Using data from the Health and Retirement Survey (HRS) and Assets and Health Dynamics of the Oldest Old (AHEAD), this paper presents estimates of the stochastic process that determines both the distribution and dynamics of health costs. We find that the data generating process for health costs is well represented by an ARMA(1,1). Furthermore, innovations to this process are close to lognormally distributed. In any given year, .1% of our sample receives a health cost shock that costs at least $80,000 in present value. Lastly, we discuss the accuracy of numerical solutions when integrating over health costs. Assuming lognormality, simple approximation rules work well.
