Disruptions to the Patient-Provider Relationship and Patient Utilization and Outcomes: Evidence from Medicaid Managed Care

Becky Staiger

doi:10.1016/j.jhealeco.2021.102574

. Author manuscript; available in PMC: 2023 Jan 1.

Published in final edited form as: J Health Econ. 2021 Dec 27;81:102574. doi: 10.1016/j.jhealeco.2021.102574

Disruptions to the Patient-Provider Relationship and Patient Utilization and Outcomes: Evidence from Medicaid Managed Care

Becky Staiger ^a

PMCID: PMC8815618 NIHMSID: NIHMS1768728 PMID: 34968786

Abstract

The patient-provider relationship is considered a cornerstone to delivering high-value healthcare. However, in Medicaid managed care settings, disruptions to this relationship are disproportionately common. In this paper, I evaluate the impact of a primary provider’s exit from a Medicaid managed care plan on adult beneficiary healthcare utilization and outcomes. Using an event study approach, I estimate a 5% decrease in the number of beneficiaries with primary care visits in the year following the exit, with slightly larger effects in terms of percentage points for patients with chronic conditions. Additionally, I observe a nearly 50% increase in the number of beneficiaries with a chronic condition who are hospitalized following a disruption.

Keywords: Medicaid managed care, patient-provider relationship, care continuity, disrupted care, chronic conditions, I12, I13, I14, I18

1. Introduction

The relationship between a patient and their doctor is widely regarded as a critical component in achieving good health outcomes (Institute of Medicine, 2001). Patient-provider relationships that are longitudinal and uninterrupted have been associated with promoting patient engagement in primary and preventive care (O’Malley et al., 1997; Saultz and Lochner, 2005) and minimizing risks of adverse health events, particularly in at-risk patients such as medically complex individuals and those with chronic conditions (Bayliss et al., 2015; Bindman et al., 1995; Hussey et al., 2014). However, in Medicaid managed care, which covers over 80% of Medicaid beneficiaries via state-contracted managed care plans (Mathematica Policy Research, 2019), institutional factors such as relatively limited provider networks and relatively high provider churn out of networks suggest disproportionately high rates of truncated patient-provider relationships (Ndumele et al., 2018). Moreover, there is sparse empirical evidence regarding the effects of these disruptions on patient utilization and wellbeing, which can be challenging for policymakers and providers who may need to consider interventions that trade off improving care delivery and access with disrupting existing relationships.

In this study, I estimate the causal effect of disruptions to the patient-provider relationship in Medicaid managed care by evaluating the effect of a provider’s exit from a Medicaid managed care network on their patients’ utilization and health outcomes. This approach is similar to several recent papers that have exploited changes in provider availability to the patient, like retirement or practice closures, to estimate the relationship between disrupted relationships and outcomes (Kwok, 2019; Sabety et al., 2021; Simonsen et al., 2019; Zhang, 2018). However, these investigations have featured settings with a single insurance provider or health system, and where alternative providers are relatively accessible. By focusing on disruptions in Medicaid managed care networks, I aim to fill in the gap in our understanding of disrupted patient-provider relationships in more restrictive settings in which a provider’s exit from a plan may significantly limit a patient’s ability to access care.

I use a robust administrative Medicaid claims database, the Medicaid Analytic eXtract (MAX), to evaluate the impact of disruptions to the patient-provider relationship among Medicaid managed care enrollees after their primary provider exits their plan’s network. Because managed care organizations discourage the use of out-of-network providers, this exit likely represents a truncation of the relationship. I focus on the exit of a beneficiary’s “key provider,” a physician or advanced practice nurse of any specialty who is responsible for the plurality of the patient’s outpatient encounters in their first six months of enrollment. This definition is intended to flexibly identify the provider who is most important to the patient, and whose exit represents a significant emotional and/or clinical disruption to care.

I use an event study to estimate the effect of an exit on a patient’s utilization of primary, inpatient, and emergency care in the following year. To improve the precision of my estimates, and to recover a single summary number, I also estimate a pre-post model on the four quarters before and after the exit. I find an average 5% decrease in the number of patients who have a primary care visit following a provider’s exit, and no significant effect on hospitalizations or emergency care use.

I then explore heterogeneity in the treatment effects for a particularly vulnerable and high-risk subset of the population, patients with chronic conditions, for whom continuity in patient-provider relationships has been shown in associational studies to be particularly beneficial. Among this population, I find a slightly higher percentage point decrease in the probability of a primary care visit after a provider’s exit (4.2 versus 3.4 percentage points for chronic and non-chronic enrollees, respectively), and a significant increase of 2.1 percentage points in the probability of hospitalization following a provider’s exit, or an increase of approximately 50% of the pre-exit mean. Finally, I examine how these effects differ along a measure that is analogous to the degree of continuity a patient has with their provider. I find evidence of increasing effects of disruptions on outcomes of interest as continuity increases, consistent with prior associational observations.

This paper contributes to three areas of research. First, it builds on and extends existing literature on access to providers in Medicaid. Medicaid beneficiaries disproportionately experience disrupted care through a variety of different pathways in addition to provider exit, including an individual’s own churn into and out of Medicaid due to state eligibility thresholds. Banerjee et al. (2010) found that 43% of Medicaid enrollees changed coverage annually between 2000 and 2004, while Sommers et al. (2016) found that even after states expanded access to Medicaid through the Affordable Care Act, approximately 25% of Medicaid beneficiaries had changed coverage within twelve months of being surveyed. Ji et al. (2017) established a link between hospitalizations and gaps in Medicaid coverage through disenrollment among specific populations such as adults with major depression, though this study did not include the Medicaid managed care population. In addition to disruptions, Medicaid enrollees face disproportionate challenges in gaining access to providers. Evidence suggests that low reimbursement rates make providers reluctant to treat Medicaid enrollees in general (Decker, 2012), and that, relative to other insurers such as Medicare and private insurance, provider participation in Medicaid is significantly lower (Health Affairs Blog, 2019). Furthermore, there is evidence that not all providers reported as participating in a Medicaid managed care network will actually treat Medicaid enrollees (Office of Inspector General, 2014; Wallace et al., 2020).

Second, this paper contributes to the ongoing discussion on the importance of the provider-patient relationship. Recent research in health economics, in which there is a growing interest in determining the relative contributions of patients and physicians on variations in healthcare utilization, has established a significant influence of primary care physicians’ (PCP) practice style on an individual patient’s spending and utilization, estimated from patients switching PCPs (Fadlon and Van Parys, 2020; Kwok, 2019). Additionally, researchers have found significant but mixed evidence of PCP exit on patient-level utilization. In Medicare, PCP retirement has been linked to short-term reductions in primary care utilization, with evidence of substitution towards urgent care and emergency departments (Kwok, 2019; Sabety et al., 2021; Zhang, 2018). In Denmark, practice closures have been found to increase chronic condition diagnoses and slightly increase emergency department utilization (Simonsen et al., 2019). These studies, which significantly contribute to our understanding of disrupted care, cannot necessarily be externalized to Medicaid, which features more vulnerable beneficiaries and more restrictive provider networks.

