The effects of Medicaid on access to care and adherence to recommended preventive services

Abstract

Objective

To quantify the impact of Medicaid enrollment on access to care and adherence to recommended preventive services.

Data Source

2005‐2015 Medical Expenditure Panel Survey Household Component.

Study Design

We examined several access measures and utilization of several preventive services within the past year and within the time frame recommended by the United States Preventive Services Task Force, if more than a year. We estimated local average treatment effects of Medicaid enrollment using a new, two‐stage regression model developed by Nguimkeu, Denteh, and Tchernis. This model accounts for both endogenous and underreported Medicaid enrollment by using a partial observability bivariate probit regression as the first stage. We identify the model with an exogenous measure of Medicaid eligibility, the simulated Medicaid eligibility rate by state, year, and parents vs childless adults. A wide range of changes in Medicaid eligibility occurred during the time period studied.

Data Collection/Extraction methods

Sample of low‐income, nonelderly adults not receiving disability benefits.

Principal Findings

Medicaid enrollment decreased the probability of having unmet needs for medical care by 7.5 percentage points and the probability of experiencing delays getting prescription drugs by 7.7 percentage points. Medicaid enrollment increased the probability of having a usual source of care by 16.5 percentage points, the probability of having a routine checkup by 17.1 percentage points, and the probability of having a flu shot in past year by 12.6 percentage points.

Conclusion

Medicaid enrollment increased access to care and use of some preventive services. Additional research is needed on impacts for subgroups, such as parents, childless adults, and the smaller and generally older populations for whom screening tests are recommended.

What is already known

• Expanded Medicaid eligibility increases access to care and use of some preventive services among the eligible population. Relatively little is known about the causal impact on the people who actually enroll in Medicaid.

• Medicaid enrollment tends to be underreported.

What this study adds

• Among nonelderly, nondisabled adults, after accounting for underreporting and unobserved factors related to Medicaid enrollment, Medicaid enrollment promotes improved access to care and increased use of some preventive services.

1. INTRODUCTION

Medicaid is a joint state and federal program providing health care to low‐income populations. Within statutory and regulatory constraints set at the federal level, states have discretion to set Medicaid eligibility criteria to meet objectives such as covering people who would otherwise lack insurance, supporting health care providers, and constraining fiscal outlays. States have often sought greater latitude over their programs. Policy makers are considering new directions for Medicaid, and policy makers may benefit from more information about the effects of the Medicaid program on its beneficiaries.

This paper presents causal estimates of the effects of Medicaid on adult beneficiaries’ access to care and preventive services. We study receipt of services recommended by the United States Preventive Services Task Force (USPSTF) within the time frames recommended by the Task Force for the appropriate populations. We estimate local average treatment effects of Medicaid enrollment using the Medical Expenditure Panel Survey Household Component (MEPS‐HC) and a new, two‐stage regression model developed by Nguimkeu, Denteh, and Tchernis. ¹ This regression model corrects for two difficulties in estimating causal effects with observational data. First, unobserved factors correlated with both Medicaid enrollment and outcomes (endogeneity) may bias conventional estimates. Second, survey respondents underreport their Medicaid coverage, ² , ³ which can also bias estimates. This regression model corrects for endogenous and underreported Medicaid enrollment by using a partial observability bivariate probit regression as the first stage. We identify the model with an exogenous measure of Medicaid eligibility, simulated Medicaid eligibility rates derived from a detailed simulation model. We use the wide range of changes in Medicaid eligibility within states from 2005 to 2015 to help broaden the understanding of the effects of Medicaid.

