For each Employer, we estimated the probability of enrollment in a managed care plan, as opposed to indemnity insurance, controlling for individual patient characteristics. The choice was posited as depending upon the following factors: patient demographics; the per se and activity-limiting condition status of the patient; whether the patient with theper se disabling condition was an employee, spouse or dependent (for Employer B only); the employment status of the primary beneficiary; and measures of past health care use (for Employer B only).
While the results of this analysis are interesting in their own right, the estimates are primarily used to formulate a proxy for unobservable individual factors (such as disease severity) that are correlated with insurance choice and may also influence use and payments. This proxy then enters into the second stage of the algorithm, in which we model utilization or expenditures. Before turning to the results of the two-stage models, we briefly discuss the determinants of the choice between managed care and indemnity insurance in the two employers.
1. Insurance Choice Results
For Employer A, the (first-stage) probability models assessed the impact of beneficiary characteristics on the probability of enrolling in the POS plan. Two models were constructed: the first included all employees with a per se disabling condition in 1995 and the second included the subsample of these employees for whom prescription drug data were available. In these models, an incremental effect (or marginal effect) of a variable is the impact of a small change in that variable on the probability of selecting the POS plan rather than the indemnity plan.
The results of the first model on insurance choice for Employer A appear in Appendix D, Table D-1. There were five statistically significant variables: gender, age, metropolitan statistical area (MSA), early retiree, and physical and mental chronic condition. The variable having the largest incremental effect was MSA: living in an MSA increased the likelihood of choosing the POS by 23.7 percentage points. The next two most important variables were early retiree and physical and mental chronic condition. Being an early retiree reduced the likelihood of enrolling in the POS plan by 11.8 percentage points, while having a physical and mental condition as opposed to just a physical condition reduced the likelihood of POS enrollment by 9.5 percentage points. Being male increased the likelihood of POS enrollment by 6.5 percentage points. Older individuals were also more likely to join the indemnity plan.
Applying the parameter estimates to the sample data to predict insurance choices yields some information on how well the model fits the data. Overall, the model correctly predicted 61.1 percent of insurance choices: 62.5 percent of POS and 60.1 percent of indemnity plan enrollees are correctly classified by this model specification.
The results from the second model, for prescription drug users for Employer A, are presented in Appendix D, Table D-9. With the exception of the variable representing the presence of a disabling mental condition only, the same variables are significant here as were for the full sample. In addition, the incremental effects have the same signs and are of similar magnitudes. Relative to having no disabling conditions, having a disabling mental condition reduced the likelihood of choosing managed care by approximately 3 percent. The model predicted the choice of indemnity insurance quite well (91.44 percent were correctly classified). However, its ability to predict the choice of the POS plan in this subsample of prescription drug users appeared to be limited: only 12.3 percent of that group were correctly classified.
For Employer B, models were run that assessed the importance of beneficiary characteristics on the probability of enrolling in the HMO plan rather than the indemnity plan. Two models were constructed: the first included data on past health care usage (which has been shown to be a good predictor of health insurance selection) and the second excluded data on past health care usage. These two specifications are presented because the second-stage results are sensitive to the inclusion or exclusion of these variables. As the results presented in Appendix D show, models without past use revealed no managed care effect.
Including past use in the first stage is somewhat methodologically controversial because one might argue that it is jointly determined with insurance choice through time. If so, parameter estimates might be biased. Alternatively, one could view past use as pre-determined (rather than jointly determined) at the time that the patient makes the decision about what plan to join in 1995. In this case, bias is not a problem. Given the non-findings for models that excluded past use, the models that include past use are our preferred specifications. For completeness, both sets of results are presented in the appendices.
Results of these two specifications for Employer B appear in Table D-13 of Appendix D. In Model 1--which included past utilization--there were six statistically significant variables: male gender, age, dependent status, early retiree status, number of outpatient visits in 1994, and the presence of both a physical and mental per se condition. The variable with the greatest impact was dependent status, followed by early retiree status. The likelihood of HMO enrollment was reduced by 24.0 percentage points for a family that had a dependent (rather than an employee) with a per se disabling condition in 1995. Being an early retiree reduced the likelihood of HMO enrollment by 14.3 percentage points. The next most important effect on this probability was for patients with both physical and mental chronic conditions. Having both conditions reduced the probability of HMO enrollment by 4.0 percentage points. The other three statistically significant variables--gender, age, and number of outpatient claim-days in 1994--produced smaller incremental effects.
