QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

This Article Abstract Full Text (PDF) Online Supporting Material Purchase Article View Shopping Cart Alert me when this article is cited Alert me if a correction is posted Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Wilde, P. E. Search for Related Content PubMed PubMed Citation Articles by Wilde, P. E.

---

Community and International Nutrition

Differential Response Patterns Affect Food-Security Prevalence Estimates for Households with and without Children¹

Gerald J. and Dorothy R. Friedman School of Nutrition Science and Policy, Tufts University, Boston, MA 02111

²To whom correspondence should be addressed. E-mail: parke.wilde{at}tufts.edu.

	ABSTRACT

TOP ABSTRACT THEORY AND METHODS RESULTS DISCUSSION LITERATURE CITED

 
To produce national prevalence estimates for "food insecurity" and "food insecurity with hunger," the USDA uses a battery of 18 survey items about symptoms of food-related hardship. Ten items refer to adults in the household, while 8 items refer to children in the household and hence are not asked of households without children. To equate food-security status in households with and without children, the USDA uses a statistical model from item response theory, known as the Rasch model. This model requires an assumption that adult-referenced items have the same severity calibrations for all households, including households with and without children. However, empirical estimates from the 2000 Current Population Survey showed significantly different severity calibrations for households with and without children. These differences have implications for observable response patterns. Holding constant the number of affirmative responses to adult-referenced items, households with children were more likely to respond that they "worried food would run out," and households without children were more likely to be unable to afford "balanced meals." In light of such differences, the Rasch model cannot be used to equate the food-security status of households with and without children. One potential solution would be to estimate household food security by using the same battery of adult-referenced survey items for all households.

KEY WORDS: • food insecurity • household hunger • Rasch model

Every year, the USDA estimates the number of American households that are "food insecure" and "food insecure with hunger." In 2000, 10.5% of American households were classified as food insecure, including 3.1% classified as food insecure with hunger (1). The federal government uses these estimates to evaluate major food assistance programs, and the media report these estimates prominently. The Food and Nutrition Service, the USDA’s largest agency, uses food-security prevalence estimates for low-income households as a performance measure in the agency’s strategic plan for the Food Stamp Program (2). Reducing the overall food-insecurity prevalence also is a goal in the Department of Health and Human Services’s "Healthy People 2010" initiative (3). Like poverty estimates from the Census Bureau and unemployment estimates from the Bureau of Labor Statistics, these annual estimates from the USDA are the most authoritative source of empirical information about a critically important social condition—the extent of food insecurity and hunger in America.

The source of these prevalence estimates is the U.S. Household Food Security Survey Module, which is fielded as a supplement to the Census Bureau’s Current Population Survey (CPS). The module uses 18 survey items that ask about a range of experiences or behaviors associated with food insecurity or hunger in the past 12 mo (4). The underlying questions may be binary choice or multiple choice but, for food-security measurement, each item is eventually coded as a binary response: "yes" or "no."

Ten items refer to the respondent and other adults in the households, while 8 items refer to children < 18 y old. For example, the least severe item is an adult-referenced item, which asks whether the household members "worried that food would run out before we got money to buy more." The most severe item is a child-referenced item, which asks whether any children in the household "[did] not eat for a whole day because there wasn’t enough money for food?" The child-referenced items are omitted for households without children. In almost all other respects, the survey treats households the same regardless of household characteristics.

The USDA determines the food-security status of households by using their raw score, the number of affirmative responses to the 18 items in the module (4). The threshold raw score for being classified as "food insecure" is 3 in households with children and 3 in households without children. The threshold raw score for being classified as "food insecure with hunger" is 8 in households with children and 6 in households without children. One could not simply use the same threshold raw score for both groups, because answering "yes" to 8 of 10 items represents a much higher degree of food insecurity than does answering "yes" to 8 of 18 items.

