|
Does Medicaid Managed Care Affect Access To Care For The Uninsured?
This study investigates whether the implementation of Medicaid managed care from 1994 to 2001 was associated with changes in access to care for the uninsured. We used regression analysis to examine relationships between changes in county-level Medicaid managed care activity over time and changes in four measures of perceived access to care. After we controlled for sex, race, ethnicity, poverty, age, health, and education and included county fixed effects to account for unobserved county characteristics that are potentially associated with the implementation of Medicaid managed care and outcome measures, we found that Medicaid managed care has had no consistent effect on access.
During the past two decades nearly all states have implemented managed care for their Medicaid populations.1 Policies vary across states, with some relying more on primary care case management (PCCM) and others using health maintenance organizations (HMOs). In PCCM, primary care providers receive a monthly fee for coordinating medical services but still collect fee-for-service (FFS) reimbursement. In HMOs, providers receive a monthly capitated fee for providing specified services, which typically include primary care and oversight of referrals to specialists. Implementation typically occurs at the county level, and enrollment is either voluntary or mandatory. Implementation of Medicaid managed care (MMC), particularly mandatory programs, could influence care for the uninsured. First, the clinical structure it imposes could limit safety-net providers’ ability to see uninsured patients. For instance, the emphasis on primary care might cause panels of safety-net providers to fill with MMC patients, thus reducing availability for uninsured patients.2 Second, in cases where Medicaid funding is used to cross-subsidize care for the uninsured, reductions in fund flows that may accompany MMC could limit access for the uninsured.3 Increased administrative burdens associated with MMC might erode resources that could subsidize uninsured care.4 Additionally, safety-net providers face more financial risk under some MMC contracts.5 Indeed, community health centers with MMC perform worse financially and serve fewer uninsured patients than their counterpart centers without MMC.6 Alternatively, other MMC plans support safety-net providers through increased payments for primary care services and case management, as well as favorable contracting mandates.7 Previous studies, of both national and single-state programs, suggest that increasing enrollment in MMC has negative effects on care for the uninsured.8 Although this notion is cause for concern, the statistical strength of this previous work is not clear. In particular, studies in single states have limited generalizability, and the national studies rely on cross-sectional comparisons in narrow time windows. While cross-sectional studies can generate useful information, the estimated associations may not reflect causal effects if they do not fully account for variation in the characteristics of different areas. Of particular concern for MMC are variations in urban and rural environments. Furthermore, several previous studies rely on state-level measures of MMC activity, which could lead to imprecise estimates of MMC’s effects on the uninsured. In this paper we revisit the relationship between MMC and access for the uninsured, using county-level data on a nationwide sample in an eight-year period during which MMC implementation and enrollment increased dramatically.9 Measuring MMC at the county level allows more (and more specific) variation with which to identify possible effects than is possible with state-level data. By using a relatively long time period, we also could use the more robust empirical strategy of incorporating county fixed effects, which relies solely on within-county variation over time to identify the effects of MMC. This method allows us to control for all potentially relevant, time-invariant county-level factors.