Third, the paper adds to our understanding of the efficacy of interpersonal continuity in care. Interpersonal continuity of care describes a patient-provider relationship that is “characterized by loyalty and trust” (Saultz and Albedaiwi, 2004), and generally is cultivated over a sustained period of time. Prior research has established consistent associations between continuity in relationships and positive patient experience, including an increase in patient satisfaction with care (Fan et al., 2005; Saultz and Albedaiwi, 2004; Tingley, 2018) and patient trust in their physician (Mainous et al., 2001). Studies that examine vulnerable populations have reported correlations between continuity and often small decreased risks of inpatient and/or emergency department admissions among seniors (Bayliss et al., 2015; Nyweide et al., 2013), veterans (Chaiyachati et al., 2014), Medicaid enrollees (Gill and Mainous, 1998; Gill et al., 2000), and other vulnerable populations (Tingley, 2018). Turnover among PCPs has been linked to decreased patient satisfaction in the Veterans Health Administration (VHA), but no significant effect on ambulatory care quality (Reddy et al., 2015). While these findings have contributed to our understanding of the importance of continuity in patient-provider relationships, there have been challenges in differentiating between association and causation of continuity and outcomes of interest (Saultz and Albedaiwi, 2004; Saultz and Lochner, 2005).

This paper proceeds as follows. In Section 2, I briefly discuss relevant institutional details of Medicaid managed care. I provide details on my data sources and sample construction in Section 3, and discuss key measures relevant to the study in Section 4. I discuss my empirical approach in Section 5, and present results and robustness checks in Section 6. Section 7 concludes with a discussion of the findings, and a back-of-the-envelope calculation of the costs of disruptions to the most affected populations.

2. Medicaid Managed Care

One in five Americans are currently covered by Medicaid, representing the most economically and socially vulnerable children, elderly, disabled, and low-income adults. Over the past several decades, states have shifted from paying providers directly for treating enrollees (on a fee-for-service basis) to contracting with managed care plans who assume responsibility for their enrollees’ healthcare. Delivery of Medicaid services through managed care plans has become the predominant form of coverage for Medicaid enrollees, growing from 9.5% (2.7 million enrollees) of all Medicaid in 1991 to 82% (66 million enrollees) in 2017 (Kaiser Family Foundation, 2001; Mathematica Policy Research, 2019). These plans are a mix of for-profit and not-for-profit, Medicaid-only and commercial insurance plans. Under the most common arrangement of comprehensive (risk-based) managed care, states pay the managed care organizations a capitated rate in exchange for the plan assuming some or all of the financial risk for covering the Medicaid beneficiary’s health care. My study sample focuses on beneficiaries enrolled in these comprehensive managed care (CMC) plans. At their time of enrollment, beneficiaries either choose or are assigned to plans and/or providers. Assignment mechanisms and frequencies vary by state, and are largely unknown with respect to providers, though slightly more is known about plan mechanisms and rates (Smith et al., 2015).

Once a plan is selected (or auto-assigned), CMC enrollees have access to the network of providers contracted by the managed care organization, though not all providers listed as participating in the network may actually be willing or available to see Medicaid enrollees. A 2014 report by the Office of the Inspector General found that less than half (49%) of providers in a random, representative sample of Medicaid managed care participants were available to new Medicaid patients, suggesting that “Medicaid managed care enrollees may not be able to make appointments with as many as half of the providers listed by their plans” (Office of Inspector General, 2014). State contracts with managed care plans generally impose provider network adequacy thresholds to ensure some minimum level of access to care for beneficiaries, though states have flexibility on determining those standards and in enforcing them (Hinton et al., 2019). For example, in 2018, eight states required a PCP-to-beneficiary ratio from between 1 to 1,200 to 1 to 2,500 (Rosenbaum et al., 2018). Providers may exit plan networks for a variety of reasons that include retirement, relocations, or preferences, as well as plans’ efforts to manage spending and quality (Flynn et al., 2002; Howard, 2014). Ndumele et al. (2018) find an annual “churn” rate of PCPs out of CMC’s of approximately 12%, compared to a 9% annual churn in the VHA (Reddy et al., 2015), with five-year churn significantly higher in narrow networks (by approximately 20-percentage points) that employ less than 30% of physicians in the plan’s market.

3. Data Sources and Sample

3.1. Medicaid MAX Data

I use Medicaid claims data from the Medicaid Analytic eXtract (MAX) for six regionally diverse states—Arizona, Indiana, Kentucky, New Jersey, New Mexico, and Washington—for 2009 through 2014. Despite being the sole national database of Medicaid administrative claims, this data has historically been underused due to concerns regarding integrity of the state-reported managed care claims, or “encounter” data. Given that the vast majority of beneficiaries are enrolled in a CMC, the usefulness of this dataset has thus been limited. However, Mathematica Policy Research (MPR) has published two reports identifying states whose encounter records are sufficiently usable (i.e., have minimal data quality concerns) for 2007–2011 (Byrd and Dodd, 2012, 2015). The six states chosen for this study were selected due to their high usability according to these briefs. While similar reports on the usability of 2012 through 2014 data are not available, both the consistent usability of these states in 2007–2011 and the general trend in improved data reporting quality suggest that it is reasonable to use these additional unvetted years (in these states) in order to generate a longer panel of available claims data. Additional information on how I clean the sample is available in Appendix H.

In addition, through developing my study sample, I employ my own validation checks and thresholds to ensure that a sufficient amount of non-missing information is available for enrollees, providers, and plans for all states in this study period. Through these validation exercises, I drop six additional states that had originally met the MPR validation thresholds for adequate encounter data integrity, but that ultimately had a large amount of missing information for providers, whom I identify using the National Provider Identifier (NPI).¹ In the remaining states, a substantial amount of missing or incorrectly reported NPI’s remain, particularly since some states identify providers using a state-specific legacy provider identifier (LPI) that is different from the NPI and not necessarily consistently linked to one provider across state and time. I undertake a three-step process to replace the LPI’s with NPI’s when possible, which includes the use of a crosswalk between NPI and LPI created by Mathematica for 2009–2011; matching the LPI to the “other provider identifier” in the National Plan and Provider Enumeration System (NPPES) downloadable file; and the creation of a crosswalk between LPI’s and NPI’s across years to fill in any remaining missing NPI’s, when appropriate. This exercise reduces missing NPI’s in a given state-year from 100% in some state-years, to generally less than 12% (with the exception of Washington in 2009, which remains missing at 20%, down from 85% missing). Detailed information on this process, and on rates of missing NPI’s at each step, is available in Appendix J.

While the MAX data offer a wealth of information regarding beneficiaries, their providers, and the plans they enroll in, I use supplemental resources to further identify provider and plan characteristics. I obtained information on provider gender and age through Doximity, a professional medical network for clinicians, and supplemented age information with year of graduation obtained from the publicly available Physician Compare files. I also utilize the NPPES Downloadable File produced by the Centers for Medicare & Medicaid Services (CMS), which contains information on all providers with a National Provider Identifier (NPI), such as name, gender, specialty, and location of practice. I use the Managed Care Crosswalks published by MPR to flag the plan identifiers in the MAX data that belong to managed care organizations, ultimately limiting my sample to enrollees in these plans and dropping those beneficiaries enrolled in non-CMC plans. Used in conjunction, these datasets provide an additional level of detail that can be used to bolster information on provider and plan characteristic and fill in gaps in data as they exist, as well as to provide additional checks on data quality.

