Statistical Analysis

We used DID analysis to estimate the average pre-post change in quality of care, health outcomes, and costs of care for SafeMed participants compared to controls. Negative binomial regression models were used for preventable hospitalizations, preventable ED visits, and all health care utilization outcomes, while expenditures were analyzed using generalized linear regression models. Additionally, we employed logistic regression and linear probability models to examine whether enrollment in SafeMed was associated with improvements in medication adherence and PCP visits, respectively. We ran all models for the overall population and by subgroup (Medicare, Medicaid, dual). Multivariate analyses controlled for all covariates except for individual chronic conditions since the Charlson Comorbidity Index were used to assess comorbidity. A significance P level of < 0.05 and 2-sided tests were used for all analyses.

In sensitivity analyses, we reran all models among SafeMed participants and propensity score–matched controls. The variables used to match controls included all covariates measured at baseline. All data analyses were conducted using SAS 9.4 (SAS Institute, Inc., Cary, NC) and Stata 13 (StataCorp 2013, College Station, TX).