Analytical methods
This paper measures the effect of variation in premium across individuals, states, and over time on insurance coverage. The cut-off rule for assigning CHIP eligible children to premium groups varies across states. We use “premium group” to denote a group of families with children who, based on state-specific income and age eligibility rules, fall into a CHIP premium bracket. We evaluate enrollment in the lowest two premium groups because of the small number of observations in higher premium groups. The income cut-off between premium groups is based on percentage of the federal poverty line (FPL) and varies across states to reflect local cost of living and budget availability. However, in every state with a two tier premium structure, the underlying principle is the same: for a given child age, families with family income
I below a state specific income cut-off pay a smaller CHIP premium compared to families with income at or above the cut-off. Thus, the premium payment, as a function of
I, contains a jump at the income cut-off between the premium groups. This discontinuity in premium payments fits within the conceptual framework of the Regression Discontinuity design. The method assumes threshold randomization, implying that, within small vicinity around the cut-off, children in the low and high premium groups near the cut-off are the same in terms of all observable and unobservable characteristics except the premium payment. The RD method has been shown to identify mean treatment effects for a subgroup of the population without having to rely on arbitrary assumptions about functional form and exclusion restrictions [
12].
Since MEPS (described in the next section) is designed to produce nationally representative estimates, we first evaluate data on all states with a two tier premium structure as of January 2003. Premium level P
i varies across premium groups within a state and across states. The cross-sectional RD equation is defined as:
where
y
is
is the insurance enrollment outcome for child
i in state
s. S
*
is the group of states with two premium groups.
is an indicator function denoting assignment of child
i to the high premium group in state
s. ℑ
s
denotes the subsample included in the estimation such that
where
h is the interval around the cut-off. The coefficient on the high premium group indicator is of primary interest and is defined as:
. This linear function captures the change in insurance status in response to premium change. The parameter
β
0 captures the average enrollment for children above the income threshold level, and
β
p
is the linear effect of the premium on the likelihood a child is insured. We control for state-specific fixed effects with a set of dummies
D
s
. The state dummies allow low premium groups to be different across states, while the high premium groups can differ only in
. Since income
I is the only systematic determinant of the premium fee and income influences insurance choice [
13,
14], the inclusion of a smooth function
which is continuous at the state income cut-off
solves the endogeneity issue [
12]. Thus, after controlling for differences in premium and income and for state-specific fixed effects, we presume no other factors affect the insurance outcomes of children. We therefore estimate a “sharp” RD model since the “simulated” insurance status is a deterministic function of family income.
We apply a Difference-in-Differences (DD) method to states with a two tier premium structure as of December 2003 to evaluate the impact of premium increases over time:
where S
**
is the relevant group of states. y
ist
is the insurance enrollment status of child i in state s at time t. ℑ
s
is the subsample included in the estimation. The state dummies (D
s
) control for state heterogeneity. Premium, time and premium-time interactions are modelled as linear specifications. The set of premium variables β
1, γ
1 and δ
1 capture the linear premium effect for being in the high premium group, for being in the second period, and their interaction. The variables β
0, γ
0 and δ
0 capture, respectively, the average enrollment for the high premium groups, in the post premium increase period, and for high premium groups in the second period.
z
is a set of other variables that affect insurance status.
We provide summary statistics on child age and health for the groups above and below the income cut-off and compare their means using data on the largest state in our data set to evaluate whether RD design is appropriate for the purposes of our study.
Data
The data for the analysis come from the 2003 MEPS panel. MEPS is designed to produce nationally representative estimates for insurance coverage, medical expenditure, and health care use. It provides detailed data on a wide range of health, demographic, and socioeconomic characteristics [
15]. We collected data on CHIP premiums and eligibility from program websites for all states and the District of Columbia. The premium data were merged by state to the 2003 full-year consolidated MEPS files. Ethical approval has been obtained from the Institutional Review Board (study number 05–0944 at the University of North Carolina at Chapel Hill).
As of January 2003, 18 CHIP programs with at least a two tier premium structure charged families up to $61 per child per month if in the low premium group and up to $77 if in the high premium group. Eleven states increased premiums over the course of 2003. By December 2003, an additional state adopted a two-tier premium structure, so these 19 states are used in the longitudinal DD analysis.
Each state’s premium information is used to assign the premium amount that the family unit will face to cover one child for one month. We have not included in our analysis states that charge annual premiums as we seek to evaluate the impact of monthly CHIP premium contributions on insurance status. The longitudinal sample is further constricted to include only children who had positive full year weights for 2003 and participated in MEPS for the entire year. We evaluate January enrollment outcomes for CHIP eligible children in the cross-sectional analysis. For our longitudinal analysis, we focus on January and December. In addition to premium, we control for family income, health and age of the child which are obtained from MEPS. Health is verified by asking parents whether the child gets sick easily, with higher scores pointing to better child health. Child age is measured as of the end of 2003.