3.2. Patient Sample

My study sample includes all CMC enrollees in the six sample states who were eligible as adults (18-65) from 2009 to 2014, and who were continuously enrolled in a CMC for at least 15 months at some point during this time.² I exclude pediatric beneficiaries who may have less agency in who their key provider is (i.e. whose key provider may reflect the preferences of their parent or guardian), and exclude any individuals who are also dually enrolled in Medicare, and for whom I would not have access to their full set of claims. I identify enrollment episodes during which time a beneficiary was recorded as being enrolled in a CMC plan. I restrict the sample to beneficiaries who were enrolled in one managed care plan throughout their entire enrollment episode in order to avoid having to disentangle discontinuities resulting from provider exits from discontinuities in plan switches. Notably, 19% of beneficiaries switch to a different plan during a given enrollment episode. For the majority of adult Medicaid enrollees in these states during these years, CMC enrollment was mandatory. See Table C.19 in Appendix C for additional details.

In a subsample analysis, I focus on patients with chronic conditions. Chronic conditions are disproportionately common among Medicaid enrollees as compared to the general non-Medicaid population (Holahan et al., 2010; Kaiser Family Foundation, 2012), and individuals with these conditions are believed in particular to benefit from a continuous relationship with a provider (Gill and Mainous, 1998; Cabana et al., 2004). I use the diagnostic definitions provided by the Chronic Condition Warehouse to flag patients with at least one of four relatively common chronic conditions: chronic heart failure (CHF), chronic obstructive pulmonary disease (COPD), diabetes, and hypertension.³ In my final study sample, between 25%-30% of the study sample has at least one chronic condition. The estimated 30% prevalence in the study sample is consistent with a recent literature review (Chapel et al., 2017), in which the authors report the following rates of chronic illness among adult Medicaid enrollees: 17.2–27.4% had hypertension (compared to my estimated 18–22%); 7.5–12.7% had diabetes (compared to my estimated 9–12%); and 55.7–62.1% had one or more chronic conditions (including more than the four I focus on).

4. Measures

4.1. Key Provider

I focus on the truncated relationship between a beneficiary and their “key provider” (“KP”). I define key providers as the physician (Medical Doctor, MD; or Doctor of Osteopathy, DO) or advanced practice nurse (Nurse Practitioner, NP; or Physician Assistant, PA) who has at least three outpatient encounters with a beneficiary within the first six months following their first outpatient encounter (defined below) in a managed care plan, and who is associated with the plurality of a beneficiary’s claims in that period. As a test of the sensitivity of this time frame with respect to identifying a key provider whose exit is significant to the beneficiary, I also identify a cohort of beneficiaries with a 12-month key provider definition period. I compare characteristics of the 6-month and 12-month cohorts in Appendix D, and discuss results for the 12-month cohort as a robustness check in Section 6.

To identify plurality, I calculate the share of encounters associated with a particular provider out of all encounters a beneficiary has with any provider during the 6-month period. This definition is similar to the Bice-Boxerman Continuity of Care index that measures the degree of continuity that a patient has with an individual provider or group of providers (Bice and Boxerman, 1977). Additional details on this definition are available in Appendix I. I define outpatient encounters as any claim that has 1) a place of service code indicating an office, walk-in retail clinic, independent clinic, federally qualified health center (FQHC), or rural health clinic; 2) a type of service code indicating the service was provided by an individual clinician, specifically: physicians, other practitioners, nurse midwife services, or nurse practitioner services (to avoid flagging an NPI or legacy provider identifier associated with an institution such as a pharmacy or lab as the key provider); and 3) a type of claim code that classifies the visit as an encounter for a service rendered through a managed care plan.

Other studies that have flagged key provider equivalents tend to use Evaluation and Management (“EM”) visits as a denominator. The definition I use is intended to more flexibly include all settings that might capture a Medicaid beneficiary’s potentially “non-traditional” and/or non-regular utilization of primary care services. I also evaluate the effect of a key provider’s exit on both outpatient encounters and EM visits as representative and relevant measures of primary care, though note that EM visits represent more conventional measures of primary care utilization. For simplicity, I refer to outpatient encounters as “encounters” below.

The share of encounters associated with a key provider, as mentioned above, is related to a continuity of (relational) care index, such that larger shares indicate that the key provider was responsible for more of the patient’s encounters during the initial 6 month period, which may be analogous to a higher degree of continuity in the relationship. Table A.2 reports the share of encounters associated with each treated patient’s key provider, for the study sample and for beneficiaries with chronic conditions. The average share of encounters is approximately the same at 0.75 and 0.74 for the study sample and chronic condition sample, respectively. Less than 10% of patients have fewer than approximately 44% of their encounters with their key provider in the 6-month period, and more than half have at least 75% of their encounters with their key provider.

4.2. Provider Exits

To identify a provider’s exit from a managed care plan, I adopt a conservative approach with the goal of avoiding the mis-characterization of providers who intermittently or irregularly see Medicaid patients as exiters. Specifically, I impose the following five restrictions. First, the provider must see a rolling average of at least three patients per year-month for all time periods prior to a potential exit, and exactly no patients thereafter. Second, the potential exit date must be at least three months prior to the end of the final year for which that state’s data is available. Third, the provider must be observably participating in the plan (per Step 1) for at least five months prior to the exit. Fourth, the provider’s potential exit date cannot be the same year as a plan’s exit date. Fifth, the potential exit date must be in a year in which the plan has no recorded data quality concerns. Additional details of this approach, as well as in my construction of provider networks, are available in Appendix J.

4.3. Outcomes

I evaluate the effect of a key provider’s exit on several measures of primary care, inpatient hospitalization, and emergency care that are generally believed to be improved through non-disrupted patient-provider relationships. Table A.3 shows the frequency and distribution of the outcomes of interest in the pre-exit period, including the share of patients that have at least one record of the outcome in a given quarter, my preferred measure.

I include two measures of primary care visits: encounters, and evaluation and management (“EM”) visits. Encounters are as defined above, and are meant to flexibly capture the potentially non-traditional type of primary care that Medicaid enrollees may have. To identify EM visits, I use Berenson-Eggers Type of Service (BETOS) codes. These types of visits are commonly used as a measure of a patient’s access to primary care, and vary across intensity of visit (by time and effort contributed by the physician) as well as reimbursement level. I focus on EM visits conducted in an office for either new or established patients.

I include a measure of all-cause hospitalizations (“hospitalizations”). As noted previously, prior research has established a negative association between continuous care and hospitalizations (Hussey et al., 2014; Bayliss et al., 2015; Tingley, 2018). An alternative measure is hospitalizations that are specifically attributable to (preventable) ambulatory care sensitive conditions (ACSCs). Such hospitalizations are considered preventable with adequate preventive care, and thus represent a measure of the quality of the preventive care a patient has access to (Bindman et al., 1995; Mkanta et al., 2016). ACSCs have been used widely to evaluate the quality of care provided to patients with chronic conditions in particular (Chopra et al., 2016). Ideally, I would be able to estimate the impact of a key provider’s exit on condition-specific ACSCs for each of the four chronic conditions I include. However, these events are relatively rare in terms of overall hospitalizations—a study by Mkanta et al. (2016) found that out of 25,581 hospitalizations among Medicaid comprehensive managed care enrollees in 2009, ACSC hospitalizations attributable to diabetes (and related complications), COPD, and heart failure accounted for only 2.53%, 0.34%, and 1.00% of all primary diagnosis codes, respectively. I similarly observe very low levels of hospitalizations for ACSCs. Among the beneficiaries in my sample with chronic conditions, only 1% of these beneficiaries ever have an ACSC hospitalization in the four quarters prior to exit. Given these low frequencies, it would be difficult to interpret a meaningful effect of disruptions on this outcome, and thus I focus on all-cause hospitalizations, which are more frequent.

