The Effects of Participation in the Supplemental Nutrition Assistance Program on the Material Hardship of Low Income Families with Children

Transcription

1 National Poverty Center Working Paper Series #11 18 May 2011 The Effects of Participation in the Supplemental Nutrition Assistance Program on the Material Hardship of Low Income Families with Children H. Luke Shaefer and Italo Gutierrez, University of Michigan This paper is available online at the National Poverty Center Working Paper Series index at: Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the view of the National Poverty Center or any sponsoring agency.

2 The Effects of Participation in the Supplemental Nutrition Assistance Program on the Material Hardship of Low-Income Families with Children April 29, 2011 H. Luke Shaefer Assistant Professor University of Michigan School of Social Work Italo Gutierrez Doctoral Candidate University of Michigan Department of Economics Abstract This study examines the effects of participation in the Supplemental Nutrition Assistance Program (SNAP) on measures of the material hardship of low-income families with children, beyond food insecurity. Using the panels of the Survey of Income and Program Participation, we utilize a bivariate probit approach that exploits variation in states SNAP eligibility policies over time as instruments to identify the treatment effects of interest. In our main models, we use changes in state recertification period lengths and use of biometric requirements as instruments. We also report on models that use state-year SNAP take-up rates as an instrument. We find a statistically significant and substantively large negative relationship between SNAP participation and both food insecurity and the ability of families to pay essential household expenses, particularly, housing expenses. We find some, although not conclusive, evidence that SNAP decreases the probability that families will fall behind in their utility costs. We thank Sheldon Danziger, Mark Nord, Jeff Smith, Alisha Colemen-Jensen, Katie Fitzpatrick, and Matthew Rutledge for comments on previous drafts. We also thank seminar participants at the University of Wisconsin-Madison s Institute for Research on Poverty and the University of Chicago School of Social Service Administration for helpful comments. This study was funded by a cooperative research contract ( ) between the National Poverty Center (NPC) at the University of Michigan and the U.S. Department of Agriculture, Economic Research Service (ERS) Food and Nutrition Assistance Research Program (FANRP). The ERS project representative is Alisha Coleman-Jensen. The views expressed are those of the authors and not necessarily those of the NPC, ERS, or USDA. 1

3 Introduction This study examines the effects of participation in the Supplemental Nutrition Assistance Program 1 (SNAP) on the material hardship of low-income families with children. A primary goal of SNAP is to reduce food insecurity among recipients, and recent studies using an instrumental variables approach have found that SNAP reduces food insecurity (Mykerezi & Mills, 2010; Ratcliffe & McKernan, 2010). Beyond food insecurity, however, there is little research on the effects of SNAP participation on other measures of material hardship. We utilize an instrumental variables approach within a bivariate probit model to examine the effects of SNAP participation on the material hardship of low-income families with children, focusing specifically on outcomes related to the ability of households to pay essential expenses such as housing and utility costs. Data are drawn from the 1996, 2001 and 2004 panels of the Survey of Income and Program Participation (SIPP). As instruments, we use changes in state SNAP policies between 1998 and 2005 in the length of recertification periods and biometric requirements. We also report on models that use state-year SNAP take-up rates as an instrument. We find that SNAP reduces the risk of food insecurity by 43% and the risk that families will experience difficulty meeting essential household expenses by 59%. In particular, SNAP reduces the risk that families will fall behind in their rent or mortgage by 39%. We find some, although not conclusive, evidence that SNAP decreases the probability that families will fall behind in their household utility costs. Our results are robust to different sensitivity checks regarding sample selection, selection of instruments, and definitions of SNAP participation. Background SNAP benefits were received by 44 million individuals in December Food security, a primary outcome used by the USDA to evaluate SNAP, is defined as access by all people at 1 Formerly the Food Stamp Program. 2

4 all times to enough food for an active, healthy life, while food insecurity is the absence of food security (Nord, Kabbani, Tiehen, Andrews, Bickel & Carlson, 2000). In 2009, 17.4 million households (14.7 percent of all households) were food insecure, and these households included 17.2 million children (23.2 percent of all children) (Nord, Coleman-Jensen, Andrews & Carlson, 2010). SNAP participation may reduce other aspects of material hardship as well by allowing recipients to reallocate resources originally directed toward the purchase of food to other essential expenses such as housing and utility costs. In recent years, scholars have increasingly analyzed measures of material hardship as alternatives to the official poverty line for assessing the well-being of low-income families (Cancian & Meyer, 2004; Heflin, Sandberg & Rafail, 2009; Mayer & Jencks, 1989; Nolan & Whelan, 2010; Sullivan, Turner & Danziger, 2008; USDHHS, 2004; Wu & Eamon, 2010). Such measures employ direct indicators of consumption and physical living conditions to examine whether families meet certain basic needs (USDHHS, 2004, p. V, see also Mayer & Jencks, 1989). Wu and Eamon (2010) examine the prevalence of material hardship among low-income families with children. They find that receipt of public benefits, including SNAP, is associated with an increased risk of material hardship among low-income households with children, based on four measures: inadequate housing, inability to meet basic expenses, unmet medical/dental care, and food insufficiency. However, Wu and Eamon (2010) do not nor does any other study to our knowledge address complicated selection issues related to the decision by families to participate, which are the likely source of these positive correlations. Selection issues are well documented in the literature on SNAP participation and food insecurity. Wilde (2007) and others report that among low-income households, those receiving 3