This article is one of the few to study the effects of Medicaid enrollment, in contrast to the many studies of the effects of Medicaid eligibility. Studies using Oregon Health Insurance Experiment (OHIE) data and studies of the Medicaid expansion under the Affordable Care Act found Medicaid eligibility increased access and often increased preventive care. ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ Studies of the effects of Medicaid enrollment have generally found that Medicaid enrollment increased access and preventive care use across range of strategies to account for endogeneity and the Medicaid undercount. ⁴ , ⁵ , ⁶ , ¹⁴ The OHIE had administrative data to accurately measure Medicaid enrollment and used two‐stage least squares (2SLS) with random assignment to instrument for endogenous enrollment. Another study used survey data to examine changes in outcomes after reported gains in Medicaid coverage, ¹⁴ but those reports may not reflect true changes, and unobserved factors may be correlated with changes in both enrollment and outcomes.

We enhance understanding of how Medicaid affects preventive care use by assessing impacts both in one year and adherence over the USPSTF‐recommended time frames. Studies of Medicaid enrollment and eligibility have investigated the impact of Medicaid on preventive services utilization within the past year, but USPSTF‐recommended time frames for mammogram, colorectal cancer screening, and blood cholesterol screening services are more than a year. Due to individuals gaining and losing Medicaid over time, ¹⁵ some people had access to preventive services in the past, so impacts on use during recommended time frames may be attenuated.

While other studies have focused on the ACA’s eligibility expansion or one state, we bring in a wider range of changes in Medicaid in all states from 2005 to 2015, which can help broaden the understanding of the effects of Medicaid on preventive care use. Medicaid eligibility criteria vary across states and over time, nearly all states had changes in eligibility during this time period, and patterns of changes were diverse. Twenty‐nine states adopted the ACA Medicaid eligibility expansion in 2014 or 2015. ¹⁶ Several states had rising eligibility rates from 2005 to 2013 for parents, childless adults, or both, partly because starting in 2010, states had more latitude to expand eligibility to most residents with incomes less than 138% of the FPL. Before 2014, other states had periods of greater eligibility followed by retrenchment. ¹⁷ Also, states that did not expand eligibility under the ACA nonetheless increased eligibility in 2014, because under the ACA, Medicaid eligibility rules in 2010 were converted to their MAGI equivalent, and then, a standard disregard (5% of the FPL) was added to the pre‐ACA income eligibility thresholds.

Lastly, our sample excludes adults with severe disabilities, who are not subject to eligibility expansions and might bias estimates. Our data source is repeated annual cross sections from the Medical Expenditure Panel Survey Household Component (MEPS‐HC), which identifies people with disability‐related income (Supplemental Security Income). In contrast, some of the prior studies relied on a survey that did not identify this population. ⁹ , ¹² , ¹³ When a study includes a population not affected by the intervention, then the local average treatment effect is downwardly biased. In our study, limiting the sample in this way is unlikely to change the composition of the analytic sample over time, because a careful study suggests that the Medicaid expansion had no impact on receipt of disability benefits. ¹⁸

2. METHODS

First, similar to the studies of Medicaid expansion, ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ we regress outcomes on a measure of Medicaid eligibility. Those studies used a binary indicator for Medicaid expansion as the measure of eligibility, interacted with indicators for the pre‐ and postexpansion period. Some studies of wider ranges of eligibility changes have used the proportion of the population eligible for Medicaid as the measure of eligibility. ¹⁹ , ²⁰ , ²¹ , ²² More specifically, using methods similar to Currie and Gruber, ²³ we simulate eligibility using a national sample for each state's rules in each year.

$$Outcome=X^{\prime }\beta +MedicaidEligibilityRat{e}_{\mathit{stj}}\times \gamma +{\pi }_{\mathit{sj}}+{\phi }_{\mathit{tj}}+\epsilon$$ (1)

where Outcome is a measure of access or preventive services use; Medicaid Eligibility Rate is the simulated eligibility rate (described in Data section below) that varies within states, s, across years, t, and by basis of potential eligibility (parents or childless adults), j. The coefficient γ measures the percentage point change in the outcome for each percentage point change in the eligibility rate. The vector X includes an indicator for basis of eligibility, which, along with fixed effects for state × eligibility group, π, and year × eligibility group, φ, ensures that identification is from variation within state and basis of eligibility group over time. Specifically, year fixed effects account for factors such as potential confounding due to reduced cost sharing for preventive services among the privately insured, because the ACA included preventive services mandates for nongrandfathered, non‐self‐insured private plans, which went into effect in 2011 and were contemporaneous with some of the early Medicaid eligibility expansions. ²⁴

