Tuesday, April 14 2020
2:00pm - 4:00pm
Masters Presentation
Variable Selection in Causal Inference via Outcome-Adaptive Lasso


More estimators of the average causal effect of an intervention require estimation of the propensity score, the outcome regression, or both. Traditionally, a 'throw in the kitchen sink' approach has used to select the covariates for inclusion into the propensity score, but recent work have shown that including the variables associated with exposure but not outcome can cause bias and inflate the standard errors. However, inclusion of the variable that related to the outcome but not exposure can improve the statistical efficiency and remain unbiased. We built on proposed adaptive lasso in the prediction and offer a modified method 'outcome-adaptive lasso' for selecting the appropriate covariates to include in the propensity score to provide the unbiased causal effect and maintain statistical efficiency. The proposed method performs variable selection and unbiased estimation in large dimension settings. We report theoretical foundation and simulation results, presenting that the outcome adaptive lasso method selects the true confounders and the variable associated with outcome but not treatment into the propensity score model, while excluding the variables unrelated with outcome and exposure and only with exposure.
Speaker:Dingxuan Zhang
Location:See Zoom link from email

Download as iCalendar