Doubly robust estimators are often viewed as protected against model misspecification, but this protection can fail in the presence of missing data. When multiple imputation is used, lack of congeniality between the imputation and analysis models can induce bias even when both the propensity score and outcome models are correctly specified. Valid inference requires imputation models to include all variables from both the propensity score and outcome models, specified in compatible functional forms. Violations of these conditions can lead to substantial bias in treatment effect estimates. A general framework for combining multiple imputation with doubly robust estimation is presented, the conditions required for congeniality are characterized, and the consequences of misspecification are illustrated. The talk concludes with practical recommendations for specifying imputation models to preserve the validity of doubly robust methods in applied causal analyses.

This workshop covers types of missing data seen in causal analyses and describes when dealing with them is necessary from both a bias and efficacy perspective. We then focus specfically on multiple imputation when estimating causal effects with a particular emphasis on which variables must be present in the imputation model (and in what functional form).

Powering a trial to detect subgroup-by-treatment interactions is, in most realistic scenarios, simply not feasible. At the design stage, investigators presume clinical equipoise, that is they do not expect treatment to harm any group of patients. The overall effect is the weighted average of subgroup-specific effects and the expectation of clinical equipoise sharply constrains how different those effects can be. Under these assumptions, the sample sizes required to detect interactions are dramatically larger than those needed to detect the overall treatment effect. In a trial designed for 80% power on the overall effect, detecting an interaction of comparable magnitude would require at least four times the sample size, and even under ideal conditions, interaction test power hovers around 29%. However, power to detect a subgroup-specific effect is not the same as power to detect an interaction, and these reflect fundamentally different scientific goals. Typically, a patient does not care whether their treatment effect is statistically distinguishable from another subgroup’s, they care whether there is a clinically and statistically meaningful benefit within their own subgroup. Meaningful, valid estimates of subgroup-specific effects are achievable without relying on formal interaction tests, and pre-specifying subgroups and reporting within-subgroup estimates directly is both statistically sound and more aligned with how patients and clinicians actually use trial results.

Linear models. Learn the theory behind linear models. Gain hands on experience with real data from a variety of disciplines. The course will focus on the statistical computing language R.

Statistical learning. Learn the theory behind cutting edge statistical and machine learning techniques. Gain hands on experience with real data from a variety of disciplines. The course will focus on the statistical computing language R.

Linear models. Learn the theory behind linear models. Gain hands on experience with real data from a variety of disciplines. The course will focus on the statistical computing language R.

Traditional measures of vaccine efficacy (VE) are inherently asymmetric, constrained above by 1 but unbounded below. As a result, VE estimates and corresponding confidence intervals (CIs) can extend far below zero, making interpretation difficult and potentially obscuring whether the apparent effect reflects true harm or simply statistical uncertainty. The proposed symmetric vaccine efficacy (SVE) is a bounded and interpretable alternative to VE that maintains desirable statistical properties while resolving these asymmetries.

Objectives Electronic health record (EHR) data are prone to missingness and errors. Previously, we devised an enriched chart review protocol where a “roadmap” of auxiliary diagnoses was used to recover missing values. Still, chart reviews are expensive and time-intensive, limiting the number of patients whose data can be reviewed. Now, we investigate the accuracy and scalability of a roadmap-driven algorithm, based on International Classification of Diseases, 10th revision (ICD-10) codes, to mimic expert chart reviews and recover missing values.

Partitioned Local Depth (PaLD) is a framework for holistic consideration of community structure for distance-based data. This paper describes an R package, pald, for calculating Partitioned Local Depth (PaLD) probabilities, implementing community analyses, determining community clusters, and creating data visualizations to display community structure. We present essentials of the PaLD approach, describe how to use the pald package, walk through several examples, and discuss the method in relation to commonly used techniques.