1. Examining the Need for Risk-Adjustment
In this section we describe the 1995 utilization and expenditures of people having selected chronic and potentially disabling conditions in 1994, by type of condition. This highlights the importance of risk-adjustment for people with chronic conditions. The analyses presented are based on the full sample of data from Employer A. The results from Employer B are very similar, and thus are not presented.
Utilization and Expenditure Comparisons
Total Payments. Figure 7-3 describes the average 1995 total payments per person, including out-of-pocket payments, for people with selected chronic conditions and for all employees with any type of utilization ("all service users"). The numbers are adjusted for differences in the age and sex distribution of the disease groups. For all conditions, the average total payments for people with the chronic conditions are statistically significant and substantially higher than the average for all employees. However, there is also a significant degree of variation across the chronic conditions. The most expensive of the ten prevalent conditions is heart failure, which is on average almost 15 times as expensive as the average for the population. The least expensive is asthma, which is 2.8 times as expensive as the average.
|FIGURE 7-3. Mean Total Payments in 1995 by Chronic Condition, Adjusted by Age and Sex, Employer A|
Inpatient differences. The average length of stay was significantly higher for the chronic conditions than for all employees who were hospitalized (Figure 7-4). For example, the average length of stay for people a cancer diagnosis (malignant neoplasms) was 11 days while the average length of all hospitalizations was about 7 days. Average inpatient payments among people with hospitalizations were also significantly greater for all chronic conditions except asthma and psychiatric conditions (Figure 7-5).
|FIGURE 7-4. Mean Length of Stay by Chronic Condition for Those Hospitalized in 1995, Adjusted for Age and Sex, Employer A|
|FIGURE 7-5. Mean Inpatient Payments by Chronic Condition for Those Hospitalized in 1995, Adjusted for Age and Sex, Employer A|
Outpatient differences. Outpatient expenditures were significantly higher for all ten of the chronic condition categories (Figure 7-6). For example, the average outpatient expenditures for people with rheumatoid arthritis who used outpatient care was$2,312 versus $946 for the average employee.
|FIGURE 7-6. Mean Outpatient Payments in 1995 for People with Any Outpatient Payments, Adjusted for Age and Sex, Employer A|
Coefficient of Variation. The preceding figures compared expenditures and hospitalizations across chronic conditions. Figure 7-7 addresses a broader issue: can we predict 1995 expenditures for people with these chronic illnesses better than for all employees as a group? If so, we may conclude that adverse selection is likely to be a problem, but also that accurate risk-adjustment systems can be successful at insulating health plans from the effects of adverse selection by chronically ill enrollees.
|FIGURE 7-7. Coefficient of Variation (x100) by Condition, 1994-1995, Employer A|
Figure 7-7 describes the coefficient of variation for total payments for people with potentially disabling chronic conditions as compared to the average employee. The coefficient is measured as the ratio of the standard error to the mean and is an indication of how much variation there is in the total payments. As shown, the coefficient of variation is much smaller for people with chronic conditions than for the entire population. The smaller coefficient of variation indicates that costs for people with chronic conditions are more predictable than for people without these conditions.
Another measure of the need for risk-adjustment is the variation of expenditures over time. Figure 7-8 presents the R2 statistic from regressing 1995 expenditures on 1994 expenditures for people having chronic conditions in 1994. For some conditions, the R2 is much higher than for the whole population. Specifically, the R2 for diabetes is nearly 14 percent and the R2 for psychiatric conditions is 12 percent, much above the R2 of 6 percent for all service users. However, other conditions, such as ulcerative colitis, chronic obstructive pulmonary disease, and heart failure, the predictability of expenditures from one year to the next isrelatively low using this measure.
|FIGURE 7-8. R2 for Regression of Total Payments in 1995 on Total Payments in 1994, by Condition, Employer A|
The results presented in this section indicate that people with the chronic conditions examined had higher and more predictable expenditures than the average employee. These findings have two important implications. The first is that there is greater risk that persons with these conditions will self-select into certain plans, as well as greater risk that health plans will try to discourage them from selecting certain plans. The second implication is that risk-adjustment may be possible because the fluctuations in costs are smaller within these condition categories and therefore it makes sense to group these people together for payment purposes.
