Estimating the causal effect of dynamic treatment strategies on pregnancy using electronic medical records or adaptive clinical trials
Embargoed until:
2022-02-14Permanent Link
https://hdl.handle.net/2144/39542Abstract
Randomized Controlled Trials (RCTs) are the gold standard of clinical research. Traditional RCT designs, however, cannot be modified after a trial has begun: sample size and randomization to treatment arms are all fixed at the outset. Given the length and cost of these trials, treatments based on novel compounds can take 10-15 years to reach patients. Flexible RCTs, or adaptive clinical trials, have been developed to reduce the time to market for new treatments and potentially reduce cost. In this dissertation, we provide an overview of adaptive RCT methods and evaluate the probability of pregnancy under dynamic infertility treatment strategies by emulating an RCT using Electronic Medical Records (EMR).
First, we review existing adaptive RCTs and compare growth in publications of novel adaptive design methodology with adaptive trials conducted by industry. However, conducting an RCT is not always ethical, feasible or timely. In particular, treatments of chronic conditions often involve a sequence of treatments that can be modified or changed based on a patient's response to previous treatments or changes in patient characteristics. Evaluating such dynamic treatment strategies requires RCTs that re-randomize each individual's treatment in a setting where resources and follow-up time are limited. Therefore, observational studies are a common alternative.
Second, we evaluate dynamic strategies for infertility treatment. Setting a woman's expectation for a successful pregnancy is challenging for clinicians as the pregnancy rate in each infertility treatment cycle varies across age, infertility diagnosis, type of treatment and history of previous treatments. We show that any successful attempt to estimate the probability of pregnancy had all women followed a particular treatment strategy should take into account that women who choose this particular strategy and continue until they are pregnant may have a different probability of pregnancy under the strategy than women who don't make this same choice: there is time-dependent confounding by indication, and some women decide to stop treatment before becoming pregnant. We show in detail how to use Marginal Structural Models to estimate the probability of pregnancy under a variety of dynamic treatment strategies, using EMRs from a network of infertility clinics across the United States.
Rights
Attribution-NonCommercial-NoDerivatives 4.0 InternationalCollections