
that the MPCf out of food stamps would be smaller after the EPR than before the EPR, because
the consumption choices of a smaller proportion of post-EPR recipients are constrained by the
coupon form of the benefit However, the three estimates that are based on post-EPR data
range from .23 to .29 and are only slightly toward the low end of the distribution of all estimates
of the MPQ out of food stamps.
2. Critique of the Estimates
We have sound reasons to believe that the estimates shown in Table IV. 1 vary in terms
of their reliability as estimates of the current MPCf out of food stamps. All but three of the
estimates are based on data that were collected prior to the EPR, and, as noted, there is a
theoretical basis for believing that the EPR led to a downward shift in the MPCf out of food
stamps; hence, were it not for the scarcity of estimates based on post-EPR data, the pre-EPR
estimates would now be of historical interest only. As it is, those estimates should be regarded
as unreliable estimates of the current MPCf and as having a high probability of containing positive
bias.
The current relevance of the estimates provided by the first five studies cited in Table IV. 1
is especially open to question. Those studies are based on data that were collected, at least in
part, prior to the adoption of uniform national standards for food stamp eligibility and benefits
(Benus, Kmenta, and Shapiro, 1976; and Hymans and Shapiro, 1976), or which are representative
of selected demographic groups in limited geographic areas (West and Price, 1976; Neenan and
Davis, 1977; and West, Price, and Price, 1978).
65
With the exception of the studies by Chen (1983), Devaney and Fraker (1989), and Fraker,
Long, and Post (1990), all of the cited studies neglect the potential problem of sample selection
bias. That is, they neglect the possibility that estimates of the MPCf out of food stamps may be
biased by the fact that decisions to participate in the FSP are made voluntarily by eligible
households, and that the underlying food expenditure patterns of those who choose to participate
may differ from those of persons who choose not to participate. Furthermore, with only a few
exceptions (e.g., Basiotis et al., 1987; Devaney and Fraker, 1989; and Fraker, Long, and Post,
1990), the effects of other food assistance programs are not explicitly incorporated into the
empirical food expenditure models, thus introducing the possibility that the estimated effect of
food stamps is positively biased because the food stamp benefit amount and the amount of other
food assistance are correlated.
None of the reviewed studies deals with the fact that the data that form the basis for the
model estimates were obtained largely from complex sample surveys in which the households that
were interviewed had varying probabilities of being selected into the survey samples. Two notable
issues are associated with analyzing data from such surveys. The first is whether the sample
weights should be used in the model estimation process. Devaney and Fraker (1989) show that
the estimate of the MPCf out of food stamps that is generated by applying a conventionally
specified model of food expenditures to data from the Low-Income Supplement to the 1977-78
NFCS is nearly twice as large when the sample weights are used as when they are not Whether
or not sample weights were used may account for much of the variability in the NFCS-based
estimates reported in Table IV. 1. In a comment on the Devaney-Fraker article, Kott (1990)
notes that the difference between weighted and unweighted estimates may be due to differences
in the MPCf out of food stamps between recipient household who live in areas that exhibit low
poverty rates and those who live in areas that exhibit higher poverty rates. The area poverty rate
was a key sample stratifier in the 1977-78 NFCS and, hence, was used to derive the sample
66
weights for the survey. Low-income households located in low-poverty areas were undersampled
in the NFCS; the sample weights can be used to increase the relative importance of such
households in statistical analyses. Kott's h