Household Food Security in the United States,
1995-1997:
Technical Issues and Statistical Report
EXECUTIVE SUMMARY
This is the Final Report for the project,
"Analysis of the Current Population Survey Data for Food Security and
Hunger Measurement" conducted by Mathematica Policy Research, Inc. (MPR)
for the USDA Food and Nutrition Service (FNS), beginning in 1997. The
project provided USDA with technical support and statistical estimation
work for analyzing the 1996 and 1997 data on food security collected in
the U.S. Census Bureau, Current Population Survey (CPS) Food Security
Supplement. More broadly, the work examined a number of analytic and
empirical issues relevant to analyzing the first three years of CPS food
security data available—those for 1995, 1996, and 1997.
It was originally
intended that the Final Report would provide the main vehicle for
dissemination of the substantive findings on the prevalence of food
insecurity based on the 1996 and 1997 data. However, because of the
importance of making these results available as early as possible, USDA
elected to issue an "Advance Report," thus making the results of
the 1996 and 1997 analyses conducted by MPR available before completion of
the overall project. In addition, since 1999, a number of publications
have become available that present estimates of food insecurity
prevalence, as well as discussions of the methods used in computing food
security estimates in general. Most important, Andrews et al. (2000)
provides a comparative analysis of the annual data for the five-year
period 1995 through 1999, while Bickel et al. (2000) provides a how-to
guide for measuring food security that incorporates relevant work done
prior to that time, including earlier work from the current project.
Selected issues in food security are also considered in Ohls et al. 1999.
In light of these developments, USDA suggested that MPR recast this Final
Report to focus on several selected topics related to the 1995-1997 data,
rather than provide comprehensive treatment of the overall research, much
of which has since been incorporated in later publications. The Final
Report has been organized around these suggestions.
Among the issues
addressed in the report are:
-
The stability of the food security
measurement scale over time
-
Temporal adjustments to the categories or designated ranges of severity
on the underlying continuous scale used to classify households by food
security status
-
Screening issues related to ensuring a strictly comparable analysis
sample over the 1995-1997 CPS food security samples
-
Alternative imputation strategies for dealing with missing data
-
The degree to which household responses to the food security questions
are "modal," in the sense that households consistently respond
affirmatively to questions involving less severe food insecurity whenever
they respond affirmatively to questions involving relatively more severe
food insecurity
-
The degree to which the estimated parameters of the model used to
measure the severity of food insecurity vary across different groups of
households, defined by ethnicity and other characteristics
The first
section below provides background information about the analysis.
Subsequent sections summarize findings on each of the above issues.
Background
The analysis in this report is based on a statistical procedure
which assigns households to food security status, based on their answers
to a series of 18 survey questions. The food security categories used are:
The data used for national-level analysis of
food security are from annual supplements to the CPS, which is fielded
monthly to more than 40,000 U.S. households.
Households are classified by
food security status in the analysis, based on a procedure, Rasch
modeling, which has a long history in the statistical literature. The
first work in applying the Rasch model to food security data was
undertaken under an earlier contract let by USDA (Hamilton et al. 1997).
The Rasch model, as used in that work, posits that there is a single,
one-dimensional attribute among households that indicates food insecurity.
The model then uses a set of assumptions and statistical methods to assign
"severity levels" to each of a series of 18 survey questions
relating to food insecurity and hunger. A continuous food security measure
is then assigned to each household in the data set, based on households’
replies to the 18 questions. Supplemental procedures developed by Hamilton
et al. are then used to translate the continuous scale score into a
limited number of discrete food security statuses.
The objectives of the
current project were to extend the analysis to 1996 and 1997 data and to
address a number of related issues associated with measuring food security
over time. Our findings in selected areas are summarized below.
Stability
of the Parameters of the Model Over Time
An important issue in examining
the validity of the Rasch modeling approach is whether the model parameter
estimates are stable over time. The underlying theory on which the Rasch
model is based posits that, if the wording of an item does not change, its
estimated level of severity should not change. For example, even if food
insecurity became more prevalent over time, a household at a given level
of insecurity this year is expected to answer each item the same way a
household at that level of insecurity did a year earlier. Due to sampling
variability and other factors, such as minor wording changes, we do not
expect estimated model parameters to remain exactly the same over time;
but a finding of major changes over time would call into question the
validity of the model. Particularly problematic would be a finding of
important changes in the ordering of the items by severity.
