Estimating the causal effects of dynamic strategies using observational data: applications to cancer survivors

Abstract

Estimating the causal effects of strategies (i.e., treatments, interventions, policies, exposures, regimes) is a key goal of public health and clinical research. In the real world, most strategies are dynamic, that is, they depend on an individual’s evolving characteristics over time. The effects of dynamic strategies would ideally be estimated in randomized trials. However, randomized trials are not always feasible, ethical, or available for the population of interest at the time when a decision must be made. Even when randomized trials do exist, it is often impractical to evaluate all dynamic strategies of interest within those studies. The causal effects of many dynamic strategies must therefore be estimated from observational data.

In this dissertation, we describe and implement methods for estimating the causal effects of dynamic strategies using observational data. Specifically, we use target trial emulation and g-methods to estimate the effects of dynamic strategies for older cancer survivors. To illustrate the wide application of these approaches, we estimate the effects of 1) treatment, 2) surveillance, and 3) lifestyle strategies.

In Chapter 1, we estimate the effects of adjuvant bone-modifying agent treatment strategies on breast cancer mortality for older postmenopausal women. We outline the protocol of a pragmatic randomized trial (the target trial) and then emulate it using cancer registry and healthcare claims data. We use a 3-step cloning, censoring, and inverse probability weighting procedure to estimate the parameters of a dynamic marginal structural model. Our findings show that bone-modifying agents reduce the risk of breast cancer mortality compared with no agent. Our results are compatible with evidence from prior randomized trials and extend this evidence to (i) estimate effects more precisely among older women and high risk subgroups and (ii) provide head-to-head comparisons of different approved agents.

In Chapter 2, we estimate the effects of post-diagnostic cystoscopy surveillance strategies on bladder cancer mortality among older adults with non-muscle invasive bladder cancer. We specify the protocols of 3 target trials, emulate them using registry and claims data, and then use the same 3-step cloning, censoring, and weighting procedure to estimate effects. We find that (i) more frequent, guideline-based cystoscopy results in reductions in bladder cancer mortality among high-risk patients and (ii) cystoscopy every 6 months as compared with every 3 months may not meaningfully increase bladder cancer mortality among low- and intermediate-risk patients. We show that an analysis that deviates from the protocol of a realistic randomized trial produces opposite results.

In Chapter 3, we describe challenges that arise when estimating the effects of lifestyle strategies. We outline a methodological framework that addresses those challenges and use that framework to estimate the effects of 2 recommendation-based lifestyle strategies on mortality among survivors of breast and prostate cancer. We use the extended parametric g-formula and data from 3 large epidemiologic cohorts. Compared with no intervention, we estimate meaningful reductions in mortality under an intervention requiring sustained adherence to 7 physical activity and dietary recommendations and meaningful increases in mortality under an intervention requiring no alcohol consumption. We show that the magnitude of our estimates vary under different assumptions.