Primary Submission Category: Machine Learning and Causal Inference
Transfer Learning With Efficient Estimators To Optimally Leverage Historical Data in Analysis of Randomized Trials
Authors: Lauren Liao, Alan Hubbard, Alejandro Schuler,
Presenting Author: Lauren Liao*
Inference from randomized control trials (RCTs) yields reliable causal interpretation but is often limited by the sample size. RCTs often answer similar questions that were done in observational studies. Opposed to RCTs, previous observational studies are often abundant in data although the causal link is less well-defined. While a typical RCT is analyzed alone, estimators have been developed to incorporate historical, observational studies to increase the estimator’s efficiency. Previous estimators developed to incorporate historical data in RCT analyses often have unrealistic additional assumptions that can create a biased estimate. Instead, we propose using no additional assumptions on the historical data, nor the connection between the historical and trial data, to increase efficiency in analysis of RCTs. This is performed by leveraging machine learning to estimate generalized prognostic scores from the observational studies of similar nature. These generalized prognostic scores are predictions on the current study using historical models, and they are included in the current analysis as covariates in an efficient estimator. We demonstrate the utility of this estimator leveraging historical data on a randomized blood transfusion study of trauma patients.