I define emergency department (“ED”) visits based on the place of service code on the claim. There is some evidence that continuous care with a provider leads to decreased use of the ED in some populations (Rosenblatt et al., 2000), while PCP retirement has been linked to a slight increase in ED use in Denmark (Simonsen et al., 2019). An increase in ED visits following a provider’s exit among Medicaid beneficiaries might reflect a decrease in patient wellbeing, or may reflect this particular population’s propensity to seek care (primary or emergent) in non-outpatient settings, as shown by Taubman et al. (2014), who observed an increase in ED utilization among individuals who were randomly given access to Medicaid coverage through the Oregon Health Insurance Experiment. Regardless of type of care needed (emergency or otherwise), ED use is generally more expensive to Medicaid agencies and enrollees than office-based primary care, and therefore any increase in ED utilization can be interpreted as a potentially avoidable increase in spending.

I include several additional outcomes of interest. I use the BETOS measure of testing to capture any additional diagnostic efforts that might follow a disruption to care. This testing category includes lab tests for routine venipuncture, urinalysis, glucose, and blood counts, and is also used in other studies as a measure of primary care utilization or potential redundancies due to information loss after a disruption (Finkelstein et al., 2016; Kwok, 2019). I use place of service codes to flag beneficiary use of urgent care facilities and federally quality health centers (FQHC). Utilization of urgent care facilities or FQHCs is not necessarily a flag for inadequate primary care or adverse health events. A report by the Centers for Studying Health System Change proposed that urgent care centers may represent a cost-saving alternative to more expensive ED use among Medicaid enrollees (Yee et al., 2013). Falik et al. (2001) found a decreased likelihood in hospitalizations and ED visits for ACSCs among Medicaid beneficiaries in traditional (non-managed care) Medicaid who received more than 50% of their primary care from FQHCs. Thus, any change in utilization of these places of service are somewhat ambiguous with regards to the implications on beneficiary health and access.

Finally, I use prescription fills and number of prescriptions as a measure of access to prescription drugs. In particular, in an attempt to capture longer-term drug coverage for which a decrease in access might represent a clinically concerning gap in coverage, I focus on prescriptions that are supplied for at least 30 days. These measures are intended to capture general changes in access to prescription drugs, though a more thoughtful examination of the effects of disrupted relationships on changes in patients’ access to clinically appropriate drugs is merited.

See Appendix I for additional details on how I identify and define these outcomes in the data, and Appendix G for a discussion of the prescription drug measures in particular.

5. Empirical Approach

5.1. Event Study

This study was approved by the Dartmouth Institutional Review Board (Committee for the Protection of Human Subjects Number STUDY0030069). My goal is to estimate the effect of a disruption to the patient-provider relationship on a patient’s utilization of primary care and health outcomes. To do this, I use the plausibly exogenous exit of beneficiary i’s key provider j from their network at time t′ = 0. Beneficiaries are “treated” if their key provider exits, and “control” otherwise. My underlying assumption is that if not for the key provider’s exit, beneficiary i’s pattern of utilization and outcomes would be similar in trends to beneficiaries whose key providers don’t exit at t′ = 0. This assumption would be supported by a lack of significant differences in trends in the pre-exit period. Additionally, I assume that the reason for the key provider’s exit is not due to an unobservable factor that also influences a particular beneficiary’s outcomes around the time of the exit. For example, if low-quality providers are more likely to exit, this assumption is less likely to hold. While I cannot directly test for balance among exiters and non-exiters in terms of provider quality, I show below that other observable characteristics of providers are balanced. I also include a robustness check in which I estimate the model using only treated beneficiaries, which will increase the likelihood that treated and controls (i.e. the treated beneficiaries, before their key provider exits) are more similar in the quality of their providers.

A potential threat to my identification strategy would be if a provider’s exit from a CMC does not effectively truncate the patient-provider relationship, which may be plausible if there is no financial penalty to the beneficiary for seeking out-of-network care. However, while financial risk for emergency out-of-network care is assumed by the CMC, non-emergency out-of-network care is generally not covered, as in any other (non-Medicaid) managed care arrangement. A 2009 report by the Lewin Group (The Lewin Group, 2009) noted that while Medicaid beneficiaries themselves cannot legally incur charges from providers, Medicaid health plans have little leverage to enforce this rule when care is provided outside of a pre-arranged contract, particularly among providers that are not Medicaid-certified. Additionally, plans and state Medicaid agencies are highly incentivized to discourage out-of-network care among their Medicaid enrollees since, in the absence of any federal law governing reimbursement for out-of-network non-emergency services, clinicians providing the out-of-network care tend to bill their “usual and customary” charges, which are often significantly higher than those that Medicaid fee-for-service would reimburse. Given the significant relative cost of out-of-network care to the managed care organization, state agency, and even the enrollee, it is reasonable to expect that there are sufficient financial disincentives to deter enrollees from seeking care outside of their network of providers. I observe fewer than 11 enrollees who continue to access their provider following the provider’s exit from their managed care plan.

My main specification is presented in Equation 1, where i indexes a beneficiary-episode (of enrollment), t indexes calendar time (year-quarter), and $1 {t^{'} = q}$ is an indicator for quarter from exit t′ equal to q, where I focus on the 10 quarters before and after a key provider’s exit:

y_{i t} = α_{i} + β_{t} + \sum_{q = - 10, q \neq - 1}^{10} θ_{q} 1 {t^{'} = q} + ε_{i t}

(1)

where ε_it is mean-zero conditional on covariates. In my preferred specification, I estimate this model using a definition of “Ever y_it,” which is equal to 0 when the beneficiary had no record of y in a particular quarter, and 1 if the beneficiary had at least one record of that outcome in quarter t. This variable is intended to capture the change in the share of beneficiaries that had any utilization in a particular quarter following a key provider’s exit, and is similar to the measure used by Hussey et al. (2014) to determine the relationship between increased continuity of relationships and a patient ever having a hospitalization or ED visit in a particular year. I prefer this definition as it provides a more straightforward interpretation of the effect of disruptions on the number of beneficiaries affected, and not simply level changes in utilization that may be difficult to interpret in terms of policy implications.

