Due to the exclusion of proxy respondents for the Patient Activation Supplement, adjusted survey weights were created to

generalize estimates to the Medicare population.3 order SAR131675 With the survey weights, this sample represents 40,729,409 Medicare beneficiaries. Cut points for high, moderate, and low activation were assigned at +/– ½ standard deviation of the unweighted mean for each question set. The unweighted score was used to determine the cut points as the distribution did not differ from the weighted scores. Sensitivity analyses included altering the survey response thresholds and the activation cut points. For more details on how the scale was created, see Appendix B. Summary scores from the supplement have been used in other research to assign levels of patient activation (Butler et al., 2012), and our method of scale creation is similar to the method demonstrated by Hibbard and Cunningham (2008). Data were analyzed using SAS survey procedures, which take into account the complex survey design of the MCBS in reporting standard errors. Activation levels

were first described across sociodemographic characteristics. Next, exploratory data analysis for the model included univariate logistic regression for all variables under consideration for the model. Missing data on covariates was less than 1% for each variable and so an “all available” data analysis was utilized, resulting in 10,512 beneficiaries included in the model, representing 40.2 million beneficiaries with the use of the cross-sectional weights. The outcome of interest was low patient activation, defined by a patient

activation score under ½ standard deviation of the mean. Multicollinearity among predictor variables was assessed by fitting a multiple linear regression model and obtaining the variance inflation factor. Several models were fit and assessed using Akaike’s Information Criteria (AIC). Final covariates in the weighted model included Medicaid eligibility, marital status, education level, race, sex, age, self-reported health status, number of functional limitations measured by self-reported difficulty with activities of daily living (ADLs), and usual place of health Entinostat care. Influence diagnostics, including plots of Pearson residuals and leverage, were used to identify potential influential data points. Deletion of these observations resulted in no noticeable change to model coefficients; therefore, the observations were retained. Goodness of fit was assessed using the Hosmer-Lemeshow test and indicated good fit. Lastly, mean service utilization and costs were compared across activation levels. Cost and utilization data was only available for the fee-for-service (FFS) population of 7,370 survey participants who completed the supplement, representing 28,326,423 Medicare beneficiaries.