Causal inference symposium

Title: On the optimality guarantees for dynamic treatment regimes
Speaker: Mats Stensrud

Abstract:
Health care providers seek to implement decision rules that, when applied to individuals in the population of interest, give the best possible outcomes. This motivation underlies the current interest in precision medicine: the search for individualized treatment decisions tailored to patients’ characteristics.
 
I will consider how to formulate, select and estimate effects that can guide individualized treatment decisions. In particular, I will introduce a class of regimes that can outperform conventional optimal regimes. I will then discuss the challenges involved in estimating optimal regimes, both conventional ones and this new class. One key difficulty is that the more covariates we include in our model, and the more finely we aim to personalize decisions, the stronger the assumptions we must impose.
 
As a solution, I will propose a different strategy for detecting and estimating customized group effects, which can be viewed as coarsenings of conventional optimal regimes. I will give explicit frequentist guarantees that these groups differ in their effects. Finally, I will show that, in realistic settings, group-based strategies can substantially outperform modern optimal regime methods, even when those methods are implemented correctly using modern (doubly robust) machine learning techniques.

--------------------------------------

Title: Estimation of Causible Parameters in Register Data
Speaker: Thomas Gerds

Abstract:
A central aim of pharmacoepidemiology is to evaluate the effectiveness and safety of medications using electronic health records. For example, studies of glucose-lowering medications in patients with type 2 diabetes can be designed to emulate a target trial comparing different drug classes, with cardiovascular disease events as the primary outcome. We define causible parameters that possess a desired clinical interpretation in a hypothetical target trial. The corresponding causal inference framework then allows us to state and discuss the assumptions for causal interpretation. Some of these assumptions are not testable and some will be violated to some extent, so that an estimate of a causible parameter based on register data does not necessarily allow a causal interpretation. Hazard ratios are not causible. In this  presentation, I will further discuss the practical challenges of applying the longitudinal targeted minimum-loss based estimator (LTMLE) to Danish register data. In particular, I will address how the data that are inherently collected in continuous time are mapped to a discrete time scale, preserving essential information, accommodating polypharmacy and competing risks, and avoiding causal confusion.

--------------------------------------
 
Title: Target trial emulation: making causal inference less casual
Speaker: Miguel Hernan

Abstract:
When randomized trials are not available, causal effects are often estimated from observational healthcare databases. Therefore, causal inference from observational data can be viewed as an attempt to emulate a hypothetical randomized trial—the target trial—that would quantify the causal effect of interest. Contrary to what is generally believed, many well-known failures of observational studies were not the result of lack of randomization but of not adequately emulating the design of a target trial. This talk discusses how the target trial framework helps articulate precise causal questions, increases transparency in the procedures used to generate the answers, and prevents design biases when using observational data.