Feb 19 Marginal Structural Models for Multi-State Outcomes

[OMS 3905] The progression of chronic disease is often described in discrete stages clinically. For example, patients with chronic kidney disease are classified into different stages based on the glomerular filtration rate (GFR), a measure of kidney function. Multi-state models are well-established for modeling transitions among different stages. However, when estimating the effect of time-varying treatments on transition rates, the model typically gives biased estimates when there is time-dependent confounding. Time-dependent confounding occurs when some time-updated covariates confound the association between subsequent treatments and outcome and are themselves affected by prior treatments. Inverse probability of treatment weighting can be used to handle time-dependent confounding. In the analysis, each subject is weighted by one over the probability of receiving the treatment that is consistent with her treatment history. Using the Cox proportional hazards model, the causal hazard ratio for the transition between any two stages can be estimated using the weighted sample without further adjustments for the time-updated covariates. We illustrate the method through simulation of data from an ongoing cohort study of patients with chronic kidney disease. This is a collaborative work with Alisa Stephens-Shields and Marshall Joffe.