For the preventive services with recommended time frames greater than a year, eligibility for Medicaid over the entire time may be important. For these variables, we conduct an additional analysis where we replace Medicaid Eligibility Rate with the sum of Medicaid Eligibility Rates over the two‐ or five‐year time frame in Equation (1). Thus, the regressor is cumulative Medicaid eligibility, and its coefficient measures the effect of an increase in the cumulative eligibility on the outcome. This approach is similar to that of studies of the long‐term effect of Medicaid eligibility. ¹⁹ , ²¹ , ²²

Second, to obtain estimates of the effect of Medicaid enrollment, we use Nguimkeu et al’s ¹ new, two‐stage, instrumental variable estimation approach that accounts for (a) endogeneity bias due to unmeasured factors that affect Medicaid enrollment and outcomes and (b) underreported Medicaid enrollment. This method yields local average treatment effects for those who enrolled because of changes in eligibility rates. The wide range of eligibility changes used in this analysis enhances generalizability, but the estimated effects may differ from those for other marginal enrollees or enrollees if coverage were automatic for everyone who was eligible.

True Medicaid enrollment is affected by the Medicaid eligibility rate and other factors.

$$TrueMedicaid=1\left(X^{\prime }\tau +MedicaidEligibilityRat{e}_{\mathit{stj}}\times \delta +{\theta }_{\mathit{sj}}+{\mu }_{\mathit{tj}}+v>0\right)$$ (2)

where 1(.) is a binary indicator function.

However, we only observe Medicaid enrollment reported by survey respondents, which reflect underreporting. Let d be an indicator for accurate reporting, which can vary with demographics X, by state, κ_s, over time, ψ_t, and by the duration of time the respondent must consider in reporting Medicaid enrollment, Recall. Then,

$$d=1(X^{\prime }\omega +Recall\times \lambda +{\kappa }_s+{\psi }_t+u>0)$$ (3)

The state and year fixed effects, κ_s and ψ_t, reflect the literature finding underreporting varies across states and over time, but it does not model underreporting that varies over time within a state. ²⁵ The literature also suggests respondents may report more accurately when the recall period is shorter. ² The vector X, which includes eligibility group, is the same in Equations (2) and (3). In contrast with Equation (2), Equation (3) omits Medicaid Eligibility Rate_stj and separate state and year fixed effects by eligibility group. Monte Carlo simulations by Nguimkeu et al ¹ of the sensitivity of their two‐stage model to misspecification suggest that the estimates are consistent even if variables are omitted from the reporting accuracy equation and even if Recall is correlated with the outcomes. Equation (2) does not contain Recall, which we assume does not affect true Medicaid enrollment.

Medicaid enrollment is observed only if the person had Medicaid and accurately reported it:

$$Medicaid=1\left(\begin{array}{c}{X}^{\prime }\tau +MedicaidEligibilityRat{e}_{\mathit{stj}}\times \delta +{\theta }_{\mathit{sj}}+{\mu }_{\mathit{tj}}+v>0\text{and}\\ X^{\prime }\omega +Recall\times \lambda +{\kappa }_s+{\psi }_t+u>0\end{array}\right)$$ (4)

Equation (4) reflects the joint determination of one outcome, reported Medicaid enrollment, arising from true enrollment, Equation (2), censored by reporting accuracy, Equation (3). Equation (4) is the first stage of the two‐stage model. We assume ν and u are jointly normally distributed, so that Equation (4) can be estimated using Poirier's partial observability bivariate probit model. ²⁶ Equation (4) is estimable because some variables are in Equation (2) but not (3), and vice versa.

