- Author: Juergen Jung
- Towson University, Department of Economics
- Email: [email protected]
- Web: https://juejung.github.io/
This repository contains some of the replication codes for my paper:
Jung, Juergen (2021), "Estimating Transition Probabilities Between Health States Using U.S. Longitudinal Survey Data," Empirical Economics, forthcoming.
Please cite this paper if these codes turn out to be useful for your research.
This paper was formerly circulated as:
- "Estimating Markov Transition Probabilities between Health States in the HRS Dataset" and
- "Estimating Transition Probabilities Between Health States Using U.S. Longitudinal Survey Data"
In this project I estimate ologit (and oprobit) models of health state transition probabilities using a combined MEPS+HRS dataset so that the entire lifecycle can be tracked.
The estimates are based on predictions from ologit (oprobit) models and the Long and Freese (2014) SPost13 Stata commands are used heavily.
References:
- S. J. Long and J. Freese, Regression Models for Categorical Dependent Variables Using Stata, 3 ed. College Station, TX: Stata Press, 2014.
- Web link: https://jslsoc.sitehost.iu.edu/spost13.htm
One of the issues that arises is:
-
MEPS is observed at annual frequency with a rotating panel design, where every 2 years the panelists are swapped and
-
HRS is a 'true' panel where a cohort is followed over multiple years and interviewed every 2! years.
This means that predictions of probabilities after an ologit from MEPS result in switching matrices based on annual frequencies and estimates based on HRS result in switching matrices with a 2 year horizon.
I use the algorithm developed in: J. Chhatwal, S. Jayasuriya, and E. H. Elbasha, “Changing Cycle Lengths in State-Transition Models: Challenges and Solutions,” Medical Decision Making, vol. 36, no. 8, pp. 952–964, 2016.
to adjust the HRS-2 year probability switching matrices to 1 year probability switching matrices in order to make the results consistent between the two surveys. This stochastic root method requires to impose the Markov property on the health transition probabilities.