The official prevalence estimates for food insecurity and hunger depend fundamentally on the selection of threshold raw scores. To set the threshold raw scores equivalently for households with and without children, the USDA relies on the Rasch model, a statistical model from the field of item response theory (IRT). Understanding this application of the Rasch model is important, not just for scholarly research in IRT but to correctly understand the food-security prevalence estimates in general. The goal of this article is to explain the specific assumptions required for using the Rasch model to equate food-security status in households with and without children and to test whether these assumptions are consistent with the data.

	THEORY AND METHODS

TOP ABSTRACT THEORY AND METHODS RESULTS DISCUSSION LITERATURE CITED

 
The USDA used the Rasch model to guide the original development of the food-security scale and still uses the model to justify key components of the ongoing measurement approach (4–6). If the model’s assumptions are correct, it offers a single metric for food insecurity in households with and without children, even though they were asked different item sets. As a USDA technical report in 2001 described: "The IRT model used here provides a way to equate the 10-item test administered to childless households with the 18-item test administered to households with children" (6).

Like logistic regression, the Rasch model is a statistical model for a binary variable, such as the "yes" or "no" response to a survey item. The odds for an affirmative response to item t can be described as the probability of a "yes" response divided by the probability of a "no" response. The natural log of the odds is a commonly used tool for expressing the probability of an affirmative response. For interpretation, log-odds of –2, 0, and 2 are equivalent to probabilities of 12%, 50%, and 88%, respectively.

The Rasch model has two key assumptions, which are summarized briefly here (further details are in the online technical appendix³). The first assumption is that the log-odds is a simple linear function of a household-specific food-insecurity score () and an item-specific severity calibration (ß_t):

(1)

where P_t() is the probability of an affirmative response to item t for a household that has a food-insecurity score of . The index t runs over the 10 adult-referenced items for childless households and over all 18 items for households with children. All else equal, the probability of an affirmative response rises as the household food-insecurity score rises and falls as the item severity calibration rises.

The second assumption, known as conditional independence, implies that the probability of an affirmative response to each item is statistically independent for all households that share the same level of food insecurity. Conditional on the level of food insecurity, learning that one item has an affirmative response does not provide the analyst with any new information about the probability that another item has an affirmative response.

The Rasch model does not assume that all households have the same unconditional probability of an affirmative response. The food-insecurity score may be related to any number of important characteristics, such as income, race/ethnicity, or region of the country. Those households with a higher food-insecurity score have a higher probability of answering "yes" to each item. However, the two key Rasch model assumptions imply that, for households that share the same food-insecurity score, the probability of an affirmative response to each item is constant across households.

The number assigned to an item severity calibration has no meaning in isolation. Only the difference between item calibrations, or the difference between an item calibration and a food-insecurity score, has meaning. In applied research with the Rasch model, an arbitrary convention is used to define the zero value in the scale. In this sense, the Rasch model’s parameters are like a temperature scale in which zero may be defined equally well as the temperature at which water freezes (as in the Celsius scale) or the temperature of a water/ice/salt mixture (as in the Fahrenheit scale), as long as the scale is internally consistent. In this article, zero was defined as the severity calibration for the least-severe item ("worried food would run out"). For each of the remaining items, the calibration parameter represents the relative severity for that item in comparison with the least-severe item.

The USDA chose the threshold raw scores for defining "food insecure" and "food insecure with hunger" by using Rasch item severity calibrations in combination with expert opinion about what degree of food insecurity deserved each label. In developing the official approach, the USDA relied on an expert working group to select the threshold items that best indicated the boundaries between successive stages of food insecurity (5,6). The threshold item selected for the "food insecure" classification asked whether the household "couldn’t afford to eat balanced meals." The threshold item for the "food insecure with hunger" classification indicated whether the adults in the household "cut the size of meals or skipped meals in three or more months" in the past year. Based on the Rasch model’s severity calibrations for these items, the USDA selected threshold food-insecurity scores, and eventually threshold raw scores, for households with and without children.

