Accommodating informative visit times for analysing irregular longitudinal data: a sensitivity analysis approach with balancing weights estimators

Irregular visit times in longitudinal studies can jeopardise marginal regression analyses of longitudinal data by introducing selection bias when the visit and outcome processes are associated. Inverse intensity weighting is a useful approach to addressing such selection bias when the visiting at random assumption is satisfied, i.e., visiting at time $t$ is independent of the longitudinal outcome at $t$, given the observed covariate and outcome histories up to $t$. However, the visiting at random assumption is unverifiable from the observed data, and informative visit times often arise in practice, e.g., when patients' visits to clinics are driven by ongoing disease activities. Therefore, it is necessary to perform sensitivity analyses for inverse intensity weighted estimators (IIWEs) when the visit times are likely informative. However, research on such sensitivity analyses is limited in the literature. In this paper, we propose a new sensitivity analysis approach to accommodating informative visit times in marginal regression analysis of irregular longitudinal data. Our sensitivity analysis is anchored at the visiting at random assumption and can be easily applied to existing IIWEs using standard software such as the coxph function of the R package Survival. Moreover, we develop novel balancing weights estimators of regression coefficients by exactly balancing the covariate distributions that drive the visit and outcome processes to remove the selection bias after weighting. Simulations show that, under both correct and incorrect model specifications, our balancing weights estimators perform better than the existing IIWEs using weights estimated by maximum partial likelihood. We applied our methods to data from a clinic-based cohort study of psoriatic arthritis and provide an R Markdown tutorial to demonstrate their implementation.