PRM17 - Methods for evaluating the effect modification in the observational studies: A retrospective analysis on the impact of simvastatina and ezetimibe and statins on acute myocardial infarction

Abstract

Fundamental potential weaknesses of observational studies are:bias and effect modification. In this situation, computing an overall estimate of association is misleading. The aim was to compare a traditional multivariable-adjusted model with a propensity score (PS) model and a cluster analysis (CA) model, in estimating the association between type of lipid modifying agent and hospitalization for Acute Myocardial Infarction (AMI). The Health Register of Emilia-Romagna (Italy; more than 10 Million records; 4.4 Milion Inhabitants) was used to select between January 1st, 2006 and December 31st, 2011 Statins and Simvastatin and Ezetimibe (SE) naïve users. A PS was constructed, predicting treatment assignment from age, gender, use of diabetic agents, different pharmacologic agents, comorbidity level and utilization of outpatient services. For analysis’ purpose, the effect of the treatment on the risk of IMA was measured by estimates of hazard rations (HR) in different fashions using: multivariable Cox regression model (CRM), CRM adjusted for the PS, CRM model within each cluster identified by a K-means method. Over 2,6 Mil inhabitants (+40 years) 57,902 (92.2%) patients were naïve statin users and 4,904 (7.8%) were SE users. Compared with Statins, the risk of IMA for SE resulted similar in the adjusted CRM and in the propensity CRM (HR=1.47 and HR=1.49 respectively). While the CRMs performed within each cluster yielded different treatment effect estimates (HR=2.39 for Cluster 1; HR=1.36 for Cluster 2; 1.37 for Cluster 3). The CA allowed to identify specific subgroups of patients, with homogeneous risk features. The CRM within each cluster yielded different treatment effect estimates that might suggest the presence of unmeasured confounding. In that case, traditional regression model and PS developed using administrative data do not necessarily balance patient characteristics contained in clinical data. Choice among different approaches for investigating effect modification should be sensitive to the circumstances of the data analysis in applying observational studies.