5 food stamps are considerably more likely to report food insecurity than nonparticipating households (See also Jensen, 2002). Gibson-Davis and Foster (2006) write, the problem with analyzing the impact of food stamps on food insecurity is that unmeasured or unobserved characteristics are likely correlated with both food stamps use and food security (p.94, see also Gundersen & Kreider, 2008; Gundersen & Oliveira, 2001; Jensen, 2002; Ribar & Hamrick, 2003, Wilde, 2007; Wilde & Nord, 2005). Several recent studies have used more sophisticated techniques, including an instrumental variables approach, and found a negative relationship between SNAP participation and food insecurity (Bartfeld & Dunifon, 2006; Borjas, 2004; Mykerezi & Mills, 2010; Nord & Golla, 2009; Ratcliffe & McKernan, 2010; Yen, Andrew, Chen, & Eastwood, 2008). Borjas (2004) uses state variation in the treatment of immigrants prior to and following the 1996 welfare reform to test the effects of participation in means-tested programs on the food insecurity of immigrants. He concludes that the evidence suggests an important [negative] causal link between public assistance and food insecurity (p.1439). However, Inoue and Solon (2010) report that Borjas based the standard errors for his instrumental variables models on the wrong asymptotic covariance matrix, so it is not clear that his instrumental variable results are statistically significant. Yen et al. (2008) use the National Food Stamp Program Survey, a small survey of income eligible households conducted by Mathematica Policy Research, Inc. They utilize an instrumental variables approach to examine the effects of SNAP participation on food insecurity, utilizing instruments measuring stigma as well as cross-sectional variation in some state SNAP policies and state-level immigrant population share (state controls are not included in their model). They find a statistically significant and negative association between SNAP 4

6 participation and food insecurity. However, their data may not be representative, as households receiving SNAP in their sample were less likely to report food insecurity than eligible households not receiving SNAP. Ratcliffe and McKernan (2010) pool data from the panels of the SIPP and take a bivariate probit approach to measure the effects of SNAP on food insecurity among households who are both below 150 percent of poverty and have low assets. As instruments, they use changes over time in state outreach spending per capita, use of biometric requirements, and a term interacting states treatment of immigrants with noncitizen immigrant status of household heads. They find that SNAP participation substantially and statistically significantly decreases the risk of household food insecurity. 2 Mykerezi and Mills (2010) use cross-sectional data from the 1999 PSID and utilize an endogenous treatment effect model using static state-level error rates in benefits payments as an instrument (without including state-level controls). They also examine the impact on food insecurity of self-reported loss of benefits reportedly due to a decision by a government office. Like Ratcliffe and McKernan, they find that SNAP participation has a substantial and statistically significant negative effect on food insecurity. To our knowledge, the current paper is the first to use an instrumental variables approach to examine the effects of SNAP participation on measures of material hardship beyond food insecurity. We analyze the ability of families with children to pay their essential expenses including housing and utility costs, as well as the probability of having a telephone line 2 Ratcliffe and McKernan (2010) construct their sample in a problematic way. SIPP households only report on the main food insecurity measures once for a four-month wave, while they report on SNAP receipt in each month of the wave. Ratcliffe and McKernan treat each reference month that respondents are in the wave as a unique observation, even though the food insecurity outcome is the same across the wave. This artificially inflates their sample and hence reduces their standard errors. They do, however, report that they re-ran models with only one observation per household, and still find that SNAP participation reduces the risk of food insecurity. But they do not report if the smaller sample is robust to sensitivity tests. 5

7 disconnected. We hypothesize that SNAP participation should reduce the material hardship of recipients by allowing them to reallocate resources originally directed toward the purchase of food to other essential expenses. In addition, previous studies have not used an instrumental variables approach to examine the effects of SNAP participation specifically on the food insecurity of families with children, the largest group of SNAP beneficiaries. During fiscal year 2009, households with children made up 50 percent of all SNAP households (USDA FNS, 2010). SNAP serves a heterogeneous population, and the program s impacts may be different for the various sub-groups, such as individuals and families without children and the elderly. We hypothesize that participation in SNAP should reduce the risk of food insecurity faced by low-income families with children. Data and Methods Data are drawn from public use files of the SIPP, collected by the U.S. Census Bureau. SIPP interviews are conducted every four months about each individual in the household for each intervening month, gathering data on demographics, income sources, public assistance program participation, household and family structure, and jobs and work history. We pool data from the 1996, 2001, and 2004 panels of the SIPP, each of which is 3-4 years long. 3 An analysis of eight major nationally representative surveys that measure income and program participation finds that the SIPP does a superior job of measuring the income of poor households and measuring public program participation (Czajka & Denmead, 2008). Underreporting of benefits receipt in household surveys (in which respondents do not report public benefits that they have accessed) remains a limitation (Gundersen & Kreider, 2008; Meyer, Mok, & Sullivan, 2008). However, the SIPP does relatively well in terms of SNAP reporting rates. 3 A few states (Maine, Vermont, Wyoming, North Dakota and South Dakota) were not uniquely identifiable in the 1996 and 2001 panels, so observations from these states are dropped because they cannot be matched with state SNAP policy data (as is done by Gruber and Simon, 2008; and Ratcliffe et al., 2008). 6