Time t′ = 0 is defined as the last quarter in which I observed the key provider treating patients in a given plan, and can be considered a “partial” quarter since the exact exit date is unknown. For this reason, the relevant omitted relative time period is t′ = −1, the last full quarter in which providers were observed treating patients in a plan. For treated beneficiaries, exit at time t′ = 0 is defined only outside of the 6-month key provider identification period. I estimate Equation 1 using both the treated and control cohorts in order to more precisely estimate the effect of calendar time on outcomes of interest, setting all indicators for quarters from exit equal to 0 for controls. Treated beneficiaries must be enrolled in a plan for at least t′ ∈ [−1, 1] (in addition to the 6-month key provider identification period), but I do not restrict the sample to a balanced panel.⁴ Analogously, control beneficiaries must be enrolled for at least three quarters (excluding the initial 6 month key provider identification period). I cluster standard errors at the patient level, which is where I believe any serial correlation will occur. I drop the key provider definition period from this analysis (i.e., the 6 month period in which the key provider is determined).

5.2. Pre-Post

I use a pre-post approach to recover a single estimate of the effect of key provider exit on outcomes of interest. This approach also provides increased precision in my estimates. The estimated model is similar to Equation 1, but with one indicator for the post-exit period (“Post Exit”) instead of individual indicators for each quarter relative to exit (“Post Exit” is set to 0 for all controls). Observations are at the beneficiary-episode-quarter level. I estimate this model on the 4 quarters before and after a key provider’s exit in order to obtain a single estimate of the short-term effect of the provider exit, informed by where the effect is most pronounced after estimating Equation 1. I omit t′ = 0 as a partial quarter.

y_{i t} = α_{i} + β_{t} + δ {Post Exit}_{i t} + ε_{i t}

(2)

where ε_it is mean-zero conditional on covariates.

In Equation 3, I test for heterogeneous treatment effects among beneficiaries with chronic conditions:

y_{i t} = α_{i} + β_{t} + δ_{1} [{Post Exit}_{i t} \times {{No CC}_{i}}] + δ_{2} [{{Post Exit}_{i t}} \times {{CC}_{i}}] + ε_{i t}

(3)

where {No CC_i} and {CC_i} are indicators that the beneficiary does not and does have at least one chronic condition of interest, respectively. This analysis responds to concerns that a pooled pre-post approach may obscure the heightened sensitivity that patients with chronic conditions may have to disrupted care. I estimate this model instead of estimating Equation 2 on the subsample of treated beneficiaries who have a chronic condition (including all controls) for two reasons. First, this approach allows me to compare heterogeneous treatment effects for beneficiaries who have chronic conditions against those who don’t. Second, this approach allows me to include all beneficiaries in the estimation; since in an event study, beneficiaries in the period prior to the main event (key provider exit) act as controls as well, including the full sample and allowing the non-chronic beneficiaries in their pre-exit period to act as controls is more consistent with the main (non-heterogeneous) specification. Notably, estimating Equation 2 on the subsample of treated chronic condition beneficiaries (and full controls) returns the same results as the specification in Equation 3.

Given prior evidence that increasing continuity in relationships is associated with improved patient outcomes, I also estimate heterogeneous treatment effects across quartiles of the share of encounters a key provider is responsible for in the 6-month definition period to determine if there is any relationship between increasing continuity and outcomes of interest in my sample. To do this, I estimate Equations 2 and 3 separately on each subset of beneficiaries whose key provider’s share of encounters falls into the relevant quartile.

6. Results

6.1. Descriptive Statistics

There are nearly 2 million unique adult beneficiaries enrolled in CMC in the six sample states from 2009 to 2014, representing nearly 3 million beneficiary-enrollment episodes (the “full sample”) (Table 1). My study sample includes 103,969 beneficiaries (105,548 beneficiary-episodes), of which 101,487 beneficiary-episodes do not include a key provider’s exit (“control”) and 4,061 do (“treated”). The majority of beneficiaries are female and white, with an average age at enrollment between 32 to 35. Approximately 12% of beneficiaries in the full sample have one of the four chronic conditions of interest. This percent is slightly over twice as large (25%) in the study sample, with 30% of the treated cohort having a chronic condition, a difference that is significant at the 0.01% level, relative to the control cohort. Notably, the prevalence of chronic conditions (overall and by specific condition) in the study sample is more in line with recent estimates of the burden of chronic disease in this population, as discussed in Section 3. While beneficiary fixed effects will control for any unobservable, time-invariant differences between treated and control cohorts that might be reflected in the slightly higher prevalence of chronic conditions in the former, I address any limitations to the external validity of my results that might be suggested by the different rates of chronic disease between the full and study samples in Section 7.

Table 1:

Description of Beneficiary Characteristics

Beneficiary Characteristic	Full Sample	(SD)	Study Sample	(SD)	Treated	(SD)	Control	(SD)	T-stat
Share Male	.27		.21		.24		.21		4.55
Share White	.56		.56		.56		.56		−.62
Mean Age at Enrollment	32	(10)	34	(10)	35	(11)	34	(10)	5.21
Share Chronic Condition	.12		.25		.3		.25		6.89
Share CHF	.01		.01		.02		.01		2.47
Share COPD	.02		.04		.05		.03		5.02
Share Diabetes	.04		.1		.12		.09		4.77
Share Hypertension	.08		.18		.22		.18		5.49
Mean Months per Episode	16	(15)	30	(12)	37	(13)	30	(12)	41.93
Median Num. of Episodes	2	(1.2)	1	(.78)	1	(.68)	1	(.78)	−6.59
Share Metropolitan			.89		.87		.89		−4.84
Share Micropolitan			.07		.08		.07		2.02
Share Small/Rural			.04		.06		.04		5.1
Share Male KP			.68		.67		.68		−.59
Share OP Encounter			.76	(.22)	.75	(.22)	.76	(.22)	−1.83
Share MD/DO KP			.93		.92		.93		−1.8
Share PCP			.89		.9		.89		1.65
Share NP/PA KP			.07		.08		.07		1.8
Share PCP			.9		.91		.9		.59
Mean KP Age at Enrollment			49	(9.5)	48	(10)	49	(9.4)	−5.23
Total Beneficiary-Episodes	2928685		105548		4061		101487
Total Beneficiaries	1837044		103969		4061		99962

Key Provider Characteristic	Full Sample	(SD)	Study Sample	(SD)	Exiters	(SD)	Non-Exiters	(SD)	T-stat
Share Male	.6		.65		.66		.65		.12
Share MD/DO	.68		.92		.91		.92		−2.4
Share PCP	.39		.87		.88		.87		2.71
Share NP/PA	.14		.08		.09		.08		2.4
Share PCP	.31		.9		.91		.9		.22
Mean Age at Enrollment	44	(10)	48	(9.6)	48	(10)	48	(9.5)	−2.86
Share OP Encounter			.7	(.21)	.7	(.21)	.7	(.21)	.16
Total	226504		15376		875		15107

	(1) Ever EM Visit	(2) Ever Encounter	(3) Ever ED Visit	(4) Ever Hospitalization
Panel A. Study Sample
Post Exit	−0.037^*** (0.007)	−0.038^*** (0.007)	−0.004 (0.006)	0.004 (0.003)

Observations	702113	702113	702113	702113
Adjusted R²	0.279	0.291	0.228	0.066
Pre-Exit Mean	.679	.683	.212	.036
Panel B. Chronic Condition Subsample
No Chronic Condition	−0.034^*** (0.009)	−0.041^*** (0.009)	−0.008 (0.007)	−0.004 (0.003)
Chronic Condition	−0.042^*** (0.011)	−0.032^** (0.012)	0.005 (0.010)	0.021^*** (0.005)