This use of the Rasch model to equate responses across household types is only valid if the calibrations for adult-referenced items are the same for all households (7). A limited body of previous research has tested this assumption empirically for households with and without children. An appendix to a USDA technical report presented Rasch item calibrations for a variety of subgroups, based on race, region, metropolitan/nonmetropolitan residence, the presence of children, and the presence of elderly persons (6). The appendix showed significantly different calibrations for three household types: households with children, households with elderly persons and no children, and other households. The accompanying discussion treated the differences as negligible and argued that they arise from an "arbitrary scaling convention" and from the fact that child-referenced items tend to be more severe.

A recent article in this journal investigated interactions between Rasch model parameters and many household characteristics, one of which was the presence of children (8). Including an interaction term is equivalent to allowing item calibrations to vary across household types. This approach offers a method for testing the model’s assumptions by comparing the Rasch model, which permits no such interactions, with a more general statistical model that does have interactions. The article found interactions between the presence of children and several calibration parameters.

This study used two methods to investigate whether the adult-referenced item calibrations were constant across different types of households. The first method was to compare the Rasch model’s item severity calibrations by using separate samples of households with and without children. The second method was to compare the frequency of affirmative responses with each item, for households that had the same adult-referenced raw score, which was defined as the total number of affirmative responses out of the 10 adult-referenced items. If households with the same adult-referenced raw score have systematically different response probabilities, depending on whether they contain children, then an assumption of the Rasch model is violated.

Not all violations of the Rasch model’s assumptions necessarily affect the main food-security prevalence estimates. In principle, even if the item calibrations differed across households with and without children, one might still find justification for the raw score cutpoints used in the USDA approach to prevalence estimation. After estimating and comparing item calibrations for households with and without children, this study investigated whether the ranking of the threshold items still agreed with the USDA cutpoints for the two household types.

    Data. The CPS is representative of the civilian noninstitutionalized population in the United States. In September 2000, the U.S. Household Food Security Survey Module was administered to 41,060 households in 8 CPS rotation groups, of which one rotation group was used to test experimental questions. This study used a sample of 35,555 households in the remaining 7 rotation groups. The USDA’s suggested methodology was used for imputing the comparatively small number of idiosyncratic nonresponses to selected items (4). Technically, the USDA recommends this imputation for prevalence estimates but uses data without imputation for estimating Rasch model parameters. Here, the imputation was used for all results to assess the statistical model’s applicability to the actual data that are used for prevalence estimation. The imputation is believed to make little difference in any case.

The CPS has a complex sample design, with stratification and clustering. Sample weights were used for descriptive results in this article, as they are for the main USDA prevalence estimates. This study followed the common practice in Rasch model estimation of not using the sample weights for multivariate estimation. Weights have been used in some previous USDA technical reports (6) and not used in others (5). Under the assumptions of the Rasch model, weights are not necessary for consistent parameter estimation, and the USDA staff believes that the use or nonuse of the survey weights makes little difference in practice.

The CPS does not report cluster identifiers, so it is not possible to compute design-corrected standard errors by using the usual Taylor series or replication methods. Design effects are generally thought to be small for household-level statistics from the CPS. Based on previous research on standard errors, a recent USDA report used a constant design effect of 1.6 in computing standard errors (9). This means that corrected variances were about 1.6 times as large as variances computed under the assumption of simple random sampling. Rather than make an adjustment by assumption, standard errors with no correction for design effects were used in this study. To compensate, a conservative significance level of 0.01 was used for hypothesis tests, unless otherwise noted.

	RESULTS

TOP ABSTRACT THEORY AND METHODS RESULTS DISCUSSION LITERATURE CITED

 
The least severe item, asking whether the respondent "worried," was answered affirmatively by 22.2% of households with children and 11.3% of households without children (Table 1). The threshold item for food insecurity, asking about "balanced meals," was answered affirmatively by 12.5% of households with children and 8.5% of households without children. The threshold item for food insecurity with hunger, asking about "cutting or skipping meals in three or more months," was answered affirmatively by 4.8% of households with children and 3.3% of households without children. For each of the adult-referenced items, households with children exhibited more frequent indications of food-related problems.

