Mandatory Medi-Cal Managed Care: Effects on Healthcare Access and Utilization

Mandatory Medi-Cal Managed Care: Effects on Healthcare Access and Utilization Yashna Nandan April 30, 2017 Abstract Medicaid managed care is often presented as the panacea for spiraling healthcare costs. However, by changing payment structures and thus physician incentives, the managed care model has the potential to reduce patient access to healthcare. I evaluate the impact of California AB1467, which required Medi-Cal enrollees in rural counties to enroll in managed care, on measures of patient access and utilization of care. Using California Health Interview Survey data and a combination of difference-in-differences and propensity score weighted regressions, I find that AB1467 significantly increased the odds of an individual s insurance being rejected by a specialist and significantly decreased the odds of getting an appointment in a timely manner. However, no statistically significant change in emergency department use or access to primary care physicians was found. The results suggest that mandatory enrollment in Medi-Cal managed care created notable barriers to timely speciality care among California s rural population.

Introduction Medicaid, the national health insurance program for low income individuals, has experienced a notable increase in total annual costs, from $263 billion to $463 billion, between 2000 and 2012. 1 Medicaid managed care has been touted as the solution to this problem of runaway healthcare inflation, particularly for low-income populations who have complex health needs. Through this system, state Medicaid agencies contract with managed care organizations (MCOs), providing them with a monthly lump sum payment per patient. The insurance company, and thus the providers they contract with, must work within this budget to provide services to the patient. This capitated structure of Medicaid managed care starkly differs from traditional Medicaid fee-forservice payment models, in which physicians are reimbursed based on the volume of services provided. California has been one of the most enthusiastic supporters of the capitated managed care model for Medi-Cal, their state s Medicaid program. In 2012, California passed AB1467, which expanded Medi-Cal managed care into 28 rural counties and required enrollment of Medi-Cal beneficiaries into this program. Through this program, which was officially implemented by the end of 2013, more than 400,000 individuals were mandatorily transferred into Medi-Cal MCOs in the expansion counties (California HealthCare Foundation, 2015). Those exempt from forceful enrollment included children who are eligible for California Children s Services, seniors with both Medicaid and Medi-Cal coverage, and individuals with HIV/AIDS. Although managed-care is often portrayed as the magic bullet of healthcare delivery reform, the evidence on its effectiveness, both in terms of cost and quality, is ambiguous. Duggan, et al (2013) found that mandatory Medicaid managed-care enrollment did not significantly reduce Medicaid spending in the average state. However, states with higher baseline Medicaid provider reimbursement rates did experience larger reductions in spending compared to states with lower baseline rates. In a targeted impact evaluation of Arizona s state Medicaid managed-care program, McCall (1997) found significant cost reductions in the managed-care plans compared to traditional Medicaid plans. It is important to note that each state has considerable flexibility in administering managed-care options; thus, this Arizona- specific study is unlikely to be externally valid. The mechanism through which payment structures incentivize physicians 1 Values reported in 2012 dollars. A large portion of this increase may be attributable to higher enrollment rates. Nandan 1

to provide different levels of care is of particular interest to this paper. The effects of these incentive structures will plausibly manifest themselves in the patient s level of access to care. In a randomized control trial of physicians, in which physicians were presented with hypothetical fee-for-service (FFS) or capitated (CAP) patients, Henning-Shmidt, et al (2011) found that physicians significantly overtreat FFS patients and significantly undertreat CAP patients. Rational physicians should consider both profit and patient outcomes when making their decisions, although the relative weight placed on each is not known. The authors addressed this by including a measure of patient health in their design. Interestingly, they found that physicians overtreat FFS patients who are in good health but undertreat CAP patients who are in bad health. These results suggest that patients who have complex and costly health needs are better served in the FFS model. Backus, et al (2011) conducted a survey of California primary care providers and specialists, asking them about their intake of Medi-Cal FFS versus CAP patients, as well as their attitudes towards the Medi-Cal managed-care model. They found that specialists were significantly less likely to accept new Medi-Cal CAP patients compared to generalists. These effects remained significant after controlling for physician practice characteristics. The literature on the impact of Medicaid managed-care on patient access to care is particularly mixed. Sisk, et al (1996) presented an optimistic view of Medicaid managed-care s potential. In a random sample of Medicaid FFS and CAP patients in New York, they found that CAP patients reported significantly higher measures of satisfaction with their health plan as well as significantly shorter appointment wait times. In stark opposition, Burns (2009) found that adults with disabilities (AWDs) in the U.S. who are mandatorily transferred into managed-care faced significantly longer appointment wait times and were more likely to report difficulty accessing specialty care. This contradiction may reflect subtle differences in the samples used in the studies. Burns (2009) focused specifically on AWDs, a population with complex health conditions that needs both acute and long-term care. Thus, although capitations rates are adjusted based on health status, physicians may be weary of accepting new AWD patients who will need much more intensive care compared to an average CAP patient. Although many studies examine the effects of Medicaid managed-care on health access at the national level, to my knowledge, no study has specifically examined the effect of AB1467 on access and utilization of care among California s rural Medi-Cal population. This paper will fill the gap in the health economics literature by using difference-in-differences and propensity score techniques to estimate the impact of AB1467. This population may have Nandan 2