Observations	702113	702113	702113	702113
Adjusted R²	0.279	0.291	0.228	0.066
Pre-Exit Mean, No Chronic	.63	.636	.202	.032
Pre-Exit Mean, Chronic	.794	.793	.235	.045

	(1) Ever EM Visit	(2) Ever Encounter	(3) Ever ED Visit	(4) Ever Hospitalization
Panel A. Study Sample
Post Exit	−0.030^*** (0.007)	−0.031^*** (0.007)	−0.003 (0.006)	0.002 (0.003)

Observations	645338	645338	645338	645338
Adjusted R²	0.270	0.283	0.222	0.069
Pre-Exit Mean	.657	.659	.194	.03
Panel B. Chronic Condition Subsample
No Chronic Condition	−0.024^** (0.009)	−0.028^** (0.009)	−0.001 (0.007)	−0.002 (0.003)
Chronic Condition	−0.045^*** (0.012)	−0.039^** (0.013)	−0.010 (0.010)	0.011^* (0.005)

Observations	645338	645338	645338	645338
Adjusted R²	0.270	0.283	0.222	0.069
Pre-Exit Mean, No Chronic	.612	.616	.183	.027
Pre-Exit Mean, Chronic	.767	.768	.223	.038

	Pregnancy Flag	Percent	Total
Treated	93	2.29%	4,061
Control	4,346	4.35%	99,962
Total	4,439	4.27%	103,969

Population	Mean	SD	10th Pctile	Median	75th Pctile	90th Pctile	Beneficiary-Quarters
Pooled	.75	.22	.44	.75	1	1	4061
Chronic Conditions	.74	.22	.43	.75	1	1	1209

Outcome	Mean	SD	Ever (Mean)	SD	Median	75th Pctile	90th Pctile	Beneficiary-Quarters
EM Visit	1.88	2.1	.68	.47	1	3	5	11446
Encounter	2.21	2.7	.68	.47	1	3	6	11446
ED Visit	.35	.88	.21	.41	0	0	1	11446
Hospitalization	.04	.22	.036	.19	0	0	0	11446
Tests	.97	1.5	.48	.5	0	1	3	11446
FQHC	.004	.1	.003	.05	0	0	0	11446
Urgent Care	.006	.09	.005	.07	0	0	0	11446
Ever Fill Date	.62	.48	.62	.48	1	1	1	9257
Num. Drugs	2.46	3.2	.62	.48	1	4	7	9257
Num. Fill Dates	4.3	6.2	.62	.48	2	6	13	9257
Num. Fill Dates per Drug	1.04	.97	.62	.48	1	1.75	2.33	9257

Measure	Mean	SD	25th Pctile	50th Pctile	75th Pctile	90th Pctile	Num. Benes
Enrollment Length in CMCO	30	12	21	27	36	48	105548
Enrollment to First Encounter	3	4	0	1	3	6	105548
Enrollment to First Visit with KP	3	4	1	2	4	7	105548
Pre-Exit Period	15	9	8	12	20	28	4061
Enrollment to KP Exit	23	10	16	20	29	36	4061
KP Exit to End of Enrollment	14	10	7	12	20	27	4061
First Visit with KP to Exit	20	9	13	18	25	33	4061
Last Visit with KP to Exit	9	8	2	6	12	20	4061
First Visit with KP to Last Visit with KP	14	12	5	11	20	29	105548

Measure	Mean	SD	25th %tile	Median	75th %tile	N
Days between pre-exit encounters	28	46	6	14	30	79536
Days between last visit with KP and next encounter	105	149	15	46	131	3819
Days between KP exit and next encounter	73	103	12	34	89	3439
Days between last visit with KP and exit	263	252	74	191	363	4061
Days between last encounter with KP and exit	270	252	80	202	376	4061
Months between last visit with KP and exit	8	8	2	6	11	4061
Months between last encounter with KP and exit	8	8	2	6	12	4061

Measure	Count	Total Beneficiaries	Share
Any visit with KP within 12 months of exit	3157	4061	0.77
Encounter with KP within 12 months of exit	3119	4061	0.77
No encounters post exit	706	4061	0.17

Outcome	F Stat	P value
EM Visit	1.726	.077
Encounter	1.505	.139
ED Visit	1.228	.272
Hospitalization	.942	.487

	EM Visit				Encounter
Share	Mean	Coef.	SE	Pct. Change	Mean	Coef.	SE	Pct. Change	Bene. (N)	Share Range
1	.74	−.02	(.01)	−2.7%	.75	−.03	(.01)	−3.4%	994	(.12, .57)
2	.66	−.01	(.01)	−.9%	.67	−.01	(.01)	−.9%	1149	(.58, .75)
3	.7	−.05^**	(.02)	−7.6%	.69	−.06^**	(.02)	−8.4%	546	(.76, .98)
4	.64	−.07^***	(.01)	−10%	.64	−.07^***	(.01)	−10%	1372	(1, 1)

	Hospitalization				ED Visit
Share	Mean	Coef.	SE	Pct. Change	Mean	Coef.	SE	Pct. Change	Bene. (N)	Share Range
1	.05	−.01	(.01)	−10%	.27	−.03^*	(.01)	−11%	994	(.12, .57)
2	.03	0	(.01)	13%	.2	.01	(.01)	5.1%	1149	(.58, .75)
3	.03	.01	(.01)	18%	.21	.02	(.01)	7.8%	546	(.76, .98)
4	.03	.01^*	(0)	34%	.18	−.01	(.01)	−2.9%	1372	(1, 1)

	EM Visit				Encounter
Share	Mean	Coef.	SE	Pct. Change	Mean	Coef.	SE	Pct. Change	Bene. (N)	Share Range
1	.84	−.01	(.02)	−1.5%	.84	.01	(.02)	.9%	303	(.12, .56)
2	.81	−.03	(.02)	−4.1%	.81	−.02	(.02)	−2.8%	248	(.56, .73)
3	.78	−.03	(.02)	−3.5%	.77	0	(.02)	−.4%	284	(.73, .97)
4	.75	−.08^***	(.02)	−11%	.76	−.09^***	(.02)	−12%	374	(.98, 1)

	Hospitalization				ED Visit
Share	Mean	Coef.	SE	Pct. Change	Mean	Coef.	SE	Pct. Change	Bene. (N)	Share Range
1	.06	.02	(.01)	26%	.31	−.01	(.02)	−3%	303	(.12, .56)
2	.05	.02	(.01)	35%	.24	0	(.02)	1.7%	248	(.56, .73)
3	.04	.02	(.01)	40%	.19	.04	(.02)	18%	284	(.73, .97)
4	.03	.03^***	(.01)	110%	.2	−.01	(.02)	−3.1%	374	(.98, 1)

Outcome	F Stat	P value
EM Visit	.724	.687
Encounter	.554	.835
ED Visit	1.995	.036
Hospitalization	1.862	.053

