Bias bounds and target trials for causal inference in observational epidemiology

Abstract

Observational epidemiology is critical for understanding population health but requires careful consideration of possible biases. Tools for avoiding and managing these biases are essential. This dissertation describes and implements methods for designing, analyzing, and assessing observational studies, with a particular focus on target-trial emulation and bounds for biases.

In Chapter 1, I investigate the association between COVID-19 and preterm birth using data from a large, international pregnancy registry. The principles of target-trial emulation guide the design of an analysis that avoids immortal time bias while allowing for the evaluation of gestational age-specific effects of the disease. I show that severe COVID-19 in the third trimester increases risk of preterm birth, but carries less risk earlier in pregnancy, and mild or moderate COVID-19 confers minimal added risk at any time during pregnancy. This conclusion is confirmed with additional, complementary analyses.

Chapter 2 concerns a more complex target trial that implements sustained treatment strategies. In the setting of recurrent prostate cancer, I design a trial to estimate the optimal approach for initiating hormonal treatment based on biomarker characteristics. I then describe and conduct its emulation using two complementary methods: the parametric g-formula and inverse probability-weighted dynamic marginal structural models. I find no evidence that any of the treatment strategies I consider improves upon the approach of initiating treatment only with evidence of overt metastasis.

Finally, in Chapter 3, I improve upon existing methods for sensitivity analysis that can be used to assess one type of bias at a time. Building on the E-value approach for unmeasured confounding, as well as similar bounds for selection bias and misclassification, I consider the effects of these three biases simultaneously. I show that a bound for the bias of the observed risk ratio can be constructed as a function of sensitivity analysis parameters describing each type of bias. I apply this method for sensitivity analysis to studies of exposures in pregnancy and demonstrate the software developed to implement it.