View larger version (21K): [in this window] [in a new window]   FIGURE 1 Percentage of households responding affirmatively to the "worried" and "balanced meals" items for households that share the same adult-referenced raw score.

	DISCUSSION

TOP ABSTRACT THEORY AND METHODS RESULTS DISCUSSION LITERATURE CITED

This empirical finding is consistent with a limited body of previous research. For example, an appendix to a USDA technical report indicated that the item calibrations for "worried" and "balanced meals" were more different for households with children and more similar for households without children (4). Similarly, previous research in this journal found that the interaction between the parameter for "worried" and the presence of children was positive, while the interaction between the parameter for "balanced meals" and the presence of children was not significantly different from zero (8). Both previous studies included the presence of children as one of several household characteristics of interest, not as a primary focus. The present study emphasizes the presence of children, because this characteristic is the only one that triggers a difference in the number of items asked as part of the U.S. Household Food Security Survey Module.

The results are not an artifact of a scaling convention. For example, in Table 4, the item calibration for "worried" appears equal across household types (zero in both cases), although the calibration parameter for "balanced meals" differs by 0.93 logits. Suppose one used an alternative convention in which "balanced meals" was the reference item. In this case, the calibration parameter for "balanced meals" would be equal across household types (zero in both cases), but the calibration for "worried" would differ by 0.93 logits. As long as one recalls that the actual numbers assigned to Rasch model parameters are arbitrary and that only comparisons between items are meaningful, it is possible to identify the same substantive results under any scaling convention. Moreover, these results are not because most child-referenced items are more severe than most adult-referenced items. That pattern has no influence over the comparisons for adult-referenced items in Tables 4, and 5.

Holding constant the adult-referenced raw score, households with children seem comparatively prone to worrying, while households without children seem comparatively prone to difficulty in acquiring a balanced meal (Table 5 and Fig. 1). This pattern merits further study, because it may provide insight into the distinct nature of food-related hardship for different types of households. Parents bear weighty responsibilities, with many causes for worry, in providing food for their children. People without children also face important challenges in acquiring food, but holding constant the approximate level of food security, the manifestation of those struggles appears qualitatively different.

This study investigated whether the violation of the Rasch model’s assumptions is sufficiently large to call into question the equivalence of the threshold raw scores for households with and without children. For example, if the threshold "balanced meals" item were still ranked third for both household types, it could be argued that the USDA’s threshold raw scores still represented approximately the same degree of food security for both household types. However, when the Rasch model’s item severity calibrations are estimated separately for the two household types (Table 4, columns 3 and 4), the "balanced meals" item is ranked fourth for households with children and third for households without children. Likewise, in the simple response frequencies for the 18 items, the "balanced meals" item is ranked fourth for households with children and third for households without children (Table 1). In this respect, the violation of the Rasch model’s assumptions has implications for the main food-security prevalence estimates. Neither Rasch analysis nor a simple comparison of the item rankings confirms that the current threshold raw scores represent the same level of hardship for households with and without children.

What should be done? It is beyond the scope of this article to attempt an answer that should immediately be convincing to the several constituencies with influence over federal food-security measurement. Indeed, a strength of the USDA’s food-security measurement project is its deliberative and consultative approach, drawing on leading expertise inside and outside government. At the same time, it would be overly cautious to simply note the empirical difficulties in the USDA’s approach without offering any forward-looking suggestions.

One solution would be to modify or to eliminate selected items, in hopes that the remaining items will exhibit no differences across households with and without children. For example, previous research has questioned whether the "balanced meals" item is understood the same way by all population groups, based on evidence from Hawaii (10), Indonesia (11), and Trinidad and Tobago (12,13). A solution may be to eliminate this item.