unique healthcare needs based on their environment. Eberhardt and Pamuk (2004) showed that residents of rural counties faced higher rates of arthritis, diabetes, and premature mortality compared to their urban counterparts. Thus, it is important to assess the efficacy of cost-constraining measures such as managed-care on the provision of accessible and high quality care in rural counties. Data Data for this paper were obtained from the 2011 to 2015 waves of the California Health Interview Survey (CHIS), a repeated cross-sectional survey conducted by the UCLA Center for Health Policy Research. CHIS uses random-dial telephone surveys of residents in all 58 counties to gather information on a variety of health and socioeconomic measures. This sample was divided into a pre -period, corresponding to the years 2011-2013, before the intervention was enacted, and a post -period, corresponding to the years 2014-2015. Sample selection was determined primarily by insurance status. Because this policy change only affected individuals in rural counties who are enrolled in or eligible for Medi-Cal, I restrict my sample to adults who selfreported being enrolled in Medi-Cal. Although CHIS contains information on individuals who are currently uninsured but eligible for Medi-Cal, I do not include these data in my sample. Enrollment in Medi-Cal is not automatic and many of these individuals may not choose to actually enroll, and thus would not necessarily be affected by the policy change. Because the data is not longitudinal, I do not observe individuals over time. Thus, I must construct treatment and control groups for both the pre- and post- periods. Treatment status in the pre- period is simply defined as being enrolled in Medi-Cal and living in a rural area, as defined by Claritas. 2 In the post-period, I assume that all individuals in the pre-period treatment groups moved into HMOs, since this was a mandatory transfer. Thus, post-period treatment status is defined as living in a rural area, being enrolled 2 Claritas defines urban/rural counties based on zip codes. This is a more parsimonious definition of urban areas than other measures. Thus, the measure may leave out a few counties in which AB1467 was implemented. However, this is much less problematic than a more generous measure of urban area, which would include unaffected counties in the treatment group. As a robustness check, I run the difference-in-differences and propensity score regressions using the Indian Health Services measure of rural, which defines rural at the county level and is less parsimonious than the Claritas definition. Use of the alternative measure of rural counties does not alter the sign or significance of the coefficients (Table 3, Table 6). Nandan 3