To examine
issues of model stability, we estimated the model independently on three
CPS data sets (1995-1997), using consistent conventions as to statistical
scaling. Some variation across years was found, as expected. In general,
however, the estimated parameters of the model were quite stable. Also,
the estimated order of severity of the different questions remained
largely constant, with the only changes in severity order occurring among
questions that were very close to each other on the severity scale in the
original estimation work. The conclusion of this component of the research
is that the food security model is sufficiently temporally stable to make
it a reasonable tool to use in time series analysis.
Adjusting "Cut
Points" Used to Classify Households into a Limited Number of Food
Security Status Categories
The Rasch model places each household on a
continuous numeric food security scale. For purposes of policy analysis,
it is also useful to establish numerical "cut points" that
assign households to a small number of designated categories which
summarize their food security status. To create this categorical measure,
Hamilton et al. (1997a and 1997b) specified four categories: food secure,
food insecure without hunger, food insecure with moderate hunger, and food
insecure with severe hunger. More recently, the latter two categories have
usually been collapsed to form a single category, while additional scale
development work has identified a new nested category, food insecure with
children’s hunger (Nord and Bickel 1999 and 2001).
A key issue that
arises in this work is whether it is appropriate to keep the same
continuous scale cut points over time, or whether, alternatively, some
temporal adjustments may be needed. The analysis of the body of the report
concludes that, at least in some situations, it is not optimal to attempt
to classify households based on the same cutpoints over time.
While the
Rasch model places households on a continuous food security scale, due to
certain statistical properties of the model substantial numbers of
households tend to be clustered at certain points in the scale. If cut
points are held constant, there is a risk that, because of chance
statistical variation, the score assigned to one of these clusters of
households might accidentally cross one of the cutpoints in a given year,
causing considerable instability in estimates of food security prevalence.
Chapter V of the Final Report identifies several technical approaches for
avoiding this difficulty. The discussion is based on the principle that a
household with a given pattern of survey answers should always be
classified into the same food security grouping, independent of when the
data are collected.
Screening Household into the Sample in the 1995-1997
Surveys
The food security supplements in the 1995-1997 CPS had two general
sections. The first section gathered information about food expenditures,
participation in several programs aimed at providing food to needy
families (for example, food stamps and school meal programs), and the
sufficiency of food eaten during the preceding 12 months. The second
section gathered more-detailed information about food insecurity and
coping behaviors during the previous 12 months and prior 30 days. Not all
households were asked this second set of questions, which includes the
questions used to construct the food security scale. In order to minimize
respondent burden, households who, on the basis of earlier questions,
appeared to have a high likelihood of being food secure were excluded from
the more detailed questions and were assumed to be food secure in the
analysis. This prescreening applied to higher income households in all
three years, 1995-1997, and in one year, 1996, it was applied to
lower-income households as well. Beginning in 1998 and continuing
consistently since then, the CPS Food Security Supplement has included a
new, less restrictive, pre-screen applied to higher-income households.
To
ensure comparability in the analysis samples for the three years, the
current research developed a common screen, such that any households
giving survey answers that passed this common screen would have been
tracked into the food security module in any of the three years.
Households that did not pass the common screen were, for purposes of the
analysis, treated as if they had not been tracked into the food security
module of the survey—essentially, they were assumed to be food secure.
Technical details concerning how this common screen was constructed are
provided in Appendix B of the Final Report.
While use of the common screen
has the desired effect of ensuring consistency in the 1995-1997 analysis
samples, it also has the effect of treating as food secure a number of
households who, during the survey, gave indications of experiencing food
insecurity. Across the three years, use of the common screen was found to
result in estimates of the prevalence of food insecurity which are between
1.0 and 1.5 percentage points lower than those that are obtained when the
maximum available samples are used in the estimation.
Imputing Missing Data
Most households gave complete answers to the food security questions
they were asked in the CPS; however, a limited number did not. Appendix C
of the report examines a number of alternative approaches for including
households with partially missing data in the analysis. One approach is
reliance on the Rasch model itself, which has the capacity to assign food
security scale scores to observations with incomplete data. However, as is
noted in Appendix C, in some instances, the determinations made within the
model for cases with substantial amounts of missing data may lack face
validity. Also, as a practical matter, many researchers may not have easy
access to the software needed to implement the model.
