Choosing Covariate Balancing Methods for Causal Inference: Practical Insights From a Simulation Study

Weighting methods are widely used for confounding adjustment in observational studies, but their finite-sample behavior depends on implementation choices and empirical overlap. We compare IPTW, just- and over-identified covariate balancing propensity score (CBPS), CBPS by tailored-loss function (CBPS-TLF), energy balancing (EB), and kernel optimal matching (KOM).

We conducted Monte Carlo simulations across 36 main scenarios varying sample size, treatment prevalence, and a complexity factor increasing confounding and reducing overlap. The main simulation considered a null constant treatment effect, with non-null constant effects examined as sensitivity analyses. Average treatment effects and average treatment effects on the treated were estimated using weighted least squares (WLS) and doubly robust (DR) estimators. Inference followed published recommendations when feasible. An empirical illustration used PROBITsim.

Performance depended on the estimator and scenario complexity. Under WLS, IPTW and CBPS-TLF were more sensitive to complexity, while standard CBPS often behaved similarly to IPTW but with less deterioration in some high-prevalence settings. EB and KOM showed more stable point-estimation patterns across scenarios. DR estimation reduced differences between weighting methods when all confounders were included in the outcome model, although confidence-interval performance remained heterogeneous. PROBITsim results were consistent with simulation patterns.

The study should be read as practical guidance rather than a ranking of methods. It identifies settings where weighting analyses become sensitive to prevalence, overlap, tuning, and variance estimation. Confidence intervals that account for weight construction and tuning remain an important open practical issue.