in Medi-Cal, and being enrolled in an HMO. For purposes of the difference-indifferences method, a control group must be constructed. I choose my control group to be individuals who live in rural areas and are enrolled in employersponsored HMOs. These individuals should not be affected whatsoever by the policy change, yet have similar access constraints pre-intervention due to living in a rural location. Additionally, neither the treatment nor control groups should be significantly affected by the Affordable Care Act policies, which expanded Medicaid eligibility to people up to 138 percent of the poverty line. Individuals in the pre-period treatment group were covered under the less generous Medicaid eligibility cutoffs and were unaffected by Medicaid expansion. People in the control group were covered by their employer-sponsored insurance and I assume they continue to be covered after the expansion. A complication with the CHIS data is the addition of new variables measuring healthcare access in the 2013 wave of the survey. These variables are far superior to existing variables as they capture specialist-specific access to care, self-reported measures of difficulty obtaining an appointment with primary care physicians (PCP), and appointment wait times. When presenting summary statistics for these variables, pre is considered to be 2013 (no data for these variables exist before 2013) and post is considered to be 2014-2015 (Table 1). Because of the limited availability of data from multiple years before the treatment, I am unable to test the parallel trend assumption for these variables, and thus will use a propensity score weighted regression to estimate the treatment effects rather than a difference-in-differences estimation. The variable ER Visits is observed for the entire sample (2011-2015), and takes on values from 1 to 5. Any individual with more than five ER visits in the past month will be recorded as having only five ER visits. Other factors that might affect healthcare access and utilization are health status, age, education, and race, all of which are summarized in Table 1 by period and treatment status. The dependent variables of interest are ER Visits, whether a patient s insurance was rejected by a specialist in the last year(spec Rej), whether a patient was able to recieve an appointment in a timely manner(appt Time), and whether a patient had difficulty obtaining an appointment wiht a primary care physician(diff PCP). The independent variable of interest varies based on the empirical strategy(discussed below). In the difference-in-differences regression, the variable of interest is an interaction between program status and time period. 3 In the propensity-score regressions, the variable of interest 3 Here, program status is coded as 1 if the individual is in the treatment group, and 0 if the individual is in the control group. Period is coded as 1 in the years 2014-2015, and 0 in the years 2011-2013. Nandan 4

is a dummy variable indicating whether the patient is currently enrolled in a Medi-Cal HMO. Empirical Analysis I use a combination of difference-in-differences and propensity score matching techniques to analyze the impact of the policy change on access and utilization. The variable describing number of ER visits in the past 12 months exists for the entire sample (2011-2015) and thus lends itself well to a difference-indifferences analysis. Plotting the average number of ER visits by year for the treatment and control groups shows that this measure satisfies the parallel trend hypothesis, with the treatment group s trend diverging after 2013 (Figure 1). I use an ordered logistic difference-in-difference specification to account for the ordered categorical nature of the dependent variable. As a second check of the parallel trends assumption, I run another logistic difference-in-differences regression including two leading interaction terms. These terms test for any significant differences in the pre-treatment trends. If the coefficient on these leading interaction terms are found to be significant, this suggests that the parallel trends assumption is violated. Equations for the original differencein-differences specification as well as the difference-in-differences specification with leading terms are included below. (1)ERV isits = β 0 + β 1 P eriod + β 2 P rogram + β 3 (P eriod P rogram) + ε (2)ERV isits = β 0 + β 1 P eriod + β 2 P rogram + β 3 (P eriod P rogram) + β 4 (P eriod t 1 P rogram) + β 5 (P eriod t 2 P rogram) + β 6 Controls + ε For the remaining three measures of access (insurance not accepted by specialist, difficulty finding PCP who accepts insurance, received appointment in timely manner), data are only available for years 2013-2015 and thus a difference-in-differences specification is not a possibility. Instead, I use a propensity score matching technique in which probability of treatment is determined using a set of observable characteristics. These characteristics are chosen to reasonably predict participation in the mandatory transfer to managed-care. This includes income level, health status, age, sex, education level, race, and eligibility for social programs. 4 Because this treatment was 4 Alternate treatment effect estimators such as nearest neighbor matching and inverse probability weighting were also attempted, which produced similar regression results. How- Nandan 5

mandatorily assigned, there is not much scope for selection into the treatment group, and thus the propensity score assumption that there are no unobserved differences between the two groups is most likely satisfied. A control group is then constructed using observations with similar probabilities of treatment, but who were not treated. I then run a logistic regression using only data from year 2015 on the sample, using weights obtained from the step above. This estimates the impact of treatment status on the binary measures of access (Equation 3). I repeat this procedure for number of ER visits to check the validity of the difference-in-differences specification. (3)HC = β 0 + β 1 T reatment + β 2 HealthStatus + β 3 Income + β 4 HighSchool + β 5 College + β 6 Age + β 7 Race + ε where HC represents Spec Rej, Diff PCP, Appt Time, or ER Visits Results Difference-in-Differences Regressions The results of the difference-in-differences regressions show no significant differential effect of the program on emergency department use between treatment and control groups. This is true even before controlling for differences between the two groups, such as health status, income, education, age, and race (Table 2). Column 2 of Table 2 shows that the coefficient on the interaction term between period and program status is negative, although insignificant, which supports the general trend seen in the graph of ER visits over time for the treatment and control groups (Figure 1). Fair and poor health are both statisically signifianct predictors of higher ER use, while age and college education are statistically significant predictors of lower ER use. The differencein-differences regression with leading interaction terms show that both leading interaction terms are insignificant, which confirms the lack of pre-treatment trends and the validity of the difference-in-differences approach. Propensity Score Weighted Regressions As a robustness check, the effect of Medi-Cal managed care on ER visits was analyzed using the propensity score matching method and the above results ever, the covariates were most balanced using propensity score matching. Post-estimation covariate balance using Spec Rej as the outcome measure are presented in Table 7. Covariate balance tests using the remaining outcome measures produce similar results. Nandan 6