In Model 2, which excluded any measure of past use, many of the incremental effects were of similar sign and magnitude. In particular, similar effects were found for gender, age, dependent status, early retiree status, and having per se disabling physical and mental conditions. Unlike Model 1, HMO membership was somewhat more likely if the family had a spouse who had a per se condition in 1995 as opposed to an employee with a per se condition.
The two models correctly predicted the same proportion of insurance choices, approximately 63 percent. Both models were better at predicting indemnity insurance choices (75 percent correctly predicted) than HMO choices (45-47 percent). In terms of this measure and the incremental effects, the models seem to be very similar. However, as previously mentioned, they led to very different results in the second stage of the model.
In sum, we found that being male and younger increased the likelihood of managed care at both Employers. In addition, early retirees and those with both mental and physical disabling conditions had a lower likelihood of managed care coverage. In Employer A, living in an MSA increased the probability of enrollment in the POS option, while in Employer B, having higher 1994 outpatient use or having the person with a disabling condition as a dependent on the insurance policy lowered the probability of HMO enrollment.
2. Utilization and Expenditure Results
A number of variables were posited to influence utilization and expenditures: demographics, insurance status (Employer B only), employment status, the number of unique MDCs in 1995, and disability status. In addition, there may be unobservable factors that affect both the first-stage (insurance choice) equations and the second-stage (use or expenditures) equations. For each Employer, we used the parameter estimates from the insurance choice equations to calculate a proxy variable for unobservable factors that may affect use and expenditures and are also correlated with insurance choice. This variable was then used in the estimations of use and expenditures in the calculation of a term that corrects the parameter estimates on the observable factors for these unobservable influences. One component of this term is a parameter that captures the correlation between unobserved factors that influence insurance selection and the outcome of interest. The estimate of this parameter, (referred to as theta in the Tables), reflects the existence (or non-existence) of the influence of these types of unobservable factors.
If the estimate of theta is statistically significant, we may reject the hypothesis that unobserved factors correlated with insurance choice do not affect the use or expenditure measure under study. In this case, unbiased parameter estimates of the determinants of use and expenditures are obtained from the two-stage technique we have outlined above. However, if theta is statistically insignificant then unbiased results on these determinants are obtained from a different algorithm, one that considers insurance choice and the determination of utilization levels and expenditures as unrelated (non-endogenous) decisions from the patient's perspective. Correct parameter estimates are obtained in this case by running simple regression models that disregard unobservable factors correlated with insurance choices. For completeness, both types of models were run for all measures and for both Employers.
Detailed results for Employer A for each utilization and expenditure measure are presented in Appendix D, Tables D-2 through D-8 and D-10 through D-12. For Employer B, results are also listed in Appendix D, in Tables D-14 through D-20. These last seven tables contain an extra column of results compared to the Tables for Employer A. As noted above, two different specifications of the insurance choice equation for Employer B were run. Model 1 includes past utilization measures in the first-stage insurance choice equation, while Model 2 excludes past utilization.
Table 6-1 and Table 6-2 below summarize the main results of these models. Specifically, they show the direction of managed care's effect on utilization and expenditures, and whether there was any evidence that unobservable factors correlated with insurance choice influenced the use and expenditure outcomes. The tables also display summary results from the simple descriptive analysis for comparison purposes. More detail about the incremental effects for each of the models is presented in Appendix D, in Table D-21 and Table D-22. The following discussion focuses on the information contained in these four tables.
In the models of rehabilitation service use and outpatient payments, the estimate of theta was statistically significant. This implies that there were unobservable factors correlated with the choice of insurance that influenced the likelihood of rehabilitation service use and the level of outpatient payments. The sign on the parameter estimate was negative. This indicates that those who were less likely to join the POS plan were also more likely (due to unobserved factors) to have used rehabilitation services and to have had higher outpatient payments. Factors such as uncontrolled-for severity of illness might have led to this result. This is consistent with the idea that sicker individuals join more generous plans. If the true health status of the population is not fully reflected in premiums, then premiums will most likely rise through time.
Controlling for the influence of such factors (as well as observable patient characteristics) yields a positive effect of POS membership on these two measures. In particular, the incremental effect of membership in the POS plan on rehabilitation service use is 33 percentage points. The effect of POS plan membership on median outpatient payments is approximately an additional $2,700 per year.
These effects are quite large. This may be an artifact of how the observable data we have interact with some of the technical features of the estimation algorithm. Note that the simple model that ignored unobservables and treated insurance choice and the level of these measures as unrelated also resulted in a positive incremental effect of POS membership. The estimate of the likelihood of rehabilitation service use from this model was much smaller--only 3.4 percent (see column 4 in Table D-21). It is likely that the true impact is somewhere between these values. For outpatient payments, the simple model produced a negative effect of POS plan membership. Ascertaining the correct sign of this effect is impossible without more comprehensive data.