To uncover the effect of Medicaid enrollment on outcomes, we use the estimated parameters $\widehat{\tau }$, $\widehat{\delta }$, ${\widehat{\theta }}_{\mathit{sj}}$ and ${\widehat{\mu }}_{\mathit{tj}}$ to predict $\widehat{TrueMedicaid}$. We estimate the effect of Medicaid participation on outcomes as:

$$Outcome=X^{\prime }\beta +\widehat{TrueMedicaid}\times \alpha +{\zeta }_{\mathit{sj}}+{\xi }_{\mathit{tj}}+\epsilon$$ (5)

The coefficient α is the local average treatment effect. As with Equation (1), the fixed effects ensure that identification is from variation within state and basis of eligibility group over time. This two‐stage method is identified if the Medicaid eligibility rate is a strong instrument that does not directly affect the outcomes. In contrast, the model would be identified if Recall did affect outcomes directly, but we argue in Data section that our measure would not be expected to affect outcomes.

All estimates are weighted to a national average of the years 2005‐2015. All standard errors and statistical tests for descriptive statistics account for the complex survey design of the MEPS‐HC. All regression‐based estimates account for state‐level clustering. The standard errors from the two‐stage regressions also account for the variance of the predicted regressor. All results described in the text are statistically significant at the 0.05 level or better.

3. DATA

3.1. Sample

The sample was drawn from the nationally representative MEPS‐HC, sponsored by the Agency for Healthcare Research and Quality. ²⁷ Each calendar year of information is derived from three interviews with an average reference period of 5 months. The MEPS‐HC asks questions about access to care in the second interview of the calendar year, and questions about preventive services are in the third and final interviews. The MEPS‐HC also collects detailed data about health insurance, as well as socioeconomic characteristics for the civilian, noninstitutionalized population.

The sample is low‐income nonelderly adult citizens. We exclude Supplemental Security Income and Medicare beneficiaries, because individuals with severe disabilities use many more services than the population affected by Medicaid eligibility expansions. We study 42, 366 adults with incomes not exceeding 138% of the FPL, because only a few states have ever had Medicaid eligibility above that threshold for adults. We exclude two groups because including them would impede convergence for the partial observability bivariate probit regression: (a) five states with fewer than 400 observations, and (b) childless adults aged 19‐20, who have their own eligibility thresholds (different from the child or adult threshold) in 16 states and would need their own state fixed effects.

3.2. Access and preventive services measures

The access measures that we examined included whether individuals had a usual source of care (other than an emergency room), had any unmet need, or experienced any delay in getting medical care or prescription drugs. The preventive services measures examined in this study came from a series of questions asked in the MEPS‐HC about receipt and timing of some preventive health services. We applied the time frame recommended by the USPSTF regarding these preventive services. ²⁸ , ²⁹ , ³⁰ We examine five measures of preventive services: flu shot, blood pressure screening, blood cholesterol screening, mammogram, and colorectal cancer screening. Colorectal cancer screening was defined by a blood stool test, a colonoscopy, or a sigmoidoscopy. First, we examine whether the adult had these services within past year. For each of these services, if applicable, we restrict the sample to the subpopulation of adults for which the service was recommended by the USPSTF. ²⁸ , ²⁹ , ³⁰ In particular, we examined cholesterol screening only for men aged 35 or older and women aged 45 or older, mammogram only for women aged 50 or older, and colorectal cancer screening only for men and women aged 50 or older. The USPSTF recommendation regarding the age at which women should start mammogram has changed within our study period, ³⁰ so we restrict our sample for mammogram to the age group for which mammogram was recommended throughout our study period.

We then applied the time frame recommended by the USPSTF regarding these preventive services, if more than one year. ²⁸ , ²⁹ , ³⁰ Three of these five preventive services had recommended time frames of more than a year: blood cholesterol screening within past five years, mammogram within past two years, and colorectal cancer screening (a blood stool test within the past year, or a sigmoidoscopy or colonoscopy within the past five years). Though the sigmoidoscopy and colonoscopy recommendations from the USPSTF are once in 5 years and once in 10 years, respectively, we chose to use the 5‐year time frame for both colonoscopy and sigmoidoscopy, because before 2009, MEPS HC did not distinguish between colonoscopy and sigmoidoscopy.