were corroborated, with treatment being statistically insignificant (Table 5). Column 2 of Table 5 shows that treatment 5 had a significant effect on the probability of having a specialist reject a patient s insurance. Specifically, treatment increased the odds of rejection by a specialist by 2.73. Interestingly, the effect of treatment on difficulty finding a primary care provider who accepts the patient s insurance is insignificant (Column 3, Table 5). The impact of treatment on probability of getting an appointment in a timely manner is presented in column 4 of Table 5. Receiving treatment significantly decreases the odds of getting an appointment in a timely manner by a factor of roughly 1.81 (calculated as 1/0.5526). Conclusion The results presented in this paper support anecdotal evidence as well as existing literature on the access constraints faced by Medicaid managed-care patients. Patients who were mandatorily transferred into Medi-Cal managed care faced 2.73 greater odds of having their insurance rejected by a specialist, yet faced no significant difference in ability to find a primary care physician accepting their insurance. This is not particularly surprising, given the research showing that specialty physicians are more likely to change their behavior when their payment structure changes (Backus, et al, 2011). Also in agreement with existing evidence is my result that receipt of treatment decreases the odds of getting a timely appointment by a factor of 1.81. Because this measure does not distinguish between appointment wait times for specialist and general physicians, I am unable to determine which is driving the overall effect. A result that differs from existing evidence is that treatment status has no significant effect on emergency department use. Given that patients have a harder time accessing specialists (both in terms of being accepted as a patient and getting a timely appointment), it is plausible that patients would increase their use of the ER as a substitute for a specialist appointment. However, the difference-in-differences and propensity score results show that this does not seem to be happening with California s rural Medi-Cal MCO population. It is possible that patients are simply shifting their care to urgent care centers or forgoing necessary care entirely, but this question is left for further research. Limitations of this paper are largely a function of data constraints. Because many of the outcome measures only exist in years 2013 and beyond, 5 Here, treatment is coded as 1 if the individual is enrolled in Medi-Cal and an HMO, and 0 otherwise. Nandan 7

difference-in-differences techniques could not be utilized. Although propensity score methods attempt to replicate a randomized control trial by constructing a control group with similar characteristics as the treatment group, I am only able to match on observed characteristics. There may be differences in unobserved characteristics between the two groups and thus the results of the propensity score regressions are strictly estimates and causality cannot be assumed. The results of this paper suggest that managed care is not the magic bullet of healthcare delivery reform, despite its cost-curtailing features. By fundamentally restructuring payment systems, managed care creates perverse incentives for physicians, which may prevent patients from accessing healthcare. Nandan 8

References Backus, L., Osmond, D., Grumbach, K., Vranizan, K., Phuong, L. and Bindman, A. B. (2001), Specialists and Primary Care Physicians Participation in Medicaid Managed Care. Journal of General Internal Medicine, 16 : 815 821. Burns, M. E. (2009), Medicaid Managed Care and Health Care Access for Adult Beneficiaries with Disabilities. Health Services Research., 44: 1521 1541. Duggan, M., et al (2013), Has the Shift to Managed Care Reduced Medicaid Expenditures? Evidence from State and Local-Level Mandates. J. Pol. Anal. Manage., 32: 505 535. Eberhardt,M. and Pamuk, E. (2004), The Importance of Place of Residence: Examining Health in Rural and Nonrural Areas. American Journal of Public Health., 94: 1682-1686. Kemper, L (2012), On the Frontier: Medi-Cal Brings Managed Care to California s Rural Counties. California Healthcare Foundation. Henning-Schmidt, et al (2011), How Payment Systems Affect Physicians Provision Behavior- An Experimental Investigation. Journal of Health Economics.,30: 637-646. McCall, N. (1997). Lessons from Arizona s Medicaid Managed Care Program. Health Affairs, 16(4), 194-9. Sisk JE., et al (1996), Evaluation of Medicaid Managed Care: Satisfaction, Access, and Use. JAMA.,276(1):50-55. Nandan 9

