The study uses Medicare Current Beneficiary Survey (MCBS) data. Cases were selected based on self-reported diabetes or the presence of an ICD-9 code for diabetes and complications (250.xx), polyneuropathy in diabetes (357.2), diabetic retinopathy (362.01, 362.02), or diabetic cataract (366.41) on one hospital, skilled nursing facility, or home health claim or any of these codes on two outpatient or physician claims following a validated protocol (10
). These selection criteria resulted in a sample of 7,441 individuals with diabetes who contributed 14,317 annual observations for the analysis.
We used MCBS prescription medication files to identify users of the following seven drug classes: older oral antidiabetes drugs (metformin and sulfonylureas), newer oral agents (thiazolidinediones, meglitinides, and α-glucosidase inhibitors), insulins, ACE inhibitors, ARBs, statins, and other lipid-lowering medications (ezetimibe, fibrates, niacin, and others). The primary explanatory variable in our analysis is the annual number of prescription fills per class per year.
We assessed the effect of prescription fill rates for users of each drug class on the risk of hospitalization, total annual hospital days, and spending on Medicare services measured in constant 2006 dollars, using the Consumer Price Index (12
). Covariates included an extensive list of demographic, socioeconomic, and health status indicators (see Table A1 in the online appendix available at http://care.diabetesjournals.org/cgi/content/full/dc08-1311/DC1
We estimated seven regression models, one per drug class, for each of the three dependent variables using person-year as the unit of analysis and the full set of covariates listed in the online appendix. Because the study subjects frequently used medications in two or more drug classes, we included fill rates for all seven drug classes in each equation. This procedure assured that the parameter coefficient on prescription fills for the subset of users of a particular drug class was conditioned on utilization of the other medication classes.
We used logistic regression for the hospitalization models and Poisson regression for the hospital day equations. For the Medicare spending models, we used a generalized linear equation with a γ distribution and log link to approximate the skewed distribution of Medicare expenditures (13
). All models were estimated in Stata (Release 9) with a robust cluster command to correct standard errors for repeated measures among subjects observed in multiple years. Results are reported as conditional marginal probabilities (hospitalization) or conditional marginal effects (dy/dx) of a unit change in prescription fills on the change in the dependent variable (hospital days and Medicare spending), with all other variables held at their mean values.