An alternative
algorithm for dealing with missing data has therefore been developed.
Depending on the exact configuration of food security module answers given
by the respondent, this alternative algorithm essentially involves
imputing missing data items based on either (a) the highest severity item,
in terms of level of food security severity, that the respondent answered
positively; or (b) the lowest severity item answered negatively.
"Modality" of Household Food Security Response Patterns
The
Rasch model implies that many households will exhibit item response
patterns that are reasonably "modal" in the sense that if a
household answers "yes" to any of the items, it will tend to
answer "yes" to the less severe items, then answer
"no" to the more severe items. A household that exhibits this
pattern exactly—a string of all "yes" answers followed by a
string of all "no" answers—is said to be a "modal"
household. There is nothing in Rasch theory that predicts that all
households will be modal; indeed, the model cannot be estimated if all
households are exactly modal. Still, it is of interest in understanding
the data to examine the degree of modality present. A large number of
strongly nonmodal response patterns could call into question the validity
of the model.
Analysis of the 1997 data indicates that most household
response patterns tend to be either exactly or approximately modal. Of
those households in the 1997 data who gave an affirmative answer to at
least one question, approximately 39 percent households provided answer
patterns that were exactly modal, while another 36 percent gave sets of
answers which had only a single nonmodal response.
Consistency of
Estimated Food Security Model Parameters Across Population Subgroups
Essentially, the analysis conducted with the aggregated CPS data sets
assumes that different subgroups of the population are similar with regard
to how they experience food insecurity. To test this assumption, the Rasch
model was estimated separately for subgroups of the population, defined
according to (a) race/ethnicity; (b) household composition; (c)
metropolitan status; and (d) region of country.
The results of this
analysis indicate considerable robustness of the analysis to this kind of
disaggregation. In general, estimated severity levels for the individual
questions were found to have consistent patterns across different
subgroups, and the magnitudes of the parameters do not change
substantially.
There is no clear statistical test of how much difference
in the estimated subpopulation models would affect confidence in the
overall modeling approach. However, the judgment of statistical experts
who have used the Rasch model extensively in other contexts is that the
findings of the subgroup analyses can reasonably be judged to be highly
consistent with one another.
Conclusion: Reflections on the Strengths and
Limitations of the Food Security Methodology
We conclude by discussing the
strengths and limitations of the use of the Rasch model as a basis for
food security measurement. Possible directions for future research are
also noted.
The food security scale reflects more than 10 years of
methodological development by both government and private groups. The use
of the Rasch model methodology has made it possible to guide the
development of the food security estimates with a thoroughly studied model
that has well understood statistical properties. In terms of
goodness-of-fit criteria, the mathematical form of the measurement model
shows strong correspondence on "fit" to the empirical data. The
approach has undergone extensive review by experts in both the public and
the private sector. In general, these experts consider the model an
appropriate application of the IRT methodology, and they have viewed the
analysis results as reasonable.
Another important strength, as established
by the current project, is that the estimated item parameters of the IRT
model are robust across time and population subgroups. The values obtained
from the 1996 and 1997 data are essentially the same as the original 1995
values. In addition to stability over time, there is stability across
subgroups, defined by such characteristics as race/ethnicity, household
composition, and region of the country.
Tempering these strengths are a
number of limitations which should also be recognized. Most of these, if
not all, are a matter of careful interpretation of what the food security
measure does and does not do. For example, the CPS indicator questions for
food insecurity and hunger and the scale developed from them are designed
to provide a household-level measure of the severity of conditions as
experienced within U.S. households. This is in line with the conceptual
understanding of food insecurity as a condition of deprivation or stress
experienced by households in meeting members’ basic food needs. However,
the experience of hunger as such, which appears only at a more severe
stage of food insecurity, is strictly individual. The household
classification, "food insecure with hunger" refers to that more
severe range where evidence of reduced food intake and hunger has appeared
for one or more household members. But this is a collective measure which
may apply to all household members, to adult members only, or to as few as
one (adult) member.
