Association Between State Supplemental Nutrition Assistance Program Policies, Child Protective Services Involvement, and Foster Care in the US, 2004-2016

Key Points Question Are state Supplemental Nutrition Assistance Program (SNAP) policy options associated with rates of Child Protective Services involvement and use of foster care services in the US? Findings This cohort study including all 50 states and the District of Columbia noted that adoption of SNAP policies increased from 2004 to 2016 and, accompanying the increases, substantiated reports of childhood neglect decreased. In instrumental variables models, policies to operate through SNAP caseloads were identified. Meaning The findings of this study suggest SNAP policy options that increase the generosity and stability of household resources may yield valuable population health returns by preventing child maltreatment and the need for costly child welfare interventions.


Data Sources and Coding
eTable 1 lists the sources of data for the study and provides details on missing years and states. We used two sources for the SNAP policy variables, the SNAP Policy Database 1 and the SNAP State Options Reports 2 for all 50 states and the District of Columbia. The SNAP Policy Database ended in 2016 thus setting a limit on our analysis sample. We hand-coded the SNAP State Options Reports 2 and discovered discrepancies between policies reported in the tables and those reported in the graphs. In some cases, information on the policy was missing for some years, and we interpolated the missing information as having the same policy as the previous year.
Identifying the effect of SNAP policies is complicated by measurement error and multiple policy changes. As a result, we were guided by previous research by Ganong and Liebman 4 who created averages of SNAP policies. They argue that there may be measurement error in the year that a state implemented a given policy. Like Ganong and Liebman, we also found that states have adopted multiple policies in the same year. Note that Ganong and Liebman used the average number of policies whereas we use the total count since our analysis focused on a subset of all state-controlled SNAP policies. Let

= ∑
where p is the given policy, s is each state and the District of Columbia, and t is the year. Exhibit A1 shows the specific policies included in these income generosity and disqualification count measures.
We also analyzed the components of for each separate state policy. The changes in these variables are listed in Table 1

Estimation Methods
All estimates were performed using STATA version 16.1. All hypothesis tests are twosided tests with statistical significance being p<.05. We use two methods to estimate the impact of state SNAP policies on child maltreatment and foster care outcomes: two-way fixed effects and Instrumental Variables models. The two-way fixed effects approach was used by Ganong and Liebman 4 to estimate the effect of SNAP policy changes on SNAP caseloads and outcomes.
Our model is specified as: Asians, and other races; share of Hispanics of any race; the log of real personal income; and child population by age. We also include controls for AR response, cases screened out, and the missing screen out variable as described above.
In Figure 3 in the paper, we regress case rates per 100,000 population of the child maltreatment and foster care outcomes (2) where is one of eight child maltreatment or foster care measures per 100,000 population of children in the state. The covariates, are the same as those used in equation 1. We also included estimates of each separate policy in place of the variable. Essentially, the estimated parameter  is the association between SNAP income generosity policies and child welfare outcomes. We posit that as a state becomes more generous, the magnitude of  will be negative and the number of children with substantiated maltreatment reports or children placed into foster care will decrease. All standard errors are clustered at the state level.
We then estimated a fixed effects instrumental variables model of the association of SNAP caseloads in households with children on our child maltreatment and foster care outcomes using two-stage least squares (TSLS) estimation. This approach assumes that SNAP income generosity policies affects child outcomes only through SNAP caseloads. The instrumental variables estimate is the local average treatment effect of SNAP income generosity policies on CPS and foster care outcomes in states that decided to adopt them. Instrumental variable methods are widely used in economics to correct for this kind of selection bias. An instrument is a variable correlated with SNAP caseloads but uncorrelated with the outcomes of involvement with child protective services and foster care placements. The first stage of the instrumental variable model regresses the logarithm of SNAP caseloads on state income generosity policies and is given in equation (3): Where is an exogenous instrument for SNAP caseloads measured as households with children. All covariates are defined as above. The fitted values from this regression are included as a regressor in the second stage regression where CPS and foster care are the outcomes given in equation (4): Essentially, ̂ is a measure of SNAP caseloads that are no longer biased by unobserved factors under the instrumental variables assumptions. All standard errors are clustered at the state level.

Impact of Policies on SNAP Caseloads
We posit that SNAP income generosity policies will increase caseloads. eTable 2 shows the results. We find that each additional income generosity policy increases SNAP caseloads and recipients by 4 to 5 percent. We find that excluding child support from income increases SNAP caseloads as much as 5 to 8%. The use of simplified reporting by states also significantly increases SNAP caseloads by 7 to 11%. Neither BBCE for income nor transitional benefits for households leaving TANF have a significant impact on caseloads. The results in eTable 2 are the first-stage of the TSLS regression.

Robustness Checks
As a first robustness check, we estimated the impact of caseloads on child maltreatment