However, there are several disadvantages to this solution. Such a change solves the problem of differences across household groups for only one item, leaving other significant differences without remedy. On a practical level, eliminating one item leaves the survey module with 9 adult-referenced items out of 17 items in total and still no assurance that the Rasch model’s assumptions will be met sufficiently to reliably equate food-security status across household groups. Moreover, this solution takes food-security research further down the path of modifying the choice of variables in the data to fit the model, rather than modifying the model to fit the data. If one finds it inherently plausible that households at the same level of food security have systematically different response frequencies to food-security items, depending on whether the households have children, then it seems preferable to reduce the reliance on a model that forbids such empirical patterns by assumption.

Thus, a second solution would be for the USDA to consider weaning food-security measurement from the Rasch model as a tool for combining data from households with and without children. There is no need to curtail the model’s use for analysis and research. Statistical models are used all the time in the analysis of important federal statistics. If one recognizes their strengths and weaknesses, such models can provide great insight. However, the Rasch model does not provide a sound justification for equating households with and without children in determining the key raw score thresholds for "food insecurity" and "food insecurity with hunger."

In pursuing this second solution, the USDA could consider basing the national prevalence estimates on adult-referenced items only. Because the same adult-referenced items are asked of all households, there would be no need to use different threshold raw scores for different types of households. This approach would sacrifice the foundation in measurement theory that the Rasch model appears to offer, but it could still be justified as a straightforward method for defining the food-security status of all households.

At the time the food-security measure was developed, there was no child hunger scale and there was less recognition that food insecurity among children might represent a separate dimension, distinct from the household-level phenomenon reflected in the adult-referenced items. It seemed very important to preserve child-referenced items as part of the main household-level scale. In recent years, studies have emphasized children’s hunger as a distinct phenomenon (14). If children’s hunger requires a separate measurement effort in any case, it seems less essential to retain child-referenced items in the main 12-mo household scale.

How much would the main national prevalences change if they were estimated by using adult-referenced items only? Interestingly, a footnote to the most recent USDA annual food security report mentioned that eliminating the child-referenced items would reduce the difference in estimated household food-security prevalences across households with and without children (9). Such a change would have no effect on prevalence rates for households without children, because these households already are asked only adult-referenced items. This change would affect the prevalence estimates among households with children, for whom the number of survey items would be reduced. National prevalence estimates are a weighted average of the separate prevalences for households with and without children, so the national estimates could change as a result of changes for households with children.

To estimate new prevalence estimates for households with children, the natural raw score thresholds are the same thresholds used for households without children: a household with an adult-referenced raw score of 3 or more would be "food insecure" and a household with an adult-referenced raw score of 6 or more would be "food insecure with hunger." By using these thresholds, the 2000 CPS estimates for the prevalence of "food insecurity" would fall from 16.2 to 12.7% of households with children (Table 3). The prevalence of "food insecurity with hunger" would fall from 3.9 to 3.7% of households with children (Table 3). For all households, the estimated national prevalence of "food insecurity" would fall from 10.5 to 9.3% and the estimated prevalence of "food insecurity with hunger" would remain nearly unchanged.

This revised food-insecurity prevalence is as justifiable as the official estimate. It would be straightforward to compute revised prevalence estimates for previous years, to produce a consistent series from 1995 onward. Such an effort would reduce the official reliance on assumptions of the Rasch model that do not appear to agree with the data and would introduce few negative consequences for the food-security measurement effort as a whole.

	FOOTNOTES

 
¹ This research was supported by the Friedman School of Nutrition Science and Policy. These results have not been presented previously.

³ Supplemental data and a technical appendix are available with the online posting of this article at www.nutrition.org.

Manuscript received 17 December 2003. Initial review completed 15 February 2004. Revision accepted 5 May 2004.

	LITERATURE CITED

TOP ABSTRACT THEORY AND METHODS RESULTS DISCUSSION LITERATURE CITED

 

1. Nord, M., Kabbani, N., Tiehen, L., Andrews, M., Bickel, G. & Carlson, S. (2002) Household Food Security in the United States 2002:2000 U.S. Department of Agriculture, Economic Research Service Washington, DC.

