Once the final PACE, HCBS, and NH samples were chosen, we used a nearest neighbor matching algorithm based on propensity score estimation--implemented separately for each of the eight states. Under this approach, we first estimated a logistic regression with a binary outcome variable for treatment status (PACE = 1; HCBS/NH = 0) and with pre-enrollment demographics, chronic conditions, and service utilization and costs included as covariates in the model. Specific covariates that we included in the model are--age; gender; indicators for race; indicators for chronic conditions (Alzheimer's disease or dementia, coronary artery disease (CAD), congestive heart failure (CHF), depression, diabetes, and stroke); number of chronic conditions; indicators for any inpatient hospitalizations, any emergency room (ER) visits, any SNF use, and any home health service use; and annualized total Medicare expenditures--all measured over the calendar year prior to the year of enrollment in PACE or HCBS waiver services or NH entry, and obtained from the Master Beneficiary Summary File (MBSF, described below).
Next, we used the propensity scores or the predicted probabilities of being a treatment group (PACE) enrollee, as obtained from the logistic regression, to implement our matching algorithm. Using the nearest neighbor algorithm, we matched each PACE enrollee to the comparison group (HCBS enrollee or NH entrant) with the closest propensity score. To allow for the best possible matches, we implemented matching with replacement, that is, the same comparison group enrollee could be matched to more than one PACE enrollee (overwhelming majority of the matched comparison group members--nearly 88 percent--were matched to a single PACE enrollee). The final matching weight of a comparison group enrollee is, therefore, the number of treatment group enrollees to whom she has been matched, and the weighted comparison group sample size is equal to the number of PACE enrollees in the sample. We estimated the propensity score model and implemented the nearest neighbor algorithm separately for each of the eight states in our sample.
Next, we repeated the matching process with only HCBS waiver enrollees in the comparison group pool, that is, without the NH entrants, in order to draw a second matched comparison group consisting of HCBS waiver enrollees alone (using matching with replacement, over 76 percent of matched HCBS enrollees were matched to a single PACE enrollee). We use this second matched group of HCBS waiver enrollees to examine all outcomes--expenditures, mortality, and NH utilization--for PACE versus matched comparison group members. Also, this second matched comparison group forms the basis of our main findings for the NH utilization outcomes since NH utilization in the first comparison group would be severely skewed upwards by the inclusion of NH entrants. Matching results, including evidence of baseline equivalence, are presented for both matched comparison samples in Section III below.