RAND Roybal Center for Health Policy Simulation

Health and Medical Spending of the Near Elderly

Objectives

The objective is to produce a demographic and economic model which allows predicting the health and retirement status of the near elderly population (age 51 and older) for the next 50 years. The model will allow simulating a variety of intervention scenarios related to changes in medical technology as well as in individual behavior.

Data to populate the model come from the Health and Retirement Survey (HRS), a longitudinal biennial survey representative of Americans over age 51. There are three main research components in this project. First we need to model transitions in health status from one year to the next. Then we need a cost model that can be used to predict health care expenditures of individuals based on their health conditions and other determinants of health. Finally, these two components need to be combined in a Markov model which generates projections of future health and retirement status.

Progress to Date

All the components of the model have been built. The model is fully operational, and it can be used to produce forecasts of a number of health and labor related outcomes, shown in Table 1, under different scenarios. The model can be used to produce two different types of outcomes:

Lifetime statistics for a cohort age 55 in year 2002. In this case we follow the individuals age 55 in year 2002 from year 2002 to their death, and sum outcomes of interest over their lifetimes. This can be used, for example, to estimate the effect of an intervention on life expectancy, lifetime medical spending, years of working life and years of SS benefit receipts.
Yearly population statistics. In this case we follow the entire population from year 2002 to year 2050, and for each year we can compute the average or the sum of the outcome of interest. This is useful, for example, to predict changes in population size, prevalence of health conditions, and rate of health spending.

Table 1: List of variables which are predicted on a bi-annual basis in the current version of the model.

Health-Related Outcomes		Economic Outcomes
Conditions		Labor market
	Heart disease		Employment
	Diabetes		Earnings
	Lung disease	Social Security
	Cancer		Benefit receipt
	Hypertension		SS benefit amount
	Stroke		Widowhood
Functional status		Spending
	ADLs and IADLs		Total medical spending
	Nursing home		Out-of-pocket spending
	Death
Risk Factors
	BMI
	Smoking (now/ever)

In order to gain confidence in the ability of the model to forecast correctly we have performed a validation experiment: we have used the model to predict key variables, such as mortality, population and prevalence of health conditions, from year 1992 to year 2002. Since for these years we do have the observed values of those variables, we were able to compare the predicted and observed values and obtain a measure of forecast accuracy. The model performed well under this test. However, the model results are sensitive to an assumption about the correlation of shocks to health across different diseases.

The model has been used to produce analyses of the 5 following scenarios:

Obesity: All individuals who are obese in year 2002 are substituted with individuals who are overweight but otherwise identical. All new cases of obesity are immediately identified and the corresponding individuals are put on a lifetime treatment, which keeps them at BMI status of overweight (or less). Treatment cost is set at $1,200 per year.
Hypertension: All individuals who have high blood pressure in year 2002 have their blood pressure restored to normal. All new cases of hypertension are immediately identified and the corresponding individuals are put on a lifetime treatment, which prevents them from developing hypertension. Treatment cost is set at $1,200 per year.
Diabetes: similar to hypertension scenario
Smoking: all smokers stop smoking, and their smoking history is erased.
Disability: In year 2002 all individuals with 1 or 2 ADL become disability-free, and all individuals with 3 or more ADL are moved to a state where they have 1 or 2 ADL. New cases of ADL are immediately identified and the corresponding individuals are put on a lifetime treatment, which prevents them from developing further disabilities. Treatment cost is set at $1,200 per year.

Results for lifetime statistics in the obesity and smoking scenarios are shown in Table 2.

While in this table we report results for the entire cohort, the model also allows computing results specific for the population who is ever treated. An example of results at aggregate level, for the entire population rather than for a single cohort, is shown in Figure 1. The figure shows the prevalence of lung disease in the status quo and under a smoking eradication scenario.

Table 2: Example of results for lifetime statistics for two scenarios.

	Status quo	Obesity	Smoking
Life expectancy at age 55	24.72	25.27	25.49
Lifetime DALY	22.71	23.32	23.46
Lifetime medical costs ($)	275,252	287,935	295,308
Lifetime SS benefits ($)	106,986	108,710	109,104
Lifetime SS taxes ($)	18,879	19,190	19,082
Lifetime working years	6.84	7.03	7.03
Lifetime years of claiming SS benefits	17.20	17.59	17.70

Total DALYs (Millions)	149.5	153.6	154.5
Total Medical costs (Billions $)	181.3	189.6	194.5
Total SS benefits received (Billions $)	70.5	71.6	71.8
Total SS tax payment (Billions $)	12.4	12.6	12.6

Figure 1: prevalence of lung disease in the population age 55 and older under the status quo and the smoking scenario (complete smoking eradication)

Research Products

F. Girosi, D. Goldman, Y. Zheng and M. Hurd. The Impact of Healthy Behavior on Future Health Status, Spending, and Retirement. 71st Annual Meeting of the Population Association of America, Los Angeles, CA, March 30th - April 2nd 2006

F. Girosi, D. Goldman, Y. Zheng: Projections of Health Status and Utilization for Older Americans. Invited talk at the 1st meeting of the IOM Committee on the Future Health Care Workforce for Older Americans. Washington, DC, March 27th 2007. A link to this IOM activity can be found at http://www.iom.edu/CMS/3809/40113.aspx

Next Steps

One limitation of the current version of the model is that the probabilities of developing health conditions (hazards) are estimated independently of each other using logit regression. Correlation among the presence of the health conditions is then introduced in the simulation stage, by imposing an ad hoc correlation structure on the health transitions. The correlation is chosen in such a way that the model reproduces accurately the time series from 1992 to 2002. We have noticed, however, that the correlation we introduce among the diseases is stronger than what observed in practice. While this problem can be solved by simply introducing a lesser degree of correlation among disease, this action leads to deterioration in the performance of the model in our validation experiment, making the model less reliable. Preliminary results suggest that this dilemma can be solved by estimating the probabilities of acquiring the different health conditions jointly, rather then independently. We have developed an estimation strategy for this task, and realized that it requires massive computations. Therefore we are planning to write the code for this part of the model in C language, rather than Stata.

We are currently lumping all cancer cases in one health condition. We are planning to disaggregate this category as much as possible. The limiting factors are the types of cancer reported in the HRS, our main data source, and the increased complexity due the increased number of health states to consider.

A new wave of the HRS has become available, and we plan to incorporate it in our model. This will allows to enlarge the population we are studying, from the individual age 55 and older to individual age 51 and older.