VARIABLE SELECTION FOR PROPENSITY SCORE MODELS IN COMPARATIVE EFFECTIVENESS RESEARCH ON CHRONIC CORONARY ARTERY DISEASE

Purpose:   To identify and compare variable selection approaches for propensity score models in studies comparing treatments for chronic coronary artery disease (CAD).

Methods:   We searched PubMed Jan 2000-Feb 2014 for studies comparing coronary artery bypass graft (CABG), percutaneous coronary intervention (PCI), or medical therapy in adult patients with CAD. We included studies that used propensity scores to compare at least two of the above treatments or two alternative types of stents (e.g., bare-metal stents, BMS, vs. drug-eluting stents, DES). We extracted information about study conduct and design, patient characteristics, treatments, and the propensity score approaches used.

Results:   108 studies met our criteria. Given the large number of CABG vs. PCI studies (n=43) and BMS vs. DES studies (n=57), for the purpose of this assessment we randomly selected 20 studies among each of these comparisons. Hence, we included a total of 47 studies: 20 CABG vs. PCI, 20 BMS vs. DES, 5 CABG vs. medical therapy, 1 PCI vs. medical therapy, and1 DES vs. balloon angiography.

   The reported method for variable selection was typically unclear (n=28), though some reported including all “clinically relevant” covariates (n=13), and others reported using statistical criteria to automatically select variables for inclusion (e.g., backward or forward stepwise selection). Few studies reported incorporating interactions of covariates in the propensity score model.

   Over 400 different variables were used to construct propensity score models. We organized these variables into 12 categories: (1) demographics, (2) CAD descriptors (e.g., number of diseased vessels, location), (3) CAD severity, (4) clinical events (e.g., acute coronary syndrome, cardiogenic shock), (5) comorbidities, (6) cardiovascular disease risk factors, (7) ventricular function, (8) hospital characteristics, (9) laboratory values, (10) medication use, (11) procedure characteristics, and (12) post-procedure characteristics. We further organized variables into finer subcategories. Few subcategories/variables were consistently selected into propensity score models across studies. See Figure for variables by study.

Conclusions:   Based on 47 studies from the current literature, propensity models for chronic CAD included over 400 different variables. Even after categorizing them, variables were not consistently used in models, which may be partially attributed to different variable selection approaches. Future empirical work should compare the identified approaches/variables in a new observational dataset and assess the impact of different propensity scores on study conclusions regarding comparative effectiveness.