Use of machine learning for comparing disease risk scores and propensity scores under complex confounding and large sample size scenarios: a simulation study

ABSTRACT Background The surge of treatments for COVID-19 in the ongoing pandemic presents an exemplar scenario with low prevalence of a given treatment and high outcome risk. Motivated by that, we conducted a simulation study for treatment effect estimation in such scenarios. We compared the performance of two methods for addressing confounding during the process of estimating treatment effects, namely disease risk scores (DRS) and propensity scores (PS) using different machine learning algorithms. Methods Monte Carlo simulated data with 25 different scenarios of treatment prevalence, outcome risk, data complexity, and sample size were created. PS and DRS matching with 1: 1 ratio were applied with logistic regression with least absolute shrinkage and selection operator (LASSO) regularization, multilayer perceptron (MLP), and eXtreme Gradient Boosting (XgBoost). Estimation performance was evaluated using relative bias and corresponding confidence intervals. Results Bias in treatment effect estimation increased with decreasing treatment prevalence regardless of matching method. DRS resulted in lower bias compared to PS when treatment prevalence was less than 10%, under strong confounding and nonlinear nonadditive data setting. However, DRS did not outperform PS under linear data setting and small sample size, even when the treatment prevalence was less than 10%. PS had a comparable or lower bias to DRS when treatment prevalence was common or high (10% - 50%). All three machine learning methods had similar performance, with LASSO and XgBoost yielding the lowest bias in some scenarios. Decreasing sample size or adding nonlinearity and non-additivity in data improved the performance of both PS and DRS. Conclusions Under strong confounding with large sample size DRS reduced bias compared to PS in scenarios with low treatment prevalence (less than 10%), whilst PS was preferable for the study of treatments with prevalence greater than 10%, regardless of the outcome prevalence. Key Messages When handling nonlinear nonadditive data with strong confounding, DRS estimated by machine learning methods outperforms PS in scenarios with low treatment prevalence (less than 10%). However, if having linear data and small sample size data with strong confounding, we did not observe DRS outperformed PS even when treatment prevalence was less than 10%. Our results suggested that using PS performed better compared to DRS in tackling strong confounding problems with treatment prevalence greater than 10%. Small sample size increased bias for both DRS and PS methods, and it affected DRS more than PS.