Wednesday, Aug 6: 8:30 AM - 10:20 AM
0423
Invited Paper Session
Music City Center
Room: CC-101C
Causal inference
Applied
No
Main Sponsor
Section on Statistics in Epidemiology
Co Sponsors
Biometrics Section
Health Policy Statistics Section
Presentations
The general theory of higher-order influence functions (HOIF) has been successfully applied to several pathwise differentiable parameters arising in causal inference, such as the expected conditional covariance and the treatment-specific mean. Such theory has yielded minimax optimal estimators in certain nonparametric models, e.g., those indexed by smooth nuisance parameters. More recently, minimax optimal, higher-order estimators have been derived for some non-pathwise differentiable causal parameters, an example of which is the conditional average treatment effect. In this work, we aim to extend the application of HOIF theory to causal parameters defined by a time-varying treatment. As a leading example, we consider the two-time point case g-formula functional in an anisotropic smoothness model where the nuisance functions can depend more smoothly on certain covariates. We also consider even more structured models, such as additive ones. In each setting, we design a higher-order estimator and calculate its bias and variance, and for some of them, we show that the convergence rates established are minimax optimal.
Keywords
TBD
In the framework of dynamic marginal structural models, {\it regimen-response curve} is a function that describes the relation between the mean outcome and the parameters in the class of decision rules. Estimation of the regimen-response curve plays a crucial role in constructing the optimal regime, as misspecification of the working model may lead to a biased estimate with questionable causal interpretability. However, the existing literature lacks methods to evaluate and compare different working models.}} To address this problem, we will leverage risk to assess the ``goodness-of-fit" of an imposed working model. We consider the counterfactual risk as our target parameter and derive inverse probability weighting and canonical gradients to map it to the observed data. We provide asymptotic properties of the resulting risk estimators, considering both fixed and data-dependent target parameters. We will show that the inverse probability weighting estimator can be efficient and asymptotic linear when the weight functions are estimated using a sieve-based estimator.
Keywords
TBD
We consider time to treatment initialization. This can commonly occur in preventive medicine, such as disease screening and vaccination; it can also occur with non-fatal health conditions such as HIV infection without the onset of AIDS; or in tech industry where items wait to be reviewed manually as abusive or not, etc. While traditional causal inference focused on `when to treat' and its effects, including their possible dependence on subject characteristics, we consider the incremental causal effect when the intensity of time to treatment initialization is intervened upon. We provide identification of the incremental causal effect without the commonly required positivity assumption, as well as an estimation framework including the efficient influence function. We illustrate our approach via simulation, and apply it to a real world data set.
Keywords
positivity
stochastic intervention
inverse probability weighting
efficient influence function
Recent years have seen a growing interest in quantifying treatment effect heterogeneity, which is vital for supporting individualized decision-making. Though black-box machine learning approaches might optimally predict treatment effect heterogeneity, in high-risk domains such as medicine, decision makers often hesitate to rely on decision support systems without understanding the underlying rationale behind the recommendations. Hence, it is crucial to offer insights into which variables best predict individualized treatment effects. Motivated by these considerations, we present model-agnostic variable importance measures for heterogeneous treatment effects. We provide efficient estimators of these measures together with corresponding confidence intervals, and introduce a Wald-type test to assess the null hypothesis of no importance. Our approach builds on recent developments in semiparametric theory for pathwise differentiable function-valued parameters, and is valid even when flexible black-box algorithms are employed to quantify treatment effect heterogeneity. We demonstrate the applicability of our methodology in the context of infectious disease prevention strategies.
Keywords
Variable importance measures
Heterogeneous treatment effects
Researchers often express interest in treatment effects that adequately account for post-treatment events (so-called intercurrent events). However, outcome contrasts that naively condition on intercurrent events lack a straightforward causal interpretation, and the practical relevance of other commonly used approaches is debated. In this presentation, I will propose strategies for formulating and choosing an estimand, beyond the marginal intention-to-treat effect, from the perspective of a decision maker and treatment developer. I will emphasize that a well-articulated, practically useful research question should either reflect decision-making at this point in time or future treatment development. A common feature of estimands that are practically useful is their correspondence to possibly hypothetical but well-defined interventions in identifiable (sub)populations. To illustrate my points, I will consider examples that have recently motivated the consideration of principal stratum estimands in clinical trials. In all of these examples, I will suggest alternative causal estimands that align with explicit research questions of practical interest and require less stringent identification assumptions.
Keywords
Causal inference
Estimands
Intercurrent events
Survival analysis
Biostatistics
Randomized experiments
Reproducible and generalizable assessment of a treatment decision requires principled handling of subsequent treatment decisions whose patterns may shift across cohorts and over time. Discontinuation and switching of treatment in clinical practice is a dynamic process that may be informative for expected outcomes. However, the information about expected outcomes has not been systematically documented or indicated in real-world health care data.
To effectively account for informative treatment switching in real-world evidence, we propose an instrumental variable approach that deals with the poorly documented expected outcomes as unmeasured confounding. Our proposed method is doubly robust, i. e. providing consistent treatment effect estimation whenever either of baseline propensity models and no-switching outcome models is consistently estimated. A co-training of treatment effect parameter and survival outcome regression model eliminates the requirement of a no-switching subset. We further develop an baseline-survival-corrected cross-fitting approach to incorporate general machine learning models for estimating nuisance models. Numerical results demonstrate the validity of proposed method in a wide range of data generating process while a basket of benchmark solutions producing biased or contradictory results. We apply our method to comparison of high-efficacy vs standard efficacy disease modifying treatment as the second line therapy of multiple sclerosis.
Keywords
Survival analysis
causal inference
Instrumental variables