8 Meyer et al. (2008) estimate that the SIPP reported 87.7 percent of SNAP benefits for 1998, 84.8 percent for 2003, and 82.9 percent for 2005, the years in our study frame that include the material hardship measures. Our sample includes families with resident children under 18 and at least one adult family member over 18. Rather than trying to simulate SNAP eligibility, we follow Mykerezi and Mills (2010) and Ratcliffe and McKernan (2010) and restrict our sample to families based on lowincome. 4 We restrict our main sample to families with an average gross income at or below 150 percent of poverty during the months covered by our outcome variables, using the monthly family-level poverty thresholds in the SIPP. We use this gross income threshold for several reasons. First, if we restricted our sample by simulated eligibility, a significant proportion of families reporting SNAP receipt would be coded as ineligible. This may relate to limitations in comparing income and assets reported in the SIPP with state eligibility calculations, or may be a result of fluctuating family incomes and assets following initial certification. More importantly, however, there are concerns that income may be endogenous to participation. Families near the eligibility threshold may modify their earnings or assets in ways that makes them eligible (Ashenfelter, 1983). Thus, restricting to eligible families may bias estimates of program effects. Our key outcome variables are drawn from the adult well-being topical modules administered once per panel in wave 8 of the 1996 panel (administered during 1998), wave 8 of the 2001 panel (administered during 2003), and wave 5 of the 2004 panel (administered during 2005). The SIPP is the primary source of nationally-representative data on material hardship in 4 Unlike Ratcliffe et al. (2008) and Ratcliffe and McKernan (2010), we and Mykerezi and Mills (2010) do not restrict by household assets. Doing so only marginally changes the sample composition and requires merging in assets data collected in other waves, which may not be representative of the household s circumstances when they applied for SNAP or when they completed the topical module with the material hardship questions. 7

9 the US (Bauman, 1999; Beverly, 2001; Heflin et al, 2009; USDHHS, 2004; Wu & Eamon, 2010). While material hardship has gained prominence in recent years as a way to assess the well-being of poor families, there remains no official measure (Cancian & Meyer, 2004; USDHHS, 2004). Our first non-food measure indicates whether a household broadly had difficulty meeting its essential household expenses. Families were asked Next are questions about difficulties people sometimes have in meeting their essential household expenses for such things as mortgage or rent payments, utility bills, or important medical care. During the past 12 months, has there been a time when (YOU/YOUR HOUSEHOLD) did not meet all of your essential expenses? Families in households that responded affirmatively were classified as having trouble meeting essential expenses. We also examine three additional, more specific measures which ask 1) whether a household reported falling behind on their rent/mortgage; 2) whether they reported falling behind on their utility bills; and 3) whether they reported having a phone line disconnected. 5 The SIPP adult food security measures do not conform exactly to the official USDA food security scale; however, they have been used in several studies and are closely related to the official food security measure (Bitler, Gundersen, & Marquis, 2005; Gundersen & Gruber, 2001; Ratcliffe & McKernan, 2010). 6 Families are classified as food insecure if they responded affirmatively to at least two of a set of questions that can be used to measure food insecurity in the Adult Well-being Topical Module in the SIPP. See the Appendix for further details. We also report results for a related outcome, food insufficiency, used in previous studies. Households are 5 We follow Beverly (2001), who argues it is better to treat material hardship variables independently rather than create an index that combines them (see also Cancian & Meyer, 2004; Heflin et al., 2009). 6 Nord (2006) reports that an assessment of the food security items using statistical methods based on the Rasch measurement model indicated that relative item severities were very nearly identical to those in the 1998 Current Population Survey Food Security Supplement, and analysis of CPS data comparing the SIPP scale with the standard U.S. Food Security Scale indicated that the SIPP scale was reasonably reliable and only moderately biased (p. 2). 8

