|Abstract #: 541-S|
|COMPARISON OF G-METHODS TO CONTROL TIME-VARYING CONFOUNDING IN A COHORT OF BONE MARROW TRANSPLANT PATIENTS.. Alexander Keil*, Jessie Edwards, Ashley Naimi, (University of North Carolina, Chapel Hill North Carolina United States)|
|Graft-versus-Host-Disease (GvHD) is a potentially serious side-effect of bone marrow transplant (BMT) among leukemia patients. However, GvHD can also occur as part of a robust immune recovery and positive prognostic indicator. Estimation of the effect of GvHD on mortality is complicated by time-varying confounding by platelet count and leukemia relapse. G-methods can appropriately address time-varying confounding and should be used to estimate the GvHD-mortality relation. The choice between g-methods should be driven by causal knowledge. But in the absence of such knowledge choice of methods is unclear. We compare three g-methods to estimate the effect of GvHD on mortality in a cohort of 137 leukemia patients followed from BMT to death or administrative censoring at 5 years: a marginal structural Cox proportional hazards model, a structural nested failure time model, and parametric g-computation. Using these methods we estimated the hazard ratio, median mortality-time ratio, and cumulative incidence ratio for the effect of GvHD on mortality, and compare the results across each method. We also use Monte Carlo simulations to compare methods when the true exposure-response relationship is known. During follow up, 73 (53%) of the patients developed GvHD, 42 (31%) experienced relapse, and 120 (88%) experienced a return to normal platelet levels. Hazard ratios (HR) marginal structural models (HR=1.18), structural nested failure time models (HR=1.20), and the parametric g-formula (HR=1.80) indicated that GvHD is associated with a slight increase in mortality, which was lower than a standard Cox model that inappropriately adjusted for time-varying confounders (HR=2.36). Time ratios and cumulative incidence ratios followed similar patterns. Simulations suggest |
that, when the correct causal model is known and fits assumptions from all three methods, that parametric g-computation yields the lowest mean squared error of the three models. These results suggest that parametric g-computation should be used when the causal model is known, and we will compare methods under model misspecification.