Comparing g-computation, propensity score-based weighting, and targeted maximum likelihood estimation for analyzing...

To have confidence in one's interpretation of treatment effects assessed by comparing trial results to external controls, minimizing bias is a critical step. We sought to investigate different methods for causal inference in simulated data sets with measured and unmeasured confounders.

The simulated data included three types of outcomes (continuous, binary, and time-to-event), treatment assignment, two measured baseline confounders, and one unmeasured confounding factor. Three scenarios were set to create different intensities of confounding effect (e.g., small and blocked confounding paths, medium and blocked confounding paths, and one large unblocked confounding path for scenario 1 to 3, respectively) caused by the unmeasured confounder. The methods of g-computation (GC), inverse probability of treatment weighting (IPTW), overlap weighting (OW), standardized mortality/morbidity ratio (SMR), and targeted maximum likelihood estimation (TMLE) were used to estimate average treatment effects and reduce potential biases.

The results with the greatest extent of biases were from the raw model that ignored all the potential confounders. In scenario 2, the unmeasured factor indirectly influenced the treatment assignment through a measured controlling factor and led to medium confounding. The methods of GC, IPTW, OW, SMR, and TMLE removed most of bias observed in average treatment effects for all three types of outcomes from the raw model. Similar results were found in scenario 1, but the results tended to be biased in scenario 3. GC had the best performance followed by OW.

The aforesaid methods can be used for causal inference in externally controlled studies when there is no large, unblockable confounding path for an unmeasured confounder. GC and OW are the preferable approaches.