2. Food and Nutrition Service (2000) Strategic Plan 2000 to 2005 2000 U.S. Department of Agriculture, Food and Nutrition Service Alexandria, VA.

3. Department of Health and Human Services (2000) Healthy People 2010: Understanding and Improving Health Second Edition 2000 U.S. Government Printing Office Washington, DC.

4. Bickel, G., Nord, M., Price, C., Hamilton, W. & Cook, J. (2000) Guide to Measuring Household Food Security 2000 U.S. Department of Agriculture, Economic Research Service Washington, DC.

5. Hamilton, W., Cook, J., Thompson, W., Buron, L., Frongillo, E., Olson, C. & Wehler, C. (1997) Household Food Security in the United States in 1995: Technical Report of the Food Security Measurement Project 1997 U.S. Department of Agriculture, Food and Nutrition Service Alexandria, VA.

6. Ohls, J., Radbill, L. & Schirm, A. (2001) Household Food Security in the United States 1995–1997: Technical Issues and Statistical Report 2001 U.S. Department of Agriculture, Food and Nutrition Service Alexandria, VA.

7. Opsomer, J., Jensen, H., Nusser, S., Drignei, D. & Amemiya, Y. (2002) Statistical Considerations for the U.S. Food Security Index. Working Paper 02-WP 307 2002 Center for Agricultural and Rural Development, Iowa State University Ames, IA.

8. Opsomer, J., Jensen, H. & Pan, S. (2002) An evaluation of the U.S. Department of Agriculture food security measure with generalized linear mixed models. J. Nutr. 133:421-427.

9. Nord, M., Andrews, M. & Carlson, S. (2003) Household Food Security in the United States 2003:2002 U.S. Department of Agriculture, Economic Research Service Washington, DC.

10. Derrickson, J. P., Sakai, M. & Anderson, J. (2001) Interpretations of the "balanced meal" household food security indicator. J. Nutr. Educ. 33:155-160.[Medline]

11. Studdert, L. J., Frongillo, E. A. & Valois, P. (2001) Measuring household food insecurity in Java during Indonesia’s economic crisis. J. Nutr. 131:2685-2691.[Abstract/Free Full Text]

12. Gulliford, M. C., Mahabir, D. & Rocke, B. (2003) Food insecurity, food choices, and body mass index in adults: nutrition transition in Trinidad and Tobago. Int. J. Epidemiol. 32:508-516.[Abstract/Free Full Text]

13. Frongillo, E. A. (2003) Commentary: assessing food insecurity in Trinidad and Tobago. Int. J. Epidemiol. 32:516-517.[Free Full Text]

14. Nord, M. & Bickel, G. (2002) Measuring Children’s Food Security in U.S. Households, 1995–99 2002 U.S. Department of Agriculture, Economic Research Service Washington, DC.

This article has been cited by other articles:

				S. I. Kirkpatrick and V. Tarasuk Food Insecurity Is Associated with Nutrient Inadequacies among Canadian Adults and Adolescents J. Nutr., March 1, 2008; 138(3): 604 - 612. [Abstract] [Full Text] [PDF]

				J. Coates, P. E. Wilde, P. Webb, B. L. Rogers, and R. F. Houser Comparison of a Qualitative and a Quantitative Approach to Developing a Household Food Insecurity Scale for Bangladesh J. Nutr., May 1, 2006; 136(5): 1420S - 1430S. [Abstract] [Full Text] [PDF]

				J. Coates, E. A. Frongillo, B. L. Rogers, P. Webb, P. E. Wilde, and R. Houser Commonalities in the Experience of Household Food Insecurity across Cultures: What Are Measures Missing? J. Nutr., May 1, 2006; 136(5): 1438S - 1448S. [Abstract] [Full Text] [PDF]

This Article Abstract Full Text (PDF) Online Supporting Material Purchase Article View Shopping Cart Alert me when this article is cited Alert me if a correction is posted Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Wilde, P. E. Search for Related Content PubMed PubMed Citation Articles by Wilde, P. E.