Although no other models yielded evidence of unobservable factors influencing use and expenditures, there were additional statistically significant managed care effects for the number of outpatient visits, the number of prescriptions, and prescription payments. POS plan coverage decreased the levels of these measures. For a patient with the average number of visits in the sample (approximately 16), enrollment in the POS plan would have had decreased the number by slightly more than 1 visit per year, relative to having indemnity insurance. Prescription use for the mean prescription user (approximately 18 prescriptions) would fall by 3.6 prescriptions per year with POS plan membership. In terms of prescription payments, a person having the median payment ($552.63) would be expected to have $75 lower payments than her counterpart in the indemnity plan.
A comparison of columns 1 and 2 in Table 6-1 reveals that some of the differences found in the descriptive analysis were eliminated once individual patient characteristics are accounted for. In fact, controlling for differences in POS and indemnity plan enrollee characteristics actually changed the sign of the effect of POS membership on the number of outpatient visits. Once these differences are accounted for, POS membership reduces rather than increases outpatient visits.
|TABLE 6-1: The Effect of POS Versus Indemnity Insurance on Health Care Utilization and Payments, Employer A|
|Outcome Measure||Sign of
| Evidence of Unobservables
|Number of hospital admissions||-||n.s.||no|
|Number of hospital days||-||n.s.||no|
|Number of outpatient visits||+||-||no|
|Any use of rehabilitation services||+||+||yes|
|Number of prescriptions among users||-||-||no|
|Prescription drug payments among users||-||-||no|
|NOTES: A minus sign indicates that utilization or payments were lower in POS than in the indemnity plan. A plus sign indicates that utilization or payments were higher in POS than in the indemnity plan. n.s. = not statistically significant.|
For Employer B, models of the number of hospital admissions, number of outpatient visits, the use of any rehabilitation services, as well as for outpatient and total payments, yielded evidence that unobservable factors correlated with insurance choice were an important factor for these outcomes. For these measures, theta was positive and statistically significant. This implies that individuals who are more likely to join the HMO plan rather than an indemnity plan (based on unmeasured factors) are also more likely to have higher levels of use and expenditures. It may be that those that were inherently sicker (and higher users of health care) were effectively screened-out of the market for indemnity insurance by price. Although Employer B does not overtly refuse coverage to employees and their dependents, in order to maintain the financial viability of the indemnity option they raised premiums significantly. The price differential between the indemnity plan and the HMO option was quite substantial (see Chapter 2 of this report).
After controlling for patient characteristics and this type of unobservable influence, the effect of HMO enrollment on these measures was negative. (Column 2 in Table D-22 in Appendix D displays these incremental effects evaluated at the mean or median levels in the sample.) HMOs would be expected to have 192 fewer admissions per year per 1000 patients than the indemnity plan. For a patient with the mean number of outpatient claim days, HMO membership would decrease use substantially--approximately 14 visits per year. The likelihood of any use of rehabilitation services during the year is expected to be lower by 35.5 percentage points with HMO membership. For the patient with median outpatient and total payments, the estimated incremental effects of HMO coverage are quite large ($1,414 and $1,425, respectively).
|TABLE 6-2: The Effect of HMO Versus Indemnity Insurance on Health Care Utilization and Payments, Employer B|
|Outcome Measure||Sign of
| Evidence of Unobservables
|Number of hospital admissions||-||-||yes|
|Number of hospital days||-||+||no|
|Number of outpatient visits||-||-||yes|
|Any use of rehabilitation services||n.s.||?||yes|
|NOTES: A minus sign indicates that utilization or payments were lower in HMO than in the indemnity plan. A plus sign indicates that utilization or payments were higher in HMO than in the indemnity plan. n.s. = not statistically significant.|
As for Employer A, using the simple model that ignores unobservables generally yielded similar results in terms of sign. An exception was found for the likelihood of any rehabilitation use. Based on the simpler model, HMO membership increased the likelihood of the use of these services slightly. For the other measures, the simple model resulted in smaller HMO effects. It is likely that the true impact of membership in the HMO as opposed to the indemnity plan is somewhere between these values (see columns 2 and 5 in Table D-22).
Also similar to Employer A, a comparison of columns 1 and 2 in Table 6-2 reveals the importance of considering differences in enrollee characteristics when making inferences regarding the effect of plan type on the levels of use and expenditures. Controlling for differences in patient characteristics changed the sign of the effect of HMO membership on the number of hospital days and eliminated the effect on inpatient payments. While the other effects were consistent in sign in the descriptive and multivariate approaches, the magnitudes did differ.