Step	N	Pct. of Total
All beneficiaries enrolled in a CMCO between 2009-2014	1,837,044	100%
All beneficiaries with a start date after Jan. 2009	1,444,216	79%
Beneficiaries with one plan per enrollment episode	1,385,430	75%
Beneficiaries with an identifiable key provider, 6-month identification period	270,498	15%
Treated
Beneficiaries with a key provider who exits (outside their KP identification period)	13,577	1%
Beneficiaries who remain enrolled for at least one month following the exit	12,762	1%
Beneficiary-episodes where key provider exited, excluding those who exited in April or May of 2012 in AZ	6,209	0.34%
Beneficiary-episodes that meet the three-quarter enrollment requirements around the exit	4,061	0.22%
Control
Beneficiaries with a key provider who doesn’t exit	265,874	14%
Beneficiaries who remain enrolled for at least one month following the KP identification period	234,146	13%
Beneficiary-episodes that meet the three-quarter enrollment requirements following the key provider identification period	101,487	6%

	(1) Ever EM Visit	(2) Ever Encounter	(3) Ever ED Visit	(4) Ever Hospitalization
Post Exit	−0.033^*** (0.007)	−0.035^*** (0.007)	−0.005 (0.005)	0.004 (0.003)

Observations	706491	706491	706491	706491
Adjusted R²	0.278	0.291	0.227	0.066
Pre-Exit Mean	.679	.686	.208	.036

	(1) Ever EM Visit	(2) Ever Encounter	(3) Ever ED Visit	(4) Ever Hospitalization
Panel A. Study Sample
Post Exit	−0.039^*** (0.010)	−0.052^*** (0.010)	−0.015 (0.008)	0.000 (0.004)

Observations	690894	690894	690894	690894
Adjusted R²	0.278	0.291	0.227	0.066
Pre-Exit Mean	.648	.652	.229	.036
Panel B. Chronic Condition Subsample
No Chronic Condition	−0.035^** (0.011)	−0.056^*** (0.012)	−0.017 (0.009)	−0.005 (0.004)
Chronic Condition	−0.053^** (0.018)	−0.041^* (0.019)	−0.008 (0.017)	0.014 (0.009)

Observations	690894	690894	690894	690894
Adjusted R²	0.278	0.291	0.227	0.066
Pre-Exit Mean, No Chronic	.604	.612	.219	.032
Pre-Exit Mean, Chronic	.778	.769	.259	.047

	(1) Ever EM Visit	(2) Ever Encounter	(3) Ever ED Visit	(4) Ever Hospitalization
Panel A. Study Sample
Post Exit	−0.037^*** (0.008)	−0.038^*** (0.010)	−0.004 (0.006)	0.004 (0.003)

Observations	702113	702113	702113	702113
Adjusted R²	0.279	0.291	0.228	0.066
Pre-Exit Mean	.679	.683	.212	.036
Panel B. Chronic Condition Subsample
No Chronic Condition	−0.034^*** (0.009)	−0.041^*** (0.010)	−0.008 (0.007)	−0.004 (0.004)
Chronic Condition	−0.042^** (0.015)	−0.032 (0.017)	0.005 (0.010)	0.021^*** (0.006)

Observations	702113	702113	702113	702113
Adjusted R²	0.279	0.291	0.228	0.066
Pre-Exit Mean, No Chronic	.63	.636	.202	.032
Pre-Exit Mean, Chronic	.794	.793	.235	.045

	(1) Ever EM Visit	(2) Ever Encounter	(3) Ever ED Visit	(4) Ever Hospitalization
Panel A. Study Sample
Post Exit	−0.053^*** (0.014)	−0.056^*** (0.014)	0.002 (0.012)	0.003 (0.006)

Observations	23275	23275	23275	23275
Adjusted R²	0.297	0.306	0.237	0.079
Pre-Exit Mean	.679	.683	.212	.036
Panel B. Chronic Condition Subsample
No Chronic Condition	−0.049^*** (0.014)	−0.059^*** (0.014)	−0.002 (0.012)	−0.004 (0.006)
Chronic Condition	−0.060^*** (0.017)	−0.050^** (0.017)	0.011 (0.015)	0.021^** (0.008)

Observations	23275	23275	23275	23275
Adjusted R²	0.297	0.306	0.237	0.080
Pre-Exit Mean, No Chronic	.63	.636	.202	.032
Pre-Exit Mean, Chronic	.794	.793	.235	.045

	(1) Ever EM Visit	(2) Ever Encounter	(3) Ever ED Visit	(4) Ever Hospitalization
Panel A. Study Sample
Post Exit	−0.050^*** (0.010)	−0.051^*** (0.011)	−0.001 (0.005)	0.002 (0.003)
Exit	0.024^** (0.008)	0.015 (0.009)	0.013^* (0.006)	0.000 (0.002)

Observations	606551	606551	606551	606551
Adjusted R²	0.043	0.046	0.033	0.012
Pre-Exit Mean	.693	.697	.212	.04
Panel B. Chronic Condition Subsample
No Chronic Condition	−0.054^*** (0.011)	−0.056^*** (0.013)	−0.007 (0.005)	−0.004 (0.003)
Chronic Condition	−0.043^*** (0.012)	−0.038^** (0.013)	0.011 (0.010)	0.014^** (0.004)
Exit	0.024^** (0.008)	0.015 (0.009)	0.013^* (0.006)	0.000 (0.002)

Observations	606551	606551	606551	606551
Adjusted R²	0.043	0.046	0.033	0.012
Pre-Exit Mean, No Chronic	.642	.649	.198	.035
Pre-Exit Mean, Chronic	.803	.801	.241	.05

State	Program	Adult Populations Enrolled	Mandatory?	Start Date	End Date
AZ	Arizona health Care Cost Containment System	Low-income adults	Mandatory	Oct. 1982

IN	Hoosier Healthwise	Poverty-level pregnant women, section 1931 adults only	Mandatory	Jan. 2008
	Hoosier Healthwise	Presumptively eligible pregnant women	Mandatory	Mar. 2011
	Healthy Indiana Plan	Low-income adults	Mandatory	Jan. 2008

KY	KY Health Care Partnership Program	Low-Income Adults	Mandatory	Nov. 1997
	Medicaid Managed Care Program	Low-income adults	Mandatory	2011

NJ	NJ Family Care	Low-income adults	Mandatory	<2010

	Salud!	Low-income adults	Mandatory	Jul. 1997	Dec. 2013
NM	State Coverage Initiative (Section 1115 Demonstration)	Childless low-income adults	Voluntary	Jan. 2010	Dec. 2013
	Centennial Care	Low-income adults	Mandatory	Jan. 2014

	Healthy Options	Low-income adults	Mandatory	Jul. 1994
WA	Washington Basic Health (Transitional Bridge Demonstration)	Low-income adults (<133% FPL and not Medicaid-eligible)	Mandatory	Jan. 2011	Dec. 2013
	Medical Care Services (Transitional Bridge Demonstration)	Low-income adults (<133% FPL and not Medicaid-eligible)	Mandatory	Jan. 2011	Dec. 2013