National Health Interview Survey. The NHIS is an annual survey of the civilian, noninstitutionalized U.S. population.10 It contains individual-level data on access to and use of health care, demographic characteristics, health status, and insurance coverage. We used repeated cross-sections of NHIS data from 1994 to 2001. We identified NHIS respondents who were under age sixty-five and uninsured at the time of the survey (126,510 people). We removed 9,802 subjects because of incomplete responses to the survey questions of interest, leaving 116,708 for analysis. We developed three dichotomous access measures for each uninsured respondent: whether the individual (1) delayed getting care because of cost in the past year, (2) reported that needed care was not obtained because of cost in the past year, and (3) had a usual source of care at the time of the survey. All measures were developed from similarly worded questions in each survey year. We also developed a measure of the number of doctor visits in the year prior to the survey. Because the NHIS characterized responses as exact numbers in some years and categories in others, we constructed our measure by assigning respondents to one of six categories (0, 1, 2–3, 4–9, 10–12, 13 or more visits), which we could do consistently across years. For our analysis, we assigned the median value from the relevant category to each respondent. Using a dichotomous variable indicating the presence or absence of any doctor visits did not affect our results. We constructed several control variables from the NHIS data: sex, race (white, black, other), Latino ethnicity, a poverty indicator (at or above versus below 100 percent of the federal poverty level), missing income data, age (0–3, 4–17, 18–24, 24–34, 35–44, 45–54, 55–64), self-perceived health status (poor, fair, good, very good, excellent), education (elementary school or less, high school, some college, college, graduate school), and county classification as a metropolitan statistical area (an MSA with one million or more residents, an MSA with fewer than one million residents, not an MSA). The overall structure of the NHIS changed in 1997, which raises questions about comparability of data from 1994–1996 with those from 1997–2001.11 We verified that we obtained consistent results when examining 1994–1996 alone and examining 1997–2001 alone. Medicaid managed care data. We linked each uninsured person in the NHIS sample to MMC data in his or her county of residence in the given year. These data came from a survey of Medicaid state program officers; they indicate the type and year of implementation of MMC plans in each county from 1994 to 2001.12 We divided counties into five categories: (1) no MMC, (2) mandatory PCCM but no mandatory HMO; (3) mandatory HMO but no mandatory PCCM; (4) both mandatory PCCM and mandatory HMO ("mixed mandatory"); and (5) voluntary PCCM or HMO but no mandatory programs. Some counties in categories 2–4 also have voluntary programs, although such programs are relatively small. Because category 5 contains only a few counties (maximum of 8 percent) and states have increasingly relied on mandatory programs, we do not present results for this category. Regression analysis. We used two sets of regression models to study relationships between living in a county with MMC and measures of access for the uninsured. First, we estimated a cross-sectional model using an approach similar to the previous national studies.13 We controlled for sex, race, ethnicity, income, age, health status, education, and county MSA classification. We used dummy variables for each year to control for secular trends over time, as well as to capture possible differences in variable means stemming from the 1997 NHIS redesign. Because we were interested in studying the relationship between changes in MMC policy and in access measures, we limited the analyses to uninsured people living in counties that appeared in more than one survey year and in which MMC policy changed over time.14 We also estimated models including uninsured people in all counties of the study sample and obtained essentially the same results. We then estimated a second model that provided a more robust test of whether MMC affected access. We used the same dependent variables and controls, but we added dummy variables for each individual county ("county fixed effects") and removed county MSA classification, which is redundant with the fixed effects. These fixed effects controlled for the unobserved county characteristics that were constant over time and might influence access to care, such as health system characteristics, underlying population preferences, and geographic attributes. Using this specification, the effects of MMC were identified only from changes in MMC activities within counties over time. We estimated the models using ordinary least squares. Because our dependent variables were dichotomous, our regression models should be interpreted as linear probability models. Although logistic regression is often considered preferable for models with binary dependent variables, the large number of fixed effects in our second model precluded estimation of logistic regression models using standard methods. Linear probability models, however, can provide statistically valid estimates of the change in the probability of having access associated with a change in county MMC status for a typical sampled person and have performed well in other related contexts.15 Furthermore, the linear functional form should be a particularly good approximation when most explanatory variables are categorical, as is the case here.16 We also verified that results from our models without fixed effects remained the same under logistic regression. Because insurance characteristics and use of health care vary by age, we separately analyzed adults (ages 25–64), young adults (ages 18–24), and children (under age 18). We excluded those over age sixty-four because most have Medicare coverage. We also analyzed adults with self-reported fair or poor health because they have high health care use rates and may be more vulnerable than healthier adults to changes in health care policies. Sample-size limitations prevented analysis of young adults and children with poor health. Additionally, because some uninsured patients acquire Medicaid through use of the health care system, we analyzed a group of respondents with a low likelihood of qualifying for Medicaid: adult males without self-reported disabilities. As appropriate for each subgroup in the analyses, we omitted some independent variables from the regression model (such as education for regressions involving only children). We considered results statistically significant at p < .05. However, because we performed tests of twenty hypotheses (one for each combination of uninsured subgroup and access measure), we considered Bonferroni adjustment (nominal p < .0025).17 We discuss the results based on both levels of significance. We used SAS to prepare all data and SUDAAN to incorporate appropriate weights and obtain standard errors adjusted for the complex survey design of the NHIS.18
Descriptive statistics. Among counties represented in this study, those with any type of MMC increased from 40 percent in 1994 to 83 percent in 2001. Counties with mandatory Medicaid HMO enrollment increased from 6 percent to 33 percent; counties with mandatory PCCM increased from 12 percent to 25 percent; and counties with mixed enrollment increased from 6 percent to 18 percent. The share of NHIS respondents, as a whole and in each subgroup, remained essentially stable in all counties.
