Multistate event data generation and analysis

In this study, we (1) evaluated the performance of multinomial logistic regression (MLR) and embedded Markov chain (EMC) models in estimating the effect of a covariate on multistate transition probabilities using panel data; (2) explored the conditions under which the estimated parameters from the two models are good approximations of the true effects of covariates on transition hazards; (3) determined the most appropriate interval for data collection and the appropriate sample size given baseline transition hazards and covariate effects on hazards; and (4) applied the MLR and EMC methods to self-assessed health status transitions.;Cox proportional hazards model (with one binary covariate using the exponential or Weibull hazard function) and the MLR model were used to calculate (1) three-state transition probabilities and (2) true MLR coefficients. These probabilities and coefficients are time-dependent, and increasing the follow-up interval usually results in an increased deviation of MLR coefficients from true Cox coefficients. However, when the probability of being in an original state is close to 1, the MLR coefficients are very close to the true Cox coefficients. The Monte Carlo simulation demonstrates a satisfactory asymptotical performance of the MLR and EMC models, and the minimal sample size that will provide sufficient statistical power decreases with increasing follow-up interval. The MLR estimates are biased asymptotically in the presence of destination-state-related but not of covariate-related censoring. Estimates are not seriously biased by using follow-up intervals that are uniformly distributed around the scheduled follow-up time.;We used the Assets and Health Dynamics Among the Oldest Old (AHEAD) data to study the effects of sociodemographic factors on self-assessed transitions in health status (good, poor) and on death in elderly Americans. The MLR and EMC models provided qualitatively similar estimates, and that age, gender, race, and education are associated with transitions in health status.