Causal Inference in Presence of Intra-Patient Correlation due to Repeated Measurements of Exposure and Outcome in Longitudinal Settings.

Abstract

Introduction: In causal inference with time-dependent confounding between an exposure and an outcome, the repeated nature of measures is likely to lead to intra-patient correlation and introduces bias in the estimation of the causal effect, even in the absence of unmeasured confounders.

Method: We evaluated the impact of intra-patient correlation on causal effect estimation with g-computation, inverse probability weighting (IPW) and longitudinal targeted maximum likelihood estimator (LTMLE), and compared two ways of accounting for it, using a fixed-effects or a mixed-effects approach. We conducted a simulation analysis under different scenarios for numerous time points and a real-life analysis to investigate the causal effect of pregnancy on neurological disability in multiple sclerosis.

Results: In simulation analyses, the presence of intra-patient correlation led to bias in the causal effect estimation with g-computation, IPW, and LTMLE when this was not accounted for; the bias was smaller for LTMLE. Taking into account intra-patient correlation with fixed-effects and mixed-effects approaches reduced the bias in g-computation, with lower standard errors for the mixed-effects approach. Regarding IPW, the fixed-effects approach suffered from weight stability issues when the number of time points increased, and the mixed-effects approach provided inconsistent estimates of the intra-patient correlation in the exposure model. Application to real-life data yielded results consistent with the simulation study, highlighting the importance of accounting for intra-patient correlation.

Conclusion: When analyzing longitudinal data in the presence of time-dependent confounding using g-methods, intra-patient correlation due to repeated measurements of exposure and outcome should be accounted for in the causal reasoning.