Appendix Nandan 10

Table 1: Averages by Period and Treatment Status Pre 1 Post a Treatment Control Treatment Control Health Status Excellent 0.089 0.2269 0.0764 0.217 Very Good 0.1835 0.3842 0.2101 0.3885 Good 0.3484 0.2906 0.3424 0.302 Fair 0.2505 0.084 0.2688 0.079 Poor 0.1286 0.0143 0.1023 0.0134 Age 39.22 47.583 42.36 48.05 High School Degree 0.7209 0.9381 0.7394 0.9389 College Degree 0.078 0.3663 0.1419 0.4131 Race White 0.6659 0.8267 0.7217 0.8412 Black 0.0209 0.0155 0.0123 0.0142 Asian 0.0176 0.0274 0.0273 0.0201 Other 0.2956 0.1304 0.2387 0.1245 Ins not accepted by specialist 0.274 0.0718 0.352 0.0679 Ins not accepted by PCP 0.1317 0.0169 0.135 0.029 Get appt in timely manner 0.862 0.927 0.767 0.889 ER Visits 2.141 1.431 1.883 1.486 Notes: 1 Pre refers to the years 2011-2013, before AB1467 was passed. a Post refers to the years 2014-2015 after AB1467 was passed. Nandan 11

Table 2: Difference-in-Differences Results for Number of ER Visits in Last 12 Months 1 2 Period -0.2615-0.0822 Program 1.2613*** 0.6991*** Period*Program -0.46-0.2306 Health Status Very Good -0.2488 Good 0.288 Fair 0.5934** Poor 1.3556*** Household Income 1.45E-06 High School Degree 0.0487 College Degree -0.3948** Age -0.0112** Race Black 0.3475 Asian 0.201 Other -0.0078 Observations 1,021 1,021 Notes: * p<0.10 ; ** p<0.05 ; *** p <0.01 1 and 2 refer to equations 1 and 2 presented in the paper. Program is coded as 1 for individuals in the treatment group and 0 for those in the control group. Period is coded as 1 in the years 2014-2015, and 0 in the years 2011-2013. Nandan 12

Table 3: Difference-in-Differences Results using Alternative Definition of Rural Counties 1 2 Period -0.001** 0.095 Program 0.939 0.691*** Period*Program -0.451-0.357 Health Status Very Good -0.2488 Good 0.288 Fair 0.5934** Poor 1.2466*** Household Income 1.21E-06 High School Degree 0.01343 College Degree -0.6573** Age -0.0034* Race Black 0.5421 Asian 0.1002 Other -0.0345 Observations 2,698 2,698 Notes: * p<0.10 ; ** p<0.05 ; *** p <0.01 1 and 2 refer to equations 1 and 2 presented in the paper. The dependent variable is number of ER visits. Program is coded as 1 for individuals in the treatment group and 0 for those in the control group. Period is coded as 1 in the years 2014-2015, and 0 in the years 2011-2013. Rural is defined using the Indian Health Services methodology, which assigns rural/urban designation based on county. This is a less parsimonious measure than the Claritas definition. Nandan 13

Table 4: Difference-in-Differences Results with Leading Interaction Terms ER Visits Period -0.0832 Program 0.5965** Period*Program 1-0.4432 Period*Program (1 lead) a -0.3531 Period*Program (2 leads) b -0.0361 Health Status Very Good -0.2423 Fair 0.2889 Good 0.5901** Poor 1.3541*** Household Income 1.48E-06 High School Degree 0.037 College Degree -0.3980** Age -0.0117** Race Black 0.3659 Asian 0.2135 Other 0.0049 Observations 1,021 Notes: * p<0.10 ; ** p<0.05 ; *** p <0.01 The dependent variable is number of ER visits. Program is coded as 1 for individuals in the treatment group and 0 for those in the control group. 1 This is the original interaction term in which period equals 1 in the years after 2013, and 0 otherwise. a Here, period equals 1 in the years after 2012, and 0 otherwise. b Here, period equals 1 in the years after 2011, and 0 otherwise. Nandan 14