10 considered food insufficient if they report that members sometimes or often did not have enough to eat. SIPP households only report on the main food insecurity measures once, in reference to the four months of the wave. One limitation of our material hardship outcomes is that they are measured at the household level rather than the family level. Households are defined as a group of persons who occupy a housing unit (SIPP User Guide, p.10-9). This includes families but also two or more unrelated families co-habiting (although it does not include group quarters). Families include multi-generational families, but exclude co-habitors, borders, or a separate, unrelated family sharing the same house. Most often, households and families are identical; however, in some cases the composition of the two will be different, and this is more likely among lower-income families than in the population as a whole. We use the family as our unit of analysis because it allows for more accurate modeling of SNAP participation. 7 In sensitivity analyses, we run our main models with an indicator for the presence of non-related individuals in the household, and our findings are robust. Instrumental variables methods attempt to correct for hidden bias or unobservable differences between participants and non-participants. Because our key variables of interest are endogenous dummy variables measuring SNAP participation and material hardship, we utilize a bivariate probit model (see Heckman 1978; Greene 1998; Angrist & Pischke, 2009). 8 We posit 7 SNAP benefits are provided to assistance units, which consist of individuals who live together and buy or prepare food together. It is our understanding that there is significant variability in the designation of these units within households, often resulting in multiple assistance units within the same household. We assume that applicants engage in profit maximizing behavior by defining their unit in the way that maximizes benefits receipt. 8 We follow the large majority of studies in the SNAP and food insecurity literature in using a non-linear IV model (Gundersen & Oliveira, 2001; Mykerezi and Mills, 2010; Ratcliffe & McKernan, 2010; Yen et al., 2008). Under the assumption that the random components of the model are normally distributed, our bivariate probit estimators will be more efficient than two-stage least square (2SLS) estimators. In fact, our 2SLS estimates of the effect of SNAP on food insecurity have standard errors that are five times as large as the ones we obtain using our bivariate probit approach, and do not achieve statistical significance. Alternatively, in order to gain efficiency in the 2SLS estimator, in a sensitivity analysis we follow a method outlined by Angrist and Krueger (2001). They suggest that non-linear 9

11 that (potentially) eligible families decide to participate in SNAP by comparing costs and benefits using a net benefit function or latent index that is linear in covariates with a random error term distributed normally with variance equal to one. 9 Families participate as long as the net benefits are positive, and thus the participation equation can be written as:,, 1,, β Z, θ ϵ 0 (1) Equation 1 models the probability of SNAP participation for family i in state j at time t. In our main specification,,, is measured as SNAP participation in the final month of the wave because respondents reporting is known to be most accurate in the month closest to the interview (Moore, 2007). 10,, represents a vector of demographic and geographic characteristics that have been shown to be related to SNAP participation and/or material hardship. We include a count variable for the number of children in the household 11 and an indicator for household headship (headed by husband/wife, single-male headed, and singlefemale headed). We also control for the highest level of schooling reported by an adult family member, and include an indicator for the presence of a full-time worker. Race and ethnicity, age (and age squared), sex, metropolitan residence and U.S. citizenship of the family head are included. We also control for the state-month unemployment rate. Finally, dummies for state, year, and calendar month are included in all models. fitted values (e.g. from a probit) of the endogenous dummy (using the excluded instruments and all other covariates in the model) may be used as an instrument in 2SLS, as long as a linear model is used to generate first-stage predictions of the endogenous dummy variable from these nonlinear fitted values and all other exogenous covariates in the second-stage equation (see Angrist & Krueger, 2001, p.80, see also Angrist & Pischke, 2009, p ). The benefit of this approach is that consistency of the estimates does not depend on the correct specification of the participation equation. Moreover, 2SLS estimates give the best linear approximation to the (local average) treatment effects. Our results using this method are very similar to our bivariate probit estimates (as reported in Table 4), although with larger standard errors, as expected. 9 The parameters in the bivariate probit model can only be identified up to scale. Placing a restriction on the variances of the random components allows for unique identification of the parameters. 10 In sensitivity analyses we utilize alternative definitions of SNAP participation, including requiring SNAP receipt in all months of the wave, any month of the wave, and just the first month of the wave. Results are robust. 11 Originally we used three age categories, but consistency in the point estimates led us to collapse this variable into one. 10

12 Vector Z, includes our state policy variables, which are predicted to increase the cost of participation. Policy data by state-year are drawn from a dataset prepared by USDA ERS researchers. Our first instrument is the proportion of assistance units within each state with a recertification period of 3 months or less, by state-year. Numerous studies have shown that the length of recertification periods has a significant effect on SNAP participation (Hanratty, 2006; Kabbanni & Wilde, 2003; Ratcliffe, McKernan and Finegold, 2008; Ribar, Edelhoch & Liu, 2008; Schmeiser, 2011). Recertification periods typically range between 1 and 12 months, and in some cases longer. As a result of federal encouragement and state policy changes, the late 1990s saw a large increase in the proportion of recipients especially those in assistance units with earners recertified within three months. This proportion, though, fell considerably after Our second instrument is the use of biometric technology (mostly fingerprinting of applicants), used with the goal of reducing fraud. We hypothesize, as in Ratcliffe et al. (2008), that this should discourage program participation. Biometrics technology was used by Texas, Arizona, and New York throughout our study period, but was introduced in California halfway through the study period. Massachusetts implemented biometrics and then ended it during our study period. While this instrument relies on change in only two states, biometric requirements have a significant impact on the probability of SNAP participation. 13 Our outcomes of interest are several measures of material hardship (, ). Conceptually, families report experiencing a hardship if an underlying latent index of material unmet needs is above a certain threshold (which can be set to zero without loss of generality). Thus, we have: 12 In 1998, an average of 28.0 percent of states caseloads had a recertification period of three months or less. This fell to 10.9 percent in 2003 and 2.8 percent in percent of our sample was subject to biometric requirements in This increased to 33.0 percent in 2003, and fell to 23.9 percent in Because of concerns with the limited nature of this instrument, we estimate alternative specifications using these two instrument separately, and a third specification utilizing a measure of stateyear take-up rates (estimates by Mathematica Policy Research, Inc.) as an instrument, discussed below. Our point estimates across these specifications are substantively similar. 11

