The key results are the coefficients on the interaction between managed care and Medicaid in and . For the linear probability model, a positive coefficient indicates an increase in the probability of use for Medicaid enrollees, after the start of the managed care period, while a positive coefficient in the second continuous part of the model indicates a greater level of service use, measured in dollars, for those who have accessed the relevant service system.
Person Fixed-Effects Linear Probability Model Results to Predict Any Monthly Expenditures
Person Fixed-Effects Model of Monthly Expenditures in 2001 Dollars, Conditional on Use
The results show a strong increase in the probability of all kinds of jail use for persons on Medicaid after the introduction of managed care. There does not seem to be a change in the quantity of use, measured in dollars, conditional on using any jail services. The results show a statistically significant increase in the probability of having some psychiatric or nonpsychiatric jail hotel costs and in the probability of receiving jail mental health services. The magnitudes may appear small, generally less than one-tenth of one percentage point, but given that the overall probabilities are also less than one percent, this is a large relative increase. For example, during the managed care period, the probability of any nonpsychiatric jail days increased by 0.19 percentage points for Medicaid enrollees and the overall average in the data was around 3 percent, an increase of over 5 percent.
In contrast to the increase in the probability of the three kinds of jail costs for Medicaid enrollees, we find no significant change in the probability of using the state mental hospital by Medicaid enrollees and a decrease in state mental hospital costs for users.
We find a significant increase in the probability of nonzero expenditures for county outpatient services after the implementation of managed care. This result does not necessarily reflect an increase in utilization, but is likely due to an artifact of the payment mechanism. After managed care, the county paid providers a capitated fee even when no services are provided and thus it is likely that payments were made for more individuals than were actually using services in any given month after the carve-out.
Because there are individual-level fixed effects, and Medicaid is a time-invariant variable as defined in our data, the Medicaid dummy variable is not included in the fixed-effects regression. Using a monthly Medicaid enrollment indicator potentially creates endogeneity problems, because individuals are precluded from Medicaid enrollment during prison or jail sentences. This policy is evidently not always enforced as we found some individuals listed as being on Medicaid when they were in prison or jail. We decided to use an indicator of ever-enrolled in Medicaid to mitigate endogeneity problems. We examined this variable for evidence that more enrollment occurred later in the sample period, which may indicate either strategic enrollment behavior by participating providers possibly in response to managed care incentives or that persons increase in the severity of their illnesses over time and thus are more likely to be eligible for Medicaid services. A trend toward more Medicaid enrollment conditional on ever enrolling did not occur in our data; in fact, the conditional probability of enrollment in a given month was smaller later in our study period over the rate earlier in the sample.
The only individual-level time-varying covariate is age. Older persons tend to have a lower probability of use, and this effect is statistically significant for nonpsychiatric care and mental health services. The gradual decline in the use of services with age for persons with severe mental illness has been found by others (e.g., Domino and Salkever 2003
; Lindrooth, Norton, and Dickey 2002
The coefficient on the uninteracted managed care term indicates the difference for non-Medicaid persons in the probability of use before and after the start of the managed care contract. This coefficient is statistically significant in the jail and county outpatient cost models (see ). The time trend indicates that the probability of any jail use increased before managed care, and decreased afterward.
In models predicting the amount of jail use, conditional on any use (), the interaction between Managed Care and Medicaid is never statistically significant in any of the jail models. Therefore, controlling for other factors, the amount of jail use for those who accessed the jail system did not change after the start of the managed care contract. This may indicate that a shift toward more severely ill jail users did not accompany the increase in the likelihood of use or that if it did occur, such persons did not receive more intensive care as measured in dollars while incarcerated in jail.
Payments for county mental health services for service users decreased substantially after managed care, especially for Medicaid enrollees, on the order of a $300 per month decrease for non-Medicaid enrollees and a $500 per month decrease for Medicaid enrollees. While this type of result may be the marker of a successful capitation program from the county's perspective, it may also fuel the cost-shifting results found in the probability models.
Combining the results from the two parts on the full sample, we can calculate the predicated effect of managed care on Medicaid enrollees' total costs. We find an overall increase in spending on nonpsychiatric jail hotel costs of $1.30 per person per month (bias-corrected bootstrapped 95 percent confidence interval of $0.45 to $1.81) and on psychiatric jail hotel costs of $0.73 per person per month (CI $0.04 to $1.62). If five percent of the approximately one million persons between ages 18 and 64 in King County during our study period are in the Medicaid program, we estimate a total annual shift of $1.13 million to the jail sector (CI $0.58M to $1.61M), or almost 2 percent of the jail's 1996 adapted budget. We also find an increase in jail mental health service costs of $0.07 per person per month (CI $0.01 to $0.11), less than 1 percent of the Department of Public Health's 1996 adopted budget. We find a decrease in spending on state mental hospital costs of $1.67 (CI decrease of $13.38 to an increase of $7.77), but since the confidence interval includes zero, we cannot rule out the hypothesis of no cost-shifting to the state psychiatric hospital system. As evidence to fuel a cost-shifting motive, we find an average reduction on county-funded outpatient mental health expenditures of $7.65 per person per month (CI $12.48 to $1.14), translating to a total annual savings of almost $4.6M (CI $0.7M to $7.5M) or 6.3 percent of the county's 1996 outpatient mental health budget.
There is obviously considerable variation in the mental health needs and utilization patterns of Medicaid enrollees, our control group. Because no independent assessment of mental illness was available in our data, we opted to use the full set of individuals enrolled in the Medicaid program, even though doing so understates our results on mental health service users. We reestimated the full set of models on a subset of our sample comprised of users of county mental health services (results not reported). This group is likely more homogenous in their mental health service needs, but it may substantially undercount mental health users, as many may seek mental health services from the primary care sector (Wells et al. 1989
). We obtain virtually identical results to those reported here in sign, significance, and magnitude of the results; the one exception is that the managed care–Medicaid interaction became positive and statistically significant in the first part of the state hospital model, indicating an increase in the probability of using state hospital services after the implementation of managed care for this population. We also reran all models on a further subset of county mental health users, those with one or more diagnoses of severe mental illness. As expected, we find similar results from the two-part model in sign and significance, although the magnitude of the coefficients from the linear probability model were 60–187 percent larger. This indicates that the effects of increasing jail use were concentrated in the severely mentally ill population.
Breakpoint Sensitivity Analysis
We conduct a breakpoint analysis, similar to Piehl and colleagues (1999)
, using the technique of Andrews (1993)
to test for the period over which a break in the parameter estimates could be determined. In other words, this procedure allows us to examine whether a significant change occurred in the probability of use of the services we examined above, but perhaps at a time period different than specified in the models (April 1995). We strongly reject the hypothesis that there was no structural break in the parameter estimates for all four sets of two-part cost shifting models during the period surrounding the actual policy shift on April 1995 using a 10 percent trimming rate, supporting our original findings