B0800
Title: On logistic regression to estimate treatment effects with observational, right censored, competing risks data.
Authors: Paul Blanche - University of Copenhagen (Denmark) [presenting]
Thomas Scheike - University of Copenhagen (Denmark)
Abstract: In medical research, the t-year risks of two groups of patients receiving different treatments are often compared using large observational data. For example, Danish registry data were recently used to compare the 33-month risk of cardiovascular death between patients who have initiated a beta-blocker treatment and those who have not among patients alive three months after myocardial infarction (MI). An unadjusted analysis is expected to be confounded, and logistic regression can be used instead to adjust for observed confounders (e.g. age, procedure during MI hospital admission, hypertension and diabetes). Unlike hazard or subdistribution hazard regression commonly employed with competing risk data, logistic regression directly models the t-year risk. It, therefore, relies on weaker assumptions and facilitates discussions between the clinical experts and the data analyst regarding how to best adjust for confounders (e.g., the relevance of including interaction terms). It presents how a simple inverse probability of censoring weighting approach can deal with right censoring to fit the model and the corresponding asymptotic properties of the estimator. How marginal risks can easily be computed via standardization (aka G-computation) and double robust estimators are further discussed from the fitted logistic model, as is commonly done in binary uncensored data. The methodology is illustrated using the Danish registry data mentioned above.
