Effect Modification of Interventions. Bridging the Gap Between Clinical Studies and the Individual Patient

Abstract

Clinical studies are usually designed to provide information on the average intervention effect. Therefore, possible differences in treatment effects among a wide range of patients will often remain undetected. When treatment effects differ between patients, this is referred to as effect modification, interaction, or heterogeneity of treatment effects. Detecting treatment effect modification is an essential step in translating results of a clinical study to individual patients. Studies described in the thesis “Effect Modification of Interventions: Bridging the Gap Between Clinical Studies and the Individual Patient” showed that the likelihood of detecting interaction effects is often low unless such effects are unrealistically large. A potential solution is to combine study results from different studies. However, a review of published studies showed that interaction effects may differ markedly between studies, notably between randomized and nonrandomized studies, which indicates the importance of taking study quality into account. A potential intuitive way to incorporate this is using Bayesian statistics to downweight biased studies. Initial simulations indeed showed an increase in power at the cost of a small increase in bias using these methods. Most assessments of interaction are based on a single variable (e.g., gender). However, patients likely differ on more than one variable. Hence, interaction tests incorporating multiple variables seem desirable. A relatively new approach to multivariable interaction testing was applied on canines suffering from osteosarcoma. Results suggested that dogs with a relatively lower baseline risk for mortality benefitted the most from chemotherapy. A literature review and empirical study of cardiovascular interventions focused on detecting the presence or absence of treatment effect modification. These studies suggested that treatment effects did not differ between studies, which implies that similar treatment effects might be expected in dissimilar patients (i.e., effects were generalizable).