Table 5: Propensity Score Matching Results ER Visits Spec Rej 1 Diff PCP a Appt Time b Treatment 1.134463 2.3013*** 1.3898 0.5526** Health status Very Good 0.6962 1.226 0.8679 0.2424 Good 1.5348 0.6823 0.7555 0.5546 Fair 0.6809 1.4576 1.1886 0.2925 Poor 0.8663 1.6279 1.9417 0.1597* Income 1 1 1 1 High School 0.645 0.9282 1.2081 0.8049 College 0.5719 1.0578 1.7109* 0.4968* Age 0.9942 0.9765*** 0.9861* 1.0049 Race Black 1 0.7874 0.6114 1 Asian 0.8432 0.9062 0.2767 1 Other 1.044 0.6593 0.6825 1.0499 Cons 5.2646***.094121*** 0.1933*** 5.2646*** Observations 313 522 406 Notes: * p<0.10 ; ** p<0.05 ; *** p <0.01 Results reported as odds ratios Treatment is defined as being enrolled in Medi-Cal and an HMO. 1 Spec Rej takes on a value of 1 if the respondent reports that her insurance was not accepted by a specialist in the last year, and 0 otherwise. a Diff PCP takes on a value of 1 if the respondent reports that she faced difficulty finding a primary care physician who accepts her insurance, and 0 otherwise. b Appt Time takes on a value of 1 if the respondent reports being able to schedule an appointment with a doctor in a timely manner, and 0 otherwise. Nandan 15

Table 6: Propensity Score Matching Results using Alternative Definition of Rural Counties ER Visits Spec Rej 1 Diff PCP a Appt Time b Treatment 1.699 1.724*** 1.408 0.9195** Health status Very Good 3.986 1.535 1.183 0.456 Good 2.588 1.079 1.218 0.573 Fair 1.128 1.804 2.283** 0.53 Poor 1.01 1.981* 3.246*** 0.239* Income 1 1.000** 1.000* 1.000*** High School 5.664** 0.897 1.151 0.9196 College 0.255** 1.129 1.677** 0.7795 Age 0.974 0.974*** 0.985** 1.005 Race Black 2.557 0.761 0.697 1.186 Asian 0.497 0.310* 0.389* 1.045 Other 7.416** 0.563*** 0.916 1.103 Cons 15.31** 0.664 0.109*** 4.796** Observations 984 1,124 2,884 916 Notes: * p<0.10 ; ** p<0.05 ; *** p <0.01 Results reported as odds ratios Treatment is defined as being enrolled in Medi-Cal and an HMO. Rural is defined using the Indian Health Services methodology, which assigns rural/urban designation based on county. This is a less parsimonious measure than the Claritas definition. 1 Spec Rej takes on a value of 1 if the respondent reports that her insurance was not accepted by a specialist in the last year, and 0 otherwise. a Diff PCP takes on a value of 1 if the respondent reports that she faced difficulty finding a primary care physician who accepts her insurance, and 0 otherwise. b Appt Time takes on a value of 1 if the respondent reports being able to schedule an appointment with a doctor in a timely manner, and 0 otherwise. Nandan 16

Table 7: Covariate Balance using Propensity Score Method Standardized Differences Variance Ratio Raw Matched Raw Matched Health Status 0.564 0.011 0.839 1.002 Eligibility for Medi-Cal -1.24 0 0.0868 1 Receiving TANF or CALWORKS 1.1 0.023 0.307 0.975 Receiving Food Stamp Benefits 0.978-0.014 0.328 0.966 College Degree -0.79-0.021 0.499 0.959 Age -0.882-0.035 0.654 1.009 Race 0.087-0.016 1.356 1.008 Sex 0.112 0.032 0.942 1.023 Notes: Results are presented for the propensity score estimation using Spec rej as the outcome variables. Covariate balance using ER Visits, Diff PCP, and Appt Time produce similar covariate balance. A standardized difference of 0 and a variance ratio of 1 imply perfect covariate balance. Figure 1 : Parallel Trend Test Notes: Test of the parallel trend hypothesis for mean ER visits before and after AB1467. The vertical line seperates pre- and post-periods. Nandan 17