Modeling Health Status for Microsimulation

1
Modeling Health Status for Microsimulation

• Julie Topoleski
• Joyce Manchester
• U.S. Congressional Budget Office
• The views expressed in this paper are those of
  the authors and should not be interpreted as
  those of the Congressional Budget Office.

2
Motivation

• The Congressional Budget Office analyzes
  long-term effects of fiscal policy for the U.S.
  Congress
• We developed a microsimulation model to examine
  Social Security over the next 75 years
• We are now exploring Medicare, Medicaid, and
  other health expenditures as well
• We need health status and transitions to model
  health expenditures

3
CBOs Long-Term Microsimulation Model

• CBOLT is based on longitudinal, administrative
  earnings from the Social Security Administration
• We impute many variables onto the administrative
  data using survey data
• CBOLT makes projections of Social Security,
  Medicare, and Medicaid
• But MM not currently projected at micro level

4
Background

• Large body of evidence indicates that an
  individuals educational attainment and earnings
  are strong predictors of health
• Relationship also works in the opposite direction
• Including health status as a factor that affects
  both demographic and economic outcomes is
  critical when projecting financial solvency of
  public pension and health programs

5
Projecting Health Status

• Use data from U.S. Survey of Income and Program
  Participation merged with administrative earnings
  and benefit data to estimate health status
  transition equations
• Predict self-reported health status
• Excellent, very good, good, fair, poor

6
Three-step approach

• First, estimate health status transitions
  assuming only age, sex, and lagged health status
  matter
• Second, add additional covariates
• Education, marital status, household earnings
  quintile
• Third, add health status to mortality equations
• At each step, examine relationship between
  predicted health status and various
  characteristics (educ, earnings, mortality, )

7
Age-centered regressions

• Allows flexibility in the relationship between
  health status and the underlying determinants
  across age groups
• Given the limited available data, estimating the
  equations using data for just a single year of
  age yields imprecise results
• Age-centered approach uses every observation for
  ages within a pre-set band (typically 4 years)
  around the specific age group being analyzed

8
Assigning health status

• Given results of ordered logits and initial
  health status assignment, each simulated
  individual receives probability of transitioning
  into each health status
• Outcome depends on combination of probability and
  a random draw
• Logit rank procedure ensures right number of
  individuals in each health status

9
Step 1a Health Status by Education(Health
status related only to age, sex, and lagged
health status)
10
Step 1b Health Status by Earnings Quintile
(Health status related only to age, sex, and
lagged health status)
11
Step 1c Average Mortality Rates(Health status
related only to age, sex, and lagged health
status)
12
Step 2a Health Status by Education(health
status depends on age, sex, lagged health and
other covariates)
13
Step 2b Health Status by Earnings Quintile
(health status depends on age, sex, lagged health
and other covariates)
14
Step 2c Average Mortality Rates(health status
depends on age, sex, lagged health and other
covariates)
15
Step 3 Average Mortality Rates(add health
status to mortality equations)
16
Comparison of Mortality Rates SIPP vs. CBOLT
17
Next Steps

• Additional work on refining the equations used in
  the microsimulation model
• Health could operate through many other pathways
• Disability status, labor force participation,
  earnings levels, fertility, and even marital
  status
• Ultimately want to add health insurance status
  and health expenditures to micro model