Estimating transition probabilities between health states using US longitudinal survey data

We use data from two representative US household surveys, the Medical Expenditure Panel Survey (MEPS) and the Health and Retirement Study (RAND-HRS) to estimate transition probability matrices between health states over the lifecycle from age 20–95. We compare nonparametric counting methods and parametric methods where we control for individual characteristics as well as time and cohort effects. We align two year transition probabilities from HRS with one-year transition probabilities in MEPS using a stochastic root method assuming a Markov structure. We find that the nonparametric counting method and the regression specifications based on ordered logit models produce similar results over the lifecycle. However, the counting method overestimates the probabilities of transitioning into bad health states. In addition, we find that young women have worse health prospects than their male counterparts but once individuals get older, being female is associated with transitioning into better health states with higher probabilities than men. We do not find significant differences of the conditional health transition probabilities between African Americans and the rest of the population. We also find that the lifecycle patterns are stable over time. Finally, we discuss issues with controlling for time effects, sample attrition, the Markov assumption, and other modeling issues that can arise with categorical outcome variables.