	Treated			Control
Beneficiary Characteristic	6-Month	12-Month	T-stat	6-Month	12-Month	T-stat
Share Male	.24	.24	−.3	.21	.22	3.32
Share Missing Sex	0	0		0	0	.05
Share White	.56	.54	−1.74	.56	.54	−9.61
Share Missing Race	.03	.03	.29	.04	.04	−.34
Mean Age at Enrollment	34.85	34.64	−.87	33.99	34.35	7.7
Share Chronic Condition	.3	.29	−1.13	.25	.26	6.37
Share CHF	.02	.02	−.53	.01	.01	1.42
Share COPD	.05	.04	−.99	.03	.04	1.82
Share Diabetes	.12	.11	−.62	.09	.1	3.21
Share Hypertension	.22	.21	−.54	.18	.19	6.33
Mean Months per Episode	37.44	42.33	17.43	29.52	35.2	110.01
Median Num. of Episodes	1.52	1.47	−3.44	1.61	1.56	−15.03
Share Metropolitan	.87	.86	−.89	.89	.9	6.68
Share Micropolitan	.08	.09	1.86	.07	.06	−4.75
Share Small/Rural	.06	.05	−.91	.04	.04	−4.5
Share Male KP	.67	.71	3.44	.68	.68	.65
Share OP Encounter	.75	.69	−12.09	.76	.69	−65.72
Share MD/DO KP	.92	.93	.82	.93	.93	3.24
Share PCP	.9	.91	1.29	.89	.9	2.46
Share NP/PA KP	.08	.07	−.82	.07	.07	−3.24
Share PCP	.91	.94	1.53	.9	.91	2.54
Mean KP Age at Enrollment	47.87	48.25	1.39	48.77	48.7	−1.63
Total Beneficiary-Episodes	4061	3709		101487	95102

	Exiters			Non-Exiters
KP Characteristic	6-Month	12-Month	T-stat	6-Month	12-Month	T-stat
Share Male KP	.66	.69	3.08	.65	.66	1.61
Share MD/DO KP	.91	.92	1.23	.92	.93	3.84
Share PCP	.88	.9	1.82	.87	.88	6.89
Share NP/PA KP	.09	.08	−1.23	.08	.07	−3.84
Share PCP	.91	.94	1.18	.9	.91	1.62
Mean KP Age at Enrollment	47.66	48	1.1	48.24	48.32	1.53
Share OP Encounter	.7	.65	−8.61	.7	.65	−43.63
Total	875	734		15107	14279

PERMALINK

Disruptions to the Patient-Provider Relationship and Patient Utilization and Outcomes: Evidence from Medicaid Managed Care

Becky Staiger

Abstract

1. Introduction

2. Medicaid Managed Care

3. Data Sources and Sample

3.1. Medicaid MAX Data

3.2. Patient Sample

4. Measures

4.1. Key Provider

4.2. Provider Exits

4.3. Outcomes

5. Empirical Approach

5.1. Event Study

5.2. Pre-Post

6. Results

6.1. Descriptive Statistics

Table 1:

Table 2:

Figure 1:

6.2. Estimates of the Effect of Exits for the Main Study Sample

Figure 2:

Table 3:

6.3. Estimates of the Effect of Exits for the Chronic Condition subsample

Figure 3:

6.4. Estimates of the Effect of Exits by Increasing Continuity of Care

Figure 4:

6.5. Estimates of the Effect of Exits for the 12-Month Key Provider Cohort

Table 4:

6.6. Additional Robustness Checks

7. Discussion

Acknowledgments

Funding:

Appendix A. Additional Tables and Figures

A.1. Tables

Table A.1:

Table A.2:

Table A.3:

Table A.4:

Table A.5:

Table A.6:

Table A.7:

Table A.8:

Table A.9:

Table A.10:

Table A.11:

Table A.12:

Table A.13:

A.2. Figures

Figure A.1:

Figure A.2:

Figure A.3:

Figure A.4:

Appendix B. Robustness Checks

Table B.14:

Table B.15:

Table B.16:

Table B.17:

Table B.18:

Figure B.5:

Appendix C. Mandatory Enrollment into Managed Care

Table C.19:

Appendix D. 12-Month Key Provider Cohort

Table D.20:

Table D.21:

Table D.22:

Table D.23:

Appendix E. Balanced Panel Robustness Check

Table E.24:

Figure E.6:

Appendix F. Exploring Bias in Excluding Beneficiary Exiters

Figure F.7:

Figure F.8:

Table F.25:

Table F.26:

Appendix G. Effect on Prescription Drug Fills

Table G.27:

Figure G.9:

Figure G.10:

	12-Month Cohort
6-Month Cohort	No	Yes	Total
No	0	1500	1500
Yes	1852	2209	4061
Total	1852	3709	5561

	12-Month Cohort
6-Month Cohort	No	Yes	Total
No	0	32933	32933
Yes	39318	62169	101487
Total	39318	95102	134420

Cohort	N	Pct. Total	Pct. Sub-Total
Treated	4551	100
Continuous	4061	89.2
Dropout	490	10.8	100
Churn	160	3.5	32.7
Full	330	7.3	67.3
Control	128511	100
Continuous	101487	79
Dropout	27024	21	100
Churn	6494	5.1	24
Full	20530	16	76

State	2009	2010	2011	Notes
AZ	Y	Y	NR	NR=file not available at time of evaluation
IN	Y	0	0	0=IN submitted 200 or fewer RX encounter records
KY	Y	Y	Y
NJ		Y	Y	NJ did not include prescription drugs in their HMO benefit package during 2009
NM	Y	Y	Y
WA	Y	Y	Y

	Years of Available Data
State	2009	2010	2011	2012	2013	2014
AZ	X	X	X	X	X
IN	X	X	X	X	X
KY	X	X	X	X
NJ	X	X	X	X	X	X
NM	X	X	X	X
WA	X	X	X	X	X

Year	State	MCCOMBO_STRG	Enrollment Episode
2009	KY	000000161616161616161616	1
2010	KY	161616000000000000000000	1
2010	IN	000000000101010101010101	2
2010	IN	010101010101010101010101	2

Year	State	MCCOMBO_STRG	Eligibility Code	Enrollment Episode
2009	KY	15 (adult)	080808080808080808080808	?
2009	IN	34 (child, poverty)	000000000000000007070707	?

Year	State	MCCOMBO_STRG	Zip Code
2009	CA	161616000000000000000000	[CA zip]
2009	AZ	000000161616161616161616	[CA zip]

Stage	Medicaid	MCO	CMCO
Baseline (unrestricted) count	2,720,964	2,720,964	2,720,964

Beneficiary Level Restriction

Drop ever dual	2,648,679	2,648,679	2,648,679
Drop ever multiple records	2,532,322	2,532,322	2,532,322

Beneficiary-Episode Level Restriction

Drop if episode is associated with non-adult eligibility	2,409,929	2,419,335	2,432,399
Drop if episode is “out-of-state”	2,399,455	2,410,390	2,423,226
Drop if state associated with episode is out-of-sample	1,811,401	1,821,664	1,837,044

Final (cleaned) beneficiaries	1,811,401	1,821,664	1,837,044

Year	IP Hosp	Pct.	IP Hosp (40)	Pct.	OT Hosp	Pct.	ED	Pct.
2009	146153	12	18546	2	−99	.	391369	32
2010	132513	11	17846	1	−99	.	409574	33
2011	146504	11	21438	2	−99	.	437591	34
2012	147071	12	19588	2	−99	.	417156	33
2013	100414	11	15345	2	−99	.	315101	34
2014	24154	8	7980	3	−99	.	121064	40