Propensity Score Analyses

To address potential confounding variables in the preintervention period that affected outcomes in the postintervention period, our models used inverse probability of treatment weights (IPTW) in the form of average treatment effect in the treated (25). This approach allowed for statistical assessment of whether comparison states were equivalent to SIM states for the confounding covariates after balancing. Propensity scores were generated by way of a multigroup weighting strategy, which used multinomial logistic regression to estimate the probability of each observation being in the treated group in the pretreatment period (26).

The following county-level variables were used to estimate the scores for the event study analyses, because they are related to diabetes prevalence and hospitalization rates: sex, age, race/ethnicity, uninsured rate, and poverty rate (income <100% of the federal poverty level). Other variables were also considered, but a parsimonious set of variables was required in order to achieve sufficient balance for the sample of 889 counties. For the CITS analyses, the following patient-level variables were used to estimate propensity scores: age, comorbidity count, and indicators for being female, Black, Hispanic, other racial/ethnic minority, having Medicaid, being a dual Medicare/Medicaid enrollee, having private insurance, being uninsured/self-paying, and being admitted through the emergency department.

The weight of each observation was then calculated to be proportional to its probability of being in the SIM group before SIM relative to its probability of being in the SIM group and the time period in which it truly occurred (i.e., SIM group during SIM, non-SIM group pre-SIM, or non-SIM group during SIM). Our goal was to have the preintervention period absolute standardized differences of the variable means between the SIM and comparison groups to be <10% (27,28). The IPTWs were used as probability weights to estimate the regressions, thus making them doubly robust (29). This approach was previously used to study the effect of Medicare hospice enrollment on costs and quality (30). The DiD regression models were weighted for a county population by multiplying the county population by the IPTW weights. SEs were estimated by clustering at the state level to allow for correlation within states across time.