A systematic approach for analyzing electronically monitored adherence data

We propose a 4-step process for analyzing medication adherence data generated by MEMS and similar electronic monitoring devices. SAS macros developed to support this analysis process are available on the Internet. An overview of these methods and macros is provided. Example analyses are presented to demonstrate these methods using MEMS data on HIV-positive subjects' adherence to antiretroviral medications. The four analysis steps are formulated including new extensions for adaptive modeling of the dispersion as well as of the expected value, i.e., variability in adherence as well as its mean. How to use the macros to conduct the example analyses is also described. The steps of the analysis process are: 1. Group MEMS opening events for each subject into opening counts and rates over disjoint intervals within that subject’s MEMS usage period. 2. Model grouped counts/rates for each subject using adaptive Poisson regression methods, fitting non-linear curves in time to the expected value and dispersion. 3. Cluster estimates of the expected value and dispersion at proportional times (e.g., every 5%) within subjects’ MEMS usage periods into adherence pattern types (e.g., high, moderate, low, improving, deteriorating). 4. Model membership in adherence pattern types in terms of available predictors. In Step 1, MEMS opening events are grouped using the grpevnts macro into opening counts and rates. These counts/rates are naturally analyzed using Poisson regression, but can change over time in a wide variety of complex patterns, and so non-linear models are required. In Step 2, count/rate data for each subject are adaptively modeled using the genreg macro. A heuristic search process is used identifying a non-linear model based on fractional polynomials (i.e., powers can be fractions) in time. Alternate models are compared using extended quasi-likelihood cross-validation (QLCV) scores, with larger QLCV scores indicating models more compatible with the data. Model selection is iterated over all subjects using the multsubj macro, which also generates estimates of the expected value and dispersion at proportional times, representing subjects' adherence patterns. In Step 3, these adherence patterns are clustered into adherence pattern types consisting of subjects with similar patterns using the LVCcluster macro. A wide variety of clustering alternatives corresponding to different clustering procedures under varying numbers of clusters are compared on the basis of likelihood cross-validation (LCV) scores. In Step 4, the properties of the resulting adherence pattern types are assessed by modeling