Second, the basic measure is designed to capture
respondents’ experiences over the course of a year, while household
circumstances can change markedly during such a period. Accordingly, the
12-month measure—designed to provide reliable benchmark and trend
figures—may not represent the current situation of given households.
Similarly, the "food insecure with hunger" designation can, in
principle, result from just one serious episode during the year, although
for most such households evidence of a repeated pattern of reduced (adult)
food intakes during the year must be established.
In addition, a number of
issues of interpretation flow from the need to have a simple categorical
measure as a means of classifying households for purposes of manageable
data reporting and monitoring, in addition to the underlying continuous
scaled measure. The categorical measure was created to make the scale more
accessible to non-technical users and more convenient to users whose needs
could be better served by a simple categorical variable than by the
detailed continuous measure. The categorical measure as such is
straightforward: it represents designated ranges of severity along the
continuous scale (i.e., qualitatively differing severity levels of
"food insecure"), plus the category of households that either
show no evidence of food problems within the CPS data set, and hence can
be deemed to be "food secure," or that show only one or two
indications of food stress, which is deemed insufficient as evidence to
establish confidently their status as "food insecure."
The
interpretive problems with the categorical measure stem from at least
three sources. First, the designation of appropriate severity ranges, and
their exact delineation in operational form based on the available set of
indicators, is inherently judgmental and thus leaves room for
disagreement.
Second, the Rasch model employs a probabilistic logic in
generating the continuous scale measure of severity of household food
insecurity; similarly, the corresponding severity-range summary categories
share this probabilistic nature. However, the naming conventions adopted
for the severity-range categories are determinate in form, which can be
misleading.
Thirdly, a misplaced specificity and determinateness can
easily be attributed to the individual indicator items as well, causing a
misunderstanding of their actual role in the measurement process.
To
illustrate this last point, straightforward names adopted for the
severity-range categories raise issues of face validity when they
seemingly contradict the clear language of particular indicator items
embedded within the measurement scale. For instance, it is technically
possible for a household to be classified "food insecure with
hunger," even though the respondent has answered "no" to
the particular question, "In the last 12 months were you ever hungry
but didn’t eat because you couldn’t afford enough food?" In this
case, the respondent either must have replied "yes" to a series
of increasingly severe indicators of food insufficiency, including at
least three items indicating reduced food intake for themselves and/or
other adult members of the household, one of which establishes a repeated
pattern of such reduced intakes over the year, or they must have replied
"yes" to most of the foregoing, plus one or more of the items
that are more severe than the explicit hunger question. The categorical
measure (and its naming convention) reflects the judgment that, on the
balance of this evidence, one or more adult members of the household has,
with high probability, experienced resource-constrained hunger sometime
during the year. Conversely, the opposite case also can occur: the
household can be classified "food insecure without hunger,"
based on its overall pattern of response and the resulting scale score,
even though the respondent has answered "yes" to the explicit
hunger question as such.
In creating the scale, a number of steps were
taken to minimize the effects of these factors. For instance the numerical
cutpoints defining the categories were set to be conservative, in the
sense that there must be three answers to questions thought to indicate
food insecurity before a household is classified as food insecure, and
similarly for the hunger classification. Also, analysis presented in the
text of the report indicates that substantial numbers of respondents
follow close-to-expected, response patterns, which do not lead to any
apparent anomalies in classification. Nevertheless, room for disagreement
remains as to what types of answers to the questions should be construed
as reflecting the language used in designating the three scale categories.
A possible solution to some of these issues would be to state the category
names in more-probabilistic terms, such as "probably food
insecure" or "a high likelihood of hunger." This would be
in keeping with the probabilistic nature of the underlying model, and it
might help ease the concerns of those who are bothered by the anomalies
posed by apparently inconsistent patterns of question responses. However,
such category name changes might also interfere with the clarity of
meaning of the categories themselves, thus reducing their effectiveness.
Overall, it is important to recognize that these limitations have not
prevented the food security scale from becoming an important, widely used
research and policy tool. Questions to support the scale have been
included in an increasing number of national surveys and scale results are
frequently cited in the policy process. This evidence suggests that many
policy analysts have found the scale to be a valuable tool for measuring
an important aspect of material deprivation among America’s poor.
December 2001
Last modified: 12/04/2008
|