1. Database Description
Data about inpatient and outpatient services were obtained for these analyses from the indemnity and managed care plans offered by two large employers. Employer A is a large Employer specializing in various forms of electronic products, media and communications equipment. This Employer has offices in over 30 cities across the United States, but the structure of the health insurance benefits offered is very similar across locations. Each location offers an indemnity plan and a POS managed care plan. The benefits offered in these plans are the same across locations, but the POS plan requires lower co-payments and deductibles and offers a wider array of preventive services than does the indemnity plan. Employer B is a large state government employer. This employer offers an indemnity plan, two PPO plans, and seven HMO plans to all employees. Like Employer A, Employer B's indemnity plan charges higher co-payments and deductibles, but makes no restrictions on out-of-network coverage. Summary features of these options were described in Chapter 2 of this report.
2. Defining Disability
A critical issue in understanding the experience of people with disabilities in managed care plans is identifying the population of interest. As part of the Private Payers Study, new methods for identifying and defining disabled populations have been developed that are based on diagnoses and utilization patterns reported in medical records. The method is described in detail in Chapter 3. The resulting list of potentially disabling chronic per se conditions appear in Appendix B.
For both the switching analysis and the utilization and expenditure analysis work, individuals identified with potentially disabling per se conditions from patient claims were used to form the basic samples. For these patients, we extracted all available health service use and expenditure information during 1994 and 1995. These data were compiled into yearly per-patient summary measures. In addition, we also collected information on patient demographics, health insurance type and enrollment duration, as well as the employment status of the primary beneficiary. Descriptive statistics and details of the analytical samples precede the discussion of each set of results.
Differences in enrollee characteristics among plans need to be accounted for in order to identify the effect of managed care on health care utilization and expenditures. Higher utilization and expenditures in one insurance type as opposed to another may not reflect differences in the plan themselves, but rather underlying differences in enrollees. For example, if older individuals tend to choose indemnity plans over managed care plans and are also less healthy, then a finding of higher utilization and payments in indemnity insurance may be solely attributable to the age variation among the plan types, not to the plan itself.
We estimated the effect of plan type on utilization and expenditures controlling for two categories of confounding influences: patient characteristics available in our data and unmeasured factors systematically related to insurance choice. This second category is important to consider since a patient's true health is not completely observable to insurers, or for that matter, researchers interested in quantifying the effect of different insurance types on health care use and expenditures. As with observable factors, if these unobserved characteristics are also correlated with the outcome variable of interest (e.g. measures of utilization and expenditures), biased estimates of the insurance effect will result. This problem is termed sample selection bias. It can result in erroneous conclusions about the impact of managed care for people with potentially disabling chronic conditions, and ultimately lead to inappropriate policy recommendations.
Heckman (1976, 1979) proposed a two-step statistical method to test for the presence of this type of bias, and correct for it if present. The first step involves estimating the probability of choosing managed care as opposed to indemnity insurance. The probability of choosing managed care is mathematically transformed and used in the second step. In particular, the regression in the second step posits that the outcome of interest (e.g. outpatient expenditures) is potentially dependent upon patient characteristics observed in the data (e.g. age), the transformed probability variable, and the plan by which the individual is covered--managed care or indemnity insurance.
The test-statistic associated with the transformed probability or "selection" variable determines whether sample selection bias is present. If the coefficient associated with this term is statistically significant, one may conclude that unobserved factors related to insurance choice do affect the outcome. The sign of this coefficient indicates the direction in which these unobserved factors influence the outcome variable. In our case, a positive coefficient implies that those who have a higher likelihood of enrolling in managed care will also tend to have higher levels of the use and expenses, based on these unobservable factors. On the other hand, a statistically significant negative coefficient--those with higher likelihood of managed care enrollment tend to have lower utilization levels--is consistent with adverse selection.
Since the selection variable captures the effect of unobserved variables systematically related to insurance choice, the parameter estimate on the insurance indicator variable will be void of these influences (as well as void of any influences of the other observable factors entered in the regression equation). The test-statistic associated with the estimated coefficient of the insurance choice variable tests the null hypothesis of no difference between managed care and indemnity insurance enrollees in terms of their levels of health care utilization and expenditures. An insignificant coefficient indicates that there is no difference between those covered under managed care and indemnity insurance in level of the outcome under study. A significant coefficient rejects the null, indicating a managed care effect--levels of the outcome are different under managed care coverage as opposed to indemnity insurance.
The specific sample selection algorithm (Terza1997, 1998) that we used is an extension of Heckman's two-step procedure. Unlike Heckman's method, this method yields consistent parameter estimates for outcomes that are counts (e.g., number of doctor visits), whose distributions may contain a large frequency of zero-valued observations (e.g., the number of hospitalizations). It is also appropriate for variables that are binary (e.g., whether or not the patient used any rehabilitation services during the year).