2011 National Meeting

1087 — Time-Dependent Survival Analysis: Its Benefits and Hazards

Zeringue AL (STL-VAMC, Washington University) , Al-Aly Z (STL-VAMC), McDonald JR (STL-VAMC)

Survival analysis is increasingly applied to cohort studies utilizing administrative data. As exposures and comorbid conditions change over time, not utilizing time-dependent covariates in survival analysis can introduce immortal time bias. However, treating an endogenous (reverse causation) variable as time-dependent can also bias results. We demonstrate the impact on analysis of misclassification of such variables.

Three longitudinal datasets (all containing intervention, outcome, and comorbid conditions) were used for these analyses: 1. A simulated dataset; 2. A study database examining the effect of immunosuppressive medications on herpes zoster (HZ); and 3. A study database examining the impact of kidney disease and care on mortality. Data were analyzed the following ways: 1) all variables treated as time-dependent, 2) all variables assume the values they had at baseline, 3) ‘exposed’ group starts at the first exposure, control group starts at beginning of study period, and covariates are defined by baseline values, 4) timing of the exposure is the same as 3, but the covariates were defined by whether they ever developed in the study period.

For the analysis of simulated data, time-dependent modeling (#1) produced results that were closest to the true hazard ratios (HRs). For each variable in model #1, the true HRs fit within the confidence intervals of the estimated HRs. The other models performed poorly on the variables which tended to develop later in the study period, but performed well on the comorbid condition which was most often present at baseline. The only exception was approach #4, which poorly estimated all variables, even reversing the direction of effect in one variable. For the HZ analysis, time dependent modeling produced covariate estimates which were the most consistent with those found in the literature. Other modeling approaches gave poorer or biologically improbable estimates. In the mortality analysis, modeling nephrology referral as time-dependent incorrectly estimated it as a significant risk factor.

Time-dependent analysis better estimated all variables, except for nephrology referral. However, because sicker patients tended to be referred, this variable is endogenous.

This study is designed to help researchers reduce bias in analyses using longitudinal data from administrative databases.