In terms of unadjusted prevalence, a delay in care because of cost and the number of doctor visits were stable during the study period, while having a usual source of care decreased (Exhibits 1
We compared the prevalence and trends of each access measure among counties that never had MMC, those that implemented MMC, and those that always had MMC during the study period. The only consistent difference is a slightly lower prevalence of a usual source of care in counties that always had MMC.19
Models without county fixed effects.
Exhibits 3
Models with county fixed effects. Exhibits 5  and 6 present the results of the second set of regressions, which control for county fixed effects, as well as sex, race, ethnicity, poverty, age, health status, and education, for 1994–2001. These results incorporate the fact that we used data over time and should be more robust to omitted variable bias than those from the models without county fixed effects.
The coefficients in these models are notably smaller and indicate many fewer statistically significant relationships between MMC and perceived access to care. We found no significant results associated with mandatory PCCM. All previously significant findings associated with mandatory HMO are no longer seen, with the exception of a decrease in the probability of having a usual source of care for uninsured adults. In counties with mixed mandatory enrollment, the decrease in the probability of a usual source of care for uninsured adult males without disabilities remains. In contrast, we found evidence of improved access with a decrease in the probability of needed care not obtained because of cost for uninsured adult males without disabilities and more doctor visits for uninsured children. Including a large number of county indicator variables could leave us with insufficient independent variation in the MMC variables to identify the effects of MMC. If this were true in our study, the standard errors should increase markedly in the county-fixed-effects specifications, but we would expect little change in the coefficients. Instead, we generally found coefficients that were closer to zero in absolute value and standard errors that showed only modest increases, if at all. We therefore believe that we have sufficient statistical power and that the change in findings across specifications reflects bias in the models without the county fixed effects because of the omission of confounding county characteristics. Our interpretation of the results thus far is based on statistical significance of p < .05. When we used the Bonferroni adjustment for multiple comparisons, the same general pattern of results emerged.