2. Predictive Ability of Risk-Adjustment Methods
Table 7-3 summarizes the prediction errors from the 30 percent samples across the various models tested using data from Employer A and Employer B. The prediction error shows the percent difference in the predicted and actual expenditures in 1995. A perfect prediction would yield a value of 0 for the prediction error. A prediction error greater than 0 indicates that the model is overestimating expenditures, and a prediction error less than 0 indicates that the model is underestimating expenditures.
These results might be evaluated in a number of ways to determine which model is most likely to reduce incentives for adverse selection based on common chronic conditions. We argue that the most important criterion is that the model minimizes the incentives for large monetary gains that result from avoiding the enrollment or treatment of people with a given condition. Thus, the evaluation criteria we select is the model with the fewest prediction errors greater than or equal to 15 percent. The 15 percent figure was arbitrarily chosen because there is no accepted benchmark for the predictive accuracy of risk-adjustment models. Ideally we would like to know the level of prediction error that eliminates incentives to avoid high cost/need patients. The shaded values in Table 7-3 indicate that the prediction error was greater than 15 percent.
According to this criterion, the model that performed the worst across all of the chronic conditions and both employers was the baseline model. The Baseline+ACG model performed significantly better, although it performed worse than the other remaining models and had prediction errors in the 40-50 percent range for some conditions. The best performing model was the HCC model. Across both employers, the HCC model only had one prediction error that was greater than 15 percent (it over predicted the cost of treating people with a stroke by 21 percent for Employer A).
The HCC-SW model performed slightly worse than the HCC model. For example, for Employer B it underestimated the costs of people with diabetes, heart failure, and stroke and overestimated the costs of people with ulcerative colitis by more than 15 percent. The HCC-SW model, developed by Ash and colleagues (1989), was based on over one million claims, allowing it to capture relatively rare high cost conditions. On the other hand, it was estimated with data from 1992 and 1993 and on a population that may differ significantly from the one being examined here. For example, technological innovations may have increased the cost of treating some diseases, such as cardiac-related illnesses, and decreased the cost of treating others, such as psychological illnesses.
|TABLE 7-3. Mean Prediction Error1 by Type of Condition, Model, and Employer, 1995 (%)2|
|TABLE 7-4. Plan's Financial Gain/Loss by Type of Condition, Model, and Employer, 19951|
The same criterion that was applied to compare models can be applied across the conditions to see whether there are certain conditions for which all the models tend to perform relatively poorly. Using this criterion, the models performed worst for heart failure and stroke. Three out of the four models underestimated or overestimated costs for heart failure and stroke for each of the two employers by more than 15 percent. It is interesting to note that heart failure and stroke had the highest average costs across the conditions examined but also had lower coefficients of variation and a higher R2 than other conditions with lower prediction error. Based on the lower coefficient of variation and higher R2, one might have guessed that heart failure and stroke expenditures would be more predictable.
Although the models tended to have large prediction errors for heart failure and stroke, the prediction errors were not always in the same direction. The HCC-SW model underestimated the cost of heart failure by 18 percent in the case of Employer B, but overestimated the cost of heart failure by 27 percent in the case of Employer A. The only condition that was consistently underestimated by all the models was chronic obstructive pulmonary disorder, although in some cases the underestimation was only 4 percent.
As described above, the ability-limiting conditions are chronic conditions identified by LaPlante (1989) as being highly associated with limitations in ability to perform activities of daily living. As shown in Table 7-3, the HCC, HCC-SW, ADG models all performed fairly well for these conditions, with an average prediction error of less than 10 percent.
In the next set of results, we simulated the impact of enrolling patients with chronic conditions with and without risk-adjusted payments. This was achieved by subtracting the mean of the actual expenditures for treating the chronic conditions from the mean predicted expenditures and multiplying the difference by the number of people with those conditions. Thus, the size of the loss is directly related to the accuracy of the risk-adjustment model, the cost of the illness, and the number of people with the condition. Table 7-4 shows the result of this exercise for employees of Employers A and B by model and type of chronic condition. Under the demographics and wage-index risk-adjustment, the losses are in the millions. For example, Employer A's indemnity plan would have lost approximately $84.7 million from treating its approximately 9,505 continuously enrolled beneficiaries who had ability-limiting conditions as defined by LaPlante (1989). Under the HCC risk model, the loss would have been reduced to only a $2 million benefit from screening out people with cancer, which would hardly be worth the financial cost of identifying the beneficiaries.