- PEISONG HAN, University of Waterloo
Multiply Robust Estimation in Causal Inference [PDF]
- Multiple robustness is a desirable property first established in the missing data literature. Estimators are multiply robust if they are consistent when any one of the multiple missingness probability models and multiple data distribution models is correctly specified. We will show how to construct multiply robust estimators for the average treatment effect for observational studies with binary treatments. The estimators are consistent if any one of the multiple propensity score models and multiple outcome regression models is correctly specified. Our estimation procedure can also achieve desired level of covariate balance between treatment and control groups by matching moments of the covariate distributions through reweighting the units in one or both groups.
- WEI LUO, Baruch College
On Estimating Regression-based Causal Effects Using Sufficient Dimension Reduction [PDF]
- In many causal inference problems, the parameter of interest is often the regression causal effect, defined as the conditional mean difference in the potential outcomes given covariates. This paper discusses how sufficient dimension reduction can be used to assist in its estimation, and proposes a new estimator for the regression causal effect inspired by a sufficient dimension reduction method called the minimum average variance estimation. The estimator requires a weaker common support condition than the traditional propensity score-based approaches. In addition, it can be easily converted to estimate the average causal effect, where it is shown to be asymptotically super efficient. The finite-sample properties of the proposed method are illustrated using simulation studies.
- MIREILLE SCHNITZER, University of Montreal
Collaborative Double Robustness and the Connection to Data-Adaptive Nuisance Model Selection in Causal Inference [PDF]
- In causal inference and censored data methods, double robust estimators such as Targeted Minimum Loss-based estimation require the specification of two model components. Estimation will be consistent under the correct specification of either nuisance component, conditional on a sufficient set of confounding variables. However, this class of estimators also has a collaborative robustness property that allows for consistent effect estimation in a wider class of settings. We describe this class of estimators in single time point and longitudinal exposure settings, describe a cohesive approach to variable selection in causal inference, and illustrate the performance of data-adaptive learning algorithms for the nuisance models.