Data sources Directory UMM :Journals:Journal of Health Economics:Vol19.Issue1.Jan2000:

includes children with siblings that are eligible for the expansions might lead to contaminated estimates of the true policy effect. For example, if comparison group children with expansion-eligible siblings were more likely to move from private insurance to uninsurance than comparison group children without such siblings as a result of the expansions, the estimated difference between the target and comparison groups in the probability of having private coverage at the first and last interviews would be too low. This downward bias would occur because any decline in the probability of having private coverage at both interviews due to crowd-out for children in the target group would be offset by reductions in this probability due to spillover for children in the comparison group. Since our crowd-out estimate is derived from coefficients in this regression equation, we could underestimate the crowd-out effect if spillover were occurring. The same type of bias could be expected in the equations modeling movement from private coverage to uninsurance at the last interview. In such a case, we would estimate a difference between the target and comparison groups that was in part attributable to the target group’s eligibility for the program and only partly attributable to the expansions’ detrimental effect upon the comparison group siblings. In order to assess whether our estimates were, in fact, contaminated by such a spillover effect, we examined each case in our sample in which a target group child had at least one sibling in the comparison group. Of the 112 families in our sample for which there were children in the target group with siblings in the comparison group, only two families exhibited a pattern similar to the one detailed above and only one of these had their children enrolled in Medicaid due the expansions per se. 12 As a result of this assessment, we concluded that the presence of target-child siblings in the comparison group does not significantly contaminate our results.

4. Data sources

Ž The SIPP’s core questionnaire questions repeated in each wave of the inter- . viewing process is built around labor force participation, public program partici- 12 The other family’s children were eligible through non-expansionrnon-AFDC Medicaid. The first family had two of four children eligible for the program, the second had three of five children eligible Ž . two in the target group and one in the comparison group . We cannot, however, determine whether the children in the comparison group became uninsured as a result of the expansions or as a result of some change in family circumstances that would have led to this movement from private coverage to Ž uninsurance even in the absence of the expansions. In only three additional families five total out of . 112 were there intrafamily differences in children’s health insurance status. In two families, certain children moved from private to Medicaid while the other child remained in private coverage, and in one family, a comparison group child moved into Medicaid while the target group child became uninsured. Ž . pation e.g., Medicaid and AFDC , and income questions. It also includes informa- tion about the health insurance coverage of each person in each sample household. The SIPP survey design is a continuous series of nationally representative panels and uses a 4-month recall period; individuals answer questions about the preceding 4 months. The 1990 panel follows individuals in 26,000 households for Ž . a period of 32 months eight interviews . The actual initial interviews of the 1990 SIPP panel are staggered over the period February through May 1990, 13 with one-fourth of the panel interviewed each month. Rather than use data from each month, we chose to use the data from the month immediately preceding actual Ž . interviews because analysts Young, 1989 have found that individuals tend to report transitions as occurring during that month even if they actually occurred during an earlier month in the recall period. This phenomenon, known as seam bias, makes data from the month immediately preceding the interview more reliable than data from the other months of the recall period. 4.1. Data preparation and modeling In order to use the SIPP data for this analysis, we create a number of new variables. First, we use data on the relationships within a household to create and characterize household units not defined on the SIPP. In particular, we created Ž . Ž filing units a subset of the family for Medicaid, and health insurance units also a . subset of the family for private insurance. We then created variables that characterize these units along a number of dimensions including family size, Ž family type two-parent, single-parent, child living with related family members, . etc. , family income, and labor force participation of the high-earning parent. In cases where both parents have exactly the same earnings, a random parent is assigned high-earner status. Second, since individuals can report multiple types of health insurance cover- age, we instituted a hierarchy to identify the primary source of coverage for people reporting more than one type. 14 We then grouped the different health insurance Ž coverage types into four groups: private coverage both employer-sponsored and . Ž non-group ; Medicaid including those who report both private coverage and . Ž . Medicaid ; other public coverage CHAMPUS and Medicare ; and uninsured. It is important to note that uninsured children are defined as those who do not report any other type of health insurance coverage. 13 Those individuals interviewed in May 1990 regarding insurance coverage in April of 1990 could, Ž . in fact, have been eligible for the expansions in the month of April their first interview . Because it generally takes 2 months to verify eligibility and become enrolled in the Medicaid program, we are not concerned that this issue seriously compromises our analysis. 14 This hierarchy was Medicaid and any other; employer-sponsored coverage and any other; other private coverage and any other; other public coverage; and uninsured. 4.2. Analysis file In developing our analysis file, we first identify children born after September 30, 1983 and children born between September 30, 1978 and September 30, 1983. We then identify children living on their own and children living with unrelated persons and exclude them from the analysis. We do this in order to identify Ž . appropriate family level characteristics e.g., family income and structure . After these exclusions and after excluding children in families with incomes above 185 of the federal poverty level and those with Medicaid coverage at the first interview, the analysis file contains 2587 children with observations at both first and last interviews of the panel. Finally, we use the longitudinal weights devel- oped by the Census Bureau to account for any SIPP attrition bias. 15

5. Results