13 ,, 1,, γ δsnap,, v 0 (2) The random component v is also assumed to be distributed normally with variance equal to one.,, is the same vector of control variables included in equation 1, with, excluded. We hypothesize that SNAP participation reduces the probability of experiencing material hardship and thus δ 0. We also hypothesize that the correlation between ϵ and v is positive ( 0). In other words, families who are more likely to experience material hardships are also more likely to participate in SNAP. This is the source of the bias that, if not accounted for, produces positive associations, as has been documented between SNAP and food insecurity. Our model is identified given that our instruments (Z, ) are independent of ϵ and v, and given the distributional assumption of the random components. The average causal effect of SNAP participation in percentage points on the probability of experiencing a given outcome of material hardship is estimated using the following formula:,,,,, SNAP,, 1,,,,, SNAP,, 0 Φ,, γ δ Φ,, γ (3) Thus, we calculate the marginal effect of SNAP participation on material hardship for each individual in our sample as the difference between the predicted hardship with and without SNAP, and then report the mean marginal effect over our sample. Alternatively, we also estimated the average causal effect of SNAP participation in percentage change on the probability of material hardship, given by 14 :,,,,,SNAP,,,,,,,SNAP,,,,,,,SNAP,, (4),,,, Results Table 1 presents weighted summary statistics. Column 1 reports means for families with incomes above 150 percent of poverty. The next three columns are restricted to families below 14 Standard errors for average causal effects were calculated using 500 bootstrap (within state) replications. 12

14 150 percent of poverty, divided into 5,421 observations for families not reporting SNAP (in column 3) and 3,145 observations for families reporting receipt of SNAP benefits (in column 4). Only 13.2 percent of families with incomes above 150 percent of poverty lived in households that reported difficulties meeting essential expenses, and only 6.2 percent reported food insecurity. Among families at or below 150 percent of poverty, 29.9 percent of those not reporting SNAP and 47.7 percent of families reporting SNAP lived in households reporting trouble meeting their essential expenses. Similarly, just over a third of low-income families reporting SNAP lived in households that were food insecure, while the same was true of only 20.8 percent of non-snap families. This positive association in the data between reported SNAP participation and measures of material hardship are likely the result of the selection process of what families decide to participate in SNAP. Examining the low-income analysis sample, the average monthly income of families reporting SNAP is 65.3 percent of poverty, compared to 90.6 percent of poverty for non-snap recipient families. Families reporting SNAP are far more likely to be female-headed (69.4 to 37.8 percent), and the heads of these families are more likely to be Black and less likely to be of Hispanic Origin than families not reporting SNAP. Table 2 reports on bivariate probit models with two key outcomes: food insecurity and trouble paying household essential expenses. Estimates are reported as probit coefficients, while average causal effects for the effects of SNAP participation on these outcomes are reported in Table 3. The SNAP participation equations are modeled jointly with food insecurity in column 1 and with problems meeting essential expenses in column 3. We find that, after controlling for other factors, each additional child in a family is associated with a higher probability of SNAP participation. Female-headed households are much more likely to participate than those headed by a married couple. Families in which the reference person is Black or Asian or Pacific 13

15 Islanders are more likely to participate than those in which the reference person is white, and families in which the reference person is a US citizen are more likely to participant than those with a non-citizen reference person. Increased education is associated with a decreased probability of SNAP participation, and families with 1 or more full-time workers are less likely to participate than families without. Our instruments are strong predictors of reported SNAP participation. 15 As a larger proportion of a state s SNAP caseloads are recertified in three months or less, the probability of participation decreases. Use of biometrics is also associated with a reduction in the probability of participation. Columns 1 and 3 of Table 2 also report on the correlation coefficient between the error components in the SNAP participation equation and in the material hardship outcomes. As we expected, the correlation coefficient is positive ( 0), considerably large, and statistically significant in both models. This result means that families that are more likely to report SNAP are also more likely to report experiencing food insecurity or difficulty meeting essential household expenses. 16 Columns 2 and 4 report on the effect of SNAP participation and other covariates on the latent indexes for food insecurity (columns 2) and trouble meeting essential expenses, referred to as non-food material hardship for the remainder of this section (column 4). There is significant consistency across the exogenous covariates shared by the two equations. Additional children are associated with increased food insecurity and non-food material hardship. Female-headed families are more likely to experience both outcomes than families headed by a married couple. 15 The Chi-square statistics for the null hypothesis that the excluded instruments coefficients are zero are (pvalue ) for column 1 and (p-value of ) for column 3. Also when the same models are run using 2SLS, the F-statistic associated with the excluded instruments in the first stage is 21.16, above the standard suggested cut-off value of 10 (Stock, Wright and Yogo,2002). 16 As would be expected given this, naive probit specifications that use the reported indicator of SNAP participation find positive associations between SNAP and both food insecurity and the inability to pay household essential expenses (results available upon request). 14

