Secondary Logo

Institutional members access full text with Ovid®

Share this article on:

Can We Train Machine Learning Methods to Outperform the High-dimensional Propensity Score Algorithm?

Karim, Mohammad, Ehsanula,b; Pang, Menglanc,d; Platt, Robert, W.c,e,f

doi: 10.1097/EDE.0000000000000787

The use of retrospective health care claims datasets is frequently criticized for the lack of complete information on potential confounders. Utilizing patient’s health status–related information from claims datasets as surrogates or proxies for mismeasured and unobserved confounders, the high-dimensional propensity score algorithm enables us to reduce bias. Using a previously published cohort study of postmyocardial infarction statin use (1998–2012), we compare the performance of the algorithm with a number of popular machine learning approaches for confounder selection in high-dimensional covariate spaces: random forest, least absolute shrinkage and selection operator, and elastic net. Our results suggest that, when the data analysis is done with epidemiologic principles in mind, machine learning methods perform as well as the high-dimensional propensity score algorithm. Using a plasmode framework that mimicked the empirical data, we also showed that a hybrid of machine learning and high-dimensional propensity score algorithms generally perform slightly better than both in terms of mean squared error, when a bias-based analysis is used.

From the aSchool of Population and Public Health, University of British Columbia, Vancouver, British Columbia, Canada; bCenter for Health Evaluation and Outcome Sciences (CHÉOS), Providence Health Care, Vancouver, British Columbia, Canada; cDepartment of Epidemiology, Biostatistics and Occupational Health, McGill University, Montreal, Quebec, Canada; dCentre For Clinical Epidemiology, Lady Davis Research Institute, Jewish General Hospital, Montreal, Quebec, Canada; eDepartment of Pediatrics, McGill University, Montreal, Quebec, Canada; and fThe Research Institute of the McGill University Health Centre, Montreal, Quebec, Canada.

Submitted April 24, 2017; accepted November 14, 2017.

Supported by a post-doctoral fellowship from the Canadian Network for Observational Drug Effect Studies (CNODES). CNODES, a collaborating centre of the Drug Safety and Effectiveness Network (DSEN), is funded by the Canadian Institutes of Health Research (CIHR). M. E. K. is a Scientist and Biostatistician at the Centre for Health Evaluation and Outcome Sciences (CHÉOS), faculty of Medicine, UBC. M. P. holds a studentship from the Fonds de Recherche du Québec - Santé (FQR-S). R. W. P. holds the Albert Boehringer I Chair in Pharmacoepidemiology and is a member of the Research Institute of the McGill University Health Centre, which is supported by core funds from FQR-S.

M. E. K. has received accommodation costs from the endMS Research and Training Network (2011, 2012), Statistical Society of Canada (2016) to present at conferences, and from Pacific Institute for the Mathematical Sciences (2013), the Canadian Statistical Sciences Institute (2016) to attend workshops. R. W. P. has received fees for service for consulting from Abbvie, Amgen, Eli Lilly, and Searchlight Pharma, for teaching from Novartis, and for scientific steering committee membership from Pfizer.

Availability of Data and Code for Replication: Software code hints are provided in the supporting material (as an eAppendix) for implementing the methods. Retrospective population-based cohort Dataset from the Clinical Practice Research Datalink is not publicly available due to patient confidentiality reasons.

Supplemental digital content is available through direct URL citations in the HTML and PDF versions of this article (

Correspondence: Mohammad Ehsanul Karim, School of Population and Public Health, University of British Columbia, 2206 East Mall, Vancouver, BC V6T 1Z3. E-mail:

Copyright © 2018 Wolters Kluwer Health, Inc. All rights reserved.