We found no evidence that Medicaid managed care has had any consistent effect on access to care for the uninsured. Our results contrast with those of previous studies, which suggest that MMC is associated with decreased access for the uninsured.20 The difference between our results and previous results appears to be related to methodology. Previous studies relied on cross-sectional comparisons in states with different levels of MMC penetration. When we used similar methods, we obtained similar results. By also examining a longer time period, we could study changes that occur when counties implement MMC, comparing access before and after implementation in the same counties. This approach effectively controls for all county characteristics that are constant over time, including factors such as population health status, provider mix, and safety-net characteristics. When we used this type of comparison in our fixed-effects models, nearly all of the formerly significant relationships disappeared. Although we did not design this study to determine which county characteristics might be responsible for this effect, we have two hypotheses. First, MMC was largely implemented in response to rising health care costs. Counties implementing MMC might face more severe financial constraints than those continuing to use traditional FFS Medicaid. These constraints could result in a weaker health care infrastructure in which clinics have insufficient primary care providers, specialist referral, or other resources. Additionally, social services, which could be vital in helping the uninsured navigate the health care system, might suffer. Second, the uninsured population might be different in counties implementing MMC. We controlled for basic demographic characteristics of the uninsured; however, we do not know what fraction of the county population they represented. These counties might have disproportionately large uninsured populations who place increased demands on the health care system. Further, the expectations of the uninsured might vary according to the county’s culture and history, thus affecting perceptions of access. Study limitations. Our study has some limitations. First, our findings are based on survey data, and all responses are subject to individual interpretation and recall bias. Perceived access to care may differ from actual access to care, although both deserve attention. Second, results here indicate average effects and are driven by the counties observed in the NHIS in which MMC policy changed during 1994–2001. Therefore, they might not be representative of all U.S. counties. Third, this analysis measures the effect of MMC programs, not the extent of MMC penetration. Although some Medicaid beneficiaries will be excluded from even the mandatory programs, we anticipate that penetration will be higher in counties with mandatory enrollment than with voluntary enrollment. Unfortunately, to our knowledge, no measures of county-level MMC penetration are available. Fourth, the study was neither designed nor intended to address the larger question of whether managed care more broadly defined, including commercial managed care, affects access and use for the uninsured. We only examined the incremental effect of implementing MMC holding fixed the average level of private managed care penetration in each county over the study period through the fixed effects. Finally, if implementation of MMC was driven by expectations about access for the uninsured in ways that vary within counties over time, our results could be biased. Discussions of MMC implementation, however, typically indicate that decisions to adopt MMC are based primarily on considerations of Medicaid beneficiaries and program costs rather than on indirect effects on the uninsured. Other programs affecting the uninsured and Medicaid populations, such as the State Children’s Health Insurance Program (SCHIP), also expanded during the study period and might have influenced health care delivery.21 The inclusion of year dummy variables in our models removed the effect of time trends in access that might be associated with implementation of these programs. At the same time, further analysis of the impacts of these programs could be valuable. Safety-net providers. Although previous work rightly led to concern about the impact of MMC on care for the uninsured, our findings suggest that safety-net providers are coping with the changes associated with MMC. Providers might have developed strategies to provide access to care despite mounting difficulties. Moreover, various policy interventions aimed at mitigating the potential for negative impact on the uninsured might have worked. Some states, for instance, required MMC plans to contract favorably with safety-net providers or provided incentives for such contracts through favorable auto-assignment of patients or allowing and encouraging providers to sponsor their own plans.22 Many states also maintained cost-based reimbursement for some safety-net providers. Indeed, some research suggests that managed care has not greatly reduced Medicaid spending and might even have increased it.23 The actual burden of MMC, therefore, might not have been onerous. We hasten to add, however, that attention to the impact of MMC on vulnerable populations must continue. We do not know if MMC has hurt other areas of the safety net, such as quality of care for the uninsured, or how MMC will evolve in the future. Further studies regarding effects on the uninsured will help policymakers determine if continued participation in MMC is warranted.
Jessica Haberer (j_haberer{at}yahoo.com) is a clinical and research adviser to the Chinese Center for Disease Control and Prevention, National Center for AIDS, in Beijing, People’s Republic of China. Bowen Garrett is a senior research associate at the Urban Institute in Washington, D.C. Loren Baker is an associate professor in the Department of Health Research and Policy, Stanford University, in Stanford, California. The Agency for Healthcare Research and Quality (AHRQ) provided funding for the design, conduct, and analysis of this study, as well as manuscript preparation. This study was presented at the AcademyHealth annual research meeting, San Diego, California, 6 June 2004. The authors thank Amber Barnato, Jayanta Bhattacharya, Christopher Rogers, and Alshadye Yemane for their assistance with the design and conduct of this study.
| ||||||||||||||||||||||||||||||||||||||||||||||||
Top
Study Data And Methods
Study Results