16 Higher levels of education and the presence of full-time workers are both associated with a lower risk of food insecurity and non-food material hardship. Families in which the reference person is black are more likely to experience both outcomes than families in which the reference person is white. Families in which the reference person is of Hispanic origin are more likely to be food insecure but not more likely to experience non-food material hardship that families in which the reference person is non-hispanic. SNAP participation has a strong statistically significant negative effect on both the latent indexes for food insecurity and for non-food material hardship. We translate those effects into average causal effects on the probability of reporting food insecurity and non-food material hardship in Table 3. We find that SNAP participation results in a statistically significant 13.2 average percentage point reduction in the probability of being food insecure, which is equivalent to an average decrease of 43 percent in the incidence of food insecurity. 17 Similarly we find a statistically significant 27.4 average percentage point reduction in the risk of non-food material hardship, which is equivalent to a 59 percent reduction in the average incidence of non-food material hardship. 18 Table 3 also reports the average causal effect (both in percentage points and in percentage change) of SNAP on a number of more specific material hardship outcomes. To our knowledge, this is the first paper to find that SNAP participation has a statistically significant negative effect on food insufficiency. In terms of percentage change, the effect reported in table 3 for food insufficiency (-42%) is substantively large and nearly identical to the effect on food insecurity. In terms of more specific non-food hardships, our results indicate that SNAP participation leads 17 The model predicts a counterfactual scenario where the average incidence of food insecurity would be if no family participates in SNAP versus an average incidence of if all families participate. 18 The average predicted incidence of non-food material hardship would be if no family participates in SNAP. It drops to only if all families were to participate in SNAP. 15

17 to a statistically significant decrease of 8.1 percentage points (or 39%) in the risk that families fall behind on their rent or mortgage. The effect associated with the risk of falling behind on household utility bills is of a similar magnitude, although not statistically significant for the effect measured in percentage points and only weakly significant (at the 10% confidence level) for the effect measured in percentage change. Finally, the point estimates for the risk of having a telephone line disconnected is positive and not statistically significant. Sensitivity Analyses In table 4, we report on a series of sensitivity tests of our estimates of the average SNAP causal effect (in percentage points) on food insecurity and difficulty meeting essential household expenses. We began by trying alternative constructions of our observed SNAP receipt variable, requiring 1) receipt in all reference months of the wave, 2) receipt in any reference month, and finally 3) receipt in the first reference month of the wave. In all cases, the point estimates remain statistically significant. Requiring participation in any reference month or the first reference month reduces the size of the point estimates somewhat. We also restricted the sample at two alternative income thresholds. Our estimates at the 175 percent threshold are highly significant. At the more-restrictive 125 percent threshold sample, the point estimates are smaller and the food insecurity outcome becomes insignificant, most likely due to the loss of statistical power because of the smaller sample size. In panel C we test the robustness of our estimates to the selection of instruments. Our findings are substantively similar and robust to the use of our two policy instruments separately. This is suggestive evidence of the validity of our estimates for the general (potentially eligible) population rather than only for a specific subset of individuals or states. Moreover, as a further test of the robustness of our estimates to the selection of instruments, we used an alternative 16

18 instrument consisting of state-year SNAP participation rates (Schirm & Castner, 2002; Cunnyngham, Castner, & Schirm, 2008; Cunnyngham, Castner, & Schirm, 2011). 19 States participation (take-up) rates might be considered to reflect comprehensive packages of state SNAP policies. States that make it easier to apply and remain on SNAP should have a higher participation rate than states than do not. Since we include state dummies in the model, identification of the causal effect is coming from changes in participation rates within states across years. Using this more comprehensive policy instrument, we find average causal effects (in percentage points) that are nearly identical to the estimates reported in table 3. Thus, we feel comfortable that our results reflect average causal effects that are valid for the population of analysis and not only for specific subgroups (or states) that would be specially affected by the selection of the instruments. In panel D, we test the robustness of our results to the functional form and distributional assumptions by reproducing our estimates using a 2SLS estimator that uses non-linear fitted values (based on all exogenous variables and our policy variables) as an instrument (as specified in footnote 7). The 2SLS estimates give the best linear approximation to the (local average) treatment effects. We find that the 2SLS results are quite similar to our bivariate probit results, although the point estimates associated with problems meeting essential expenses is somewhat smaller (20.8 percentage points compared to 27.4 points). As might be expected, though, the 2SLS estimates are less precise, leading the food insecurity estimate to be insignificant and the problems meeting essential expenses estimate to be significant only at the 10 percent level. 19 We use empirical Bayes shrinkage estimates of states SNAP participation rates published by Mathematica Policy Research, Inc. (Schirm & Castner, 2002; Cunnyngham, Castner, & Schirm, 2008; Cunnyngham, Castner, & Schirm, 2011). These estimates optimally combine direct sample estimates from Current Population Survey (CPS) and SNAP administrative data with regression estimates using predictors from decennial census, American Community Survey (ACS), income tax and other administrative data. The shrinkage estimates derived are substantially more precise than direct sample estimates from the CPS or the SIPP, the best sources of current data on households incomes and program eligibility (Mathematica 2002). 17

