Abstract
Randomized controlled trials are the gold standard approach for estimating causal effects of treatments on health outcomes, but are typically restricted to relatively short follow-up time and a subset of the eventual treatment population. Observational data on treatment use offers the possibility of estimating treatment effects over long periods of follow-up and in diverse populations. However, to estimate treatment effects from observational data we must tackle the challenge of confounding, especially by time-dependent covariates. This paper talk will explore methods for estimating the effects of treatment on survival using longitudinal observational data in which information on treatment use and covariates is obtained at approximately regular visits. We will discuss what can be done with standard analyses, followed by marginal structural models, the sequential trials method, g-computation and others. A particular focus will be on investigating what causal quantities can be estimated using the different methods. I am motivated by the aim of estimating the effect of a treatment called DNase on the survival of people with cystic fibrosis. I will describe how the methods were applied to answer this question using data from the UK Cystic Fibrosis Registry, a longitudinal database of over 10,000 people.