Causal Inference Methods for Secondary Analysis of Randomized Screening Trials

The primary objective of randomized trials is usually pre-specified in the protocol and typically adheres to the intention-to-treat (ITT) principle, allowing for simple comparisons between intervention arms. However, trials often collect high-quality data that can be utilized for secondary analysis. This thesis is focused on randomized screening trials where asymptomatic individuals are assigned to receive a series of screening examinations or standard care and subsequently followed for a pre-specified period. While the primary analysis in randomized screening trials estimates the effect of intention-to-screen (ITS) on cancer-specific mortality, among the screening-detectable subgroup we might also be interested in the causal effect of early (screening-induced) treatments compared to delayed treatments in the absence of screening. The first objective of this thesis is to develop estimators for the effect of early versus delayed cancer treatments among the screening-detectable subgroup. Using the framework of Rubin’s causal model, we consider two alternative measures, proportional and absolute mortality reductions in the subgroup. We propose estimators for these using the instrumental variable principle as well as outline their identifying assumptions. These estimators generalize existing IV estimators to allow for time-dependent exposure/latent subgroup. The existing models for screening trials, primarily proposed for planning future trials with adequate power for the ITS analysis, are unnecessarily complex for defining and estimating the causal effect of screening-induced early treatments. To address this, we formulate a simplified structural multi-state model, in which the causal effect of early treatments is summarized using a time-invariant, cause-specific structural hazard ratio. For estimating the hazard ratio, we propose two methods, based on an estimating equation and a likelihood expression. Finally, with the aim to generalize the IV methods outside of the trial setting and to allow for covariate-dependent censoring, we introduce covariate adjustment into the estimation. We consider both parametric and non-parametric covariate adjustment using hazard regression models and machine learning algorithms. For the latter, we propose a sub-sampling approach to avoid large-counting process datasets. The performance of all the proposed estimators in this thesis are illustrated through simulation studies and real data examples.Ph.D