19 Some studies on SNAP and food insecurity control for income (Yen et al., 2008), Although this is clearly an endogenous variable, we do this in panel E of table 4 (using dummies for family income falling within 0-50%, %, and % of poverty level) and our results remain robust. We ran a specification that drops all observations with imputed values. We also ran a model controlling for an unrelated adult in the same household to account for families with unrelated co-habiting adults. Finally, we ran models with family-level weights. In all cases, our results remain robust. Discussion Because SNAP participation may allow families to reallocate resources otherwise directed toward purchase of food to other essential expenses, it can affect economic well-being in many dimensions. The prominence of SNAP among means-tested programs suggests that it should be evaluated using a broader set of material hardship outcomes than food insecurity. To our knowledge, this study is the first to use instrumental variables approach to estimate the effect of SNAP benefits on non-food measures of material hardship. Under-reporting of benefits receipt in the SIPP remains a limitation of the current study, even though the SIPP does relatively well in terms of reporting rates (Gundersen & Kreider, 2008; Meyer, Mok, & Sullivan, 2008). Unfortunately, there is currently no source of nationallyrepresentative data linking the demographic characteristics of individuals with administrative data on SNAP receipt. Thus, the current study would be impossible with any existing source of administrative data. Studies that use administrative data from one or a handful of states are limited in their generalizability and ability to use changes over time in policies as a source of identification. We find that our estimates of the effects of SNAP of food insecurity are quite similar to those reported by Ratcliffe and McKernan (2010), despite the fact that we use a 18

20 different configuration of policy instruments and a different sample selection. As previously discussed, there are now a number of studies using different data and different methods that offer evidence that SNAP reduces food insecurity. We find that SNAP participation increases substantially the likelihood that low-income families with children will meet their essential household expenses. Indeed, this effect is larger than the one associated with food insecurity. We interpret this to mean that families effectively spread their SNAP benefit across both food and non-food essential household expenses. In particular, we find robust evidence that SNAP participation decreases the risk that families will fall behind in their housing costs. We also find some (although weaker) evidence that SNAP decreases the risk that families will fall behind on their utility bills, and these point estimates are of a similar magnitude to our results associated with housing costs. We find no evidence that SNAP decreases the likelihood of a phone line disconnection. However, we wonder whether with the increasing utilization of cell phones, especially among the poor this particular measure of non-food material hardship is obsolete. A number of recent articles forward a heat or eat hypothesis, finding that as utility expenditures increase due to cold weather, low-income families reduce their food expenditures (Bhattacharya, DeLeire, Haider & Currie, 2003; Nord & Kantor, 2006). Given the results of the current study, perhaps future research can identify changes in the ways families allocate resources from SNAP based on external circumstances: do they allocate more resources toward paying utility bills and less toward food during cold seasons? Or perhaps future research could examine the extent to which SNAP buffers families from increased material hardship associated with economic shocks such as job loss or marital dissolution. Overall, we hope the current study 19

21 is a first step in an examination of the effects of SNAP on a broader set of material hardship outcomes than has previously been examined. 20

22 Table 1: Sample means > 150% <=150% of poverty Characteristics of Families of Non- All poverty SNAP SNAP (1) (2) (3) (4) Sample size 23,898 8,566 5,421 3,145 Material Hardship characteristics Food Hardship Food Insecurity in past four months Sometimes or often not enough food Non-Food Hardship Problem meeting essential expenses Did not pay full rent Did not pay full gas, oil, or electricity bills Telephone line disconnected SNAP Participation Family Characteristics Family Income as % Poverty Number of children Household structure Headed by husband/wife Male Headed Family Female Headed Family Maximum education Level Less than High school High School Some college BA degree or above Full time workers in family Live in a metropolitan area State-month unemployment rate Reference person characteristics Male Female Age Race White Black American Indian Asian or Pacific Islander Hispanic Origin US citizen Source: Authors' analyses of a pooled sample from the panels of the SIPP Note 1: Estimates are weighted. 21

23 Note 2: The sample selection criteria was: a. Observations belong to the fourth reference month only. b. Families must have a positive number of children c. The family reference person must be 19 or older. d. Information is the family-year-month level. Note 3: We used the following waves: 1996w8, 2001w8, 2004w5. These are the waves in which the adult well-being topical module is asked. Table 2: Effects of SNAP Participation on Food Insecurity and Material Hardship for Low-Income Families with Children Bivariate Probit Equations (Linear index coefficients and standard errors reported) Food Hardship SNAP Participation (1) Food Insecurity (2) Non- Food Material Hardship Problems SNAP meeting essential Participation Expenses (3) (4) SNAP coverage ** *** [0.179] [0.315] Family characteristics Number of children 0.173*** 0.069*** 0.171*** 0.091*** [0.014] [0.019] [0.014] [0.021] Married couple Headed Family Male Headed Family ** ** [0.069] [0.063] [0.069] [0.060] Female Headed Family 0.467*** 0.258*** 0.478*** 0.248*** [0.035] [0.060] [0.036] [0.071] Less than High school 0.183*** 0.158*** 0.179*** 0.109*** [0.037] [0.034] [0.039] [0.035] High School Diploma Some college *** ** *** [0.034] [0.032] [0.033] [0.053] BA degree or Advanced degree *** *** *** *** [0.056] [0.066] [0.057] [0.062] 1+ full time workers *** *** *** *** [0.028] [0.057] [0.028] [0.064] Lives in a metropolitan area [0.045] [0.049] [0.044] [0.039] State-month unemployment rate *** 0.066* [0.037] [0.028] [0.038] [0.032] Reference person characteristics 22

24 Female 0.145*** 0.132*** 0.134** 0.135*** [0.055] [0.049] [0.053] [0.043] Age ** 0.026*** *** 0.016** [0.006] [0.006] [0.006] [0.008] Age Squared *** 0.000* *** [0.000] [0.000] [0.000] [0.000] White Black 0.389*** 0.133** 0.388*** 0.200*** [0.041] [0.052] [0.042] [0.056] American Indian [0.139] [0.153] [0.138] [0.090] Asian or Pacific Islander 0.286*** *** [0.086] [0.081] [0.090] [0.111] Hispanic Origin ** [0.081] [0.062] [0.088] [0.038] US citizen 0.202*** ** 0.181*** [0.078] [0.052] [0.081] [0.053] State Policies Biometrics *** *** [0.056] [0.064] Short period recertification ** ** [0.099] [0.103] Correlation of errors terms 0.425*** 0.692** [0.113] [0.193] Observations Source: Authors' analyses of a pooled sample from the panels of the SIPP Notes: Regressions run unweighted. All estimations include state dummies, year dummies and month dummies. Standard errors [in brackets] are clustered by state. *** p<0.01, ** p<0.05, * p<0.1 23

25 Table 3: Average Causal Effect of SNAP coverage on Alternative Measures of Material Hardship In percentage points In percentage change Food Hardship Food Insecurity ** *** [0.055] [14.849] Food Insufficiency * ** [0.023] [19.749] Non-Food Hardship Problems meeting essential expenses *** *** [0.095] [16.408] Did not pay full rent *** *** [0.031] [12.582] Did not pay full gas, oil, or electricity bills * [0.071] [21.462] Telephone line disconnected [0.076] [87.900] Source: Authors' analyses of a pooled sample from the panels of the SIPP Notes: All estimations include state dummies, year dummies and month dummies. Standard errors are calculated from 500 bootstrap draws within each state. *** p-value<0.01, ** p-value<0.05, * pvalue<0.1 24

26 Table 4: Average Causal Effect of SNAP coverage on Material Hardship, Sensitivity Analyses (Effects in percentage points reported) Food Insecurity Problem meeting essential expenses A. Alternative definions of SNAP participation = 1 if participation in all reference months, 0 otherwise *** *** [0.058] [0.096] = 1 if participation in any reference month, 0 otherwise * ** [0.057] [0.102] = 1 if participation in first reference month, 0 otherwise * ** [0.063] [0.107] B. Alternative samples by Income 175% of Poverty *** *** [0.040] [0.075] 125% of Poverty * [0.065] [0.113] C. Alternative set of instruments Recertification period only ** *** [0.052] [0.091] Biometrics only *** *** [0.044] [0.076] Average state SNAP participation rate *** *** [0.044] [0.074] D. Alternative estimators 2SLS results * [0.110] [0.121] E. Other sensitivity tests Controlling for family income *** ** [0.046] [0.101] Controlling for unrelated adults *** *** [0.046] [0.085] Dropping imputed values *** *** [0.051] [0.078] Weighted regressions *** *** [0.049] [0.077] Source: Authors' analyses of a pooled sample from the panels of the SIPP Note 1: All estimations include state dummies, year dummies and month dummies. Standard errors are calculated from 500 bootstrap draws within each state. Note 2: 2SLS results using non-linear predictions of SNAP participation as instruments. See footnote 8 for further details. *** p-value<0.01, ** p-value<0.05, * p-value<0.1 25