In addition, we examined whether the adult had a routine checkup within the past year. The USPSTF does not recommend an annual checkup for the general population, but many patients and physicians believe that it is important to receive annual medical checkups. ³¹ , ³²

3.3. Medicaid enrollment

Enrollment in Medicaid any time during the entire calendar year is the key explanatory variable. In our sample, 14 149 adults were enrolled and 28 217 adults were not enrolled.

3.4. Explanatory variables

The explanatory variables in X are age, sex, race/ethnicity, marital status, education, basis of potential eligibility, and fixed effects for state, year, basis of eligibility × state, and basis of eligibility × year. The bases of potential eligibility are being (a) parents of minor children and (b) childless adults aged 21 and older, who some states covered under program waivers or ACA expansions. We exclude childless adults aged 19‐20, due to their small sample size and because some states (16 in 2010) opted to cover this group.

3.5. Simulated Medicaid eligibility rate

Medicaid eligibility is derived from a detailed simulation model. ³³ , ³⁴ , ³⁵ Specifically, detailed state‐specific Medicaid eligibility rules are applied to information from the MEPS‐HC about family relationships, state of residence, and amounts and sources of income. We simulate eligibility through a comprehensive set of Medicaid eligibility categories. Our model accounts for a variety of differences across states before 2014, including whether gross or net income (or both) was used to determine eligibility, income counting rules for earned and unearned income, the income and asset thresholds for eligibility, the amount of income that was disregarded, rules for two‐parent families, and which family members counted for family size and income. For 2014 and 2015, we use federal income and family composition rules, as well as state eligibility categories and thresholds. The eligibility pathways simulated include family coverage for parents, waiver programs for parents, and waiver programs for childless adults.

Hence, the simulated Medicaid eligibility rate summarizes changes in eligibility across multiple dimensions. However, using the rate for each state‐year basis of potential eligibility could yield biased estimates if there are omitted variables correlated with both eligibility and access or preventive services use. As with prior studies, we simulate the eligibility rate for each state and year in a way that removes the impact of differences in population characteristics across states and over time. ¹⁹ , ²⁰ , ²¹ , ²² , ²³ Specifically, we apply the rules of each state‐year to the entire sample, regardless of the sample member's actual residence and year. We then create separate estimates by state, year, and parents vs childless adults. The technical Appendix S1 contains a detailed description of the instrument. A wide range of changes in Medicaid eligibility occurred during the time period studied (see Tables S1 and S2).

3.6. Recall

Calendar year information about MEPS sample members was collected in three interviews. Specifically, the first interview asked about Medicaid enrollment between January 1st and the interview, and the second interview asked about Medicaid enrollment between the first and second interviews. The timing between interviews varies across families depending on the ease with which families could be contacted for in‐person interviews and each family's availability. We measured the duration of the recall period for each interview in terms of months. Our measure of recall is the minimum months of recall across the three interviews, which removes some of the correlation between initial reluctant response and recall. Specifically, even if the initial interview was in July, at the end of the first fielding period, the subsequent interview would necessarily have a shorter recall period to complete fielding before the end of the year. In contrast to the questions about enrollment, the questions about access to care and preventive services were fielded in the second interview of the calendar year and asked about fixed time frames, for example, 12 months.

4. RESULTS

Table 1 presents descriptive statistics on the association between increases in Medicaid eligibility and access to care and preventive services utilization. Eligibility expansions were divided into two groups. Medium eligibility expansions were defined as 20‐50 percentage point year‐to‐year within‐state increases in the simulated Medicaid eligibility rate for parents or childless adults. Large eligibility expansions were defined as 50 percentage point or more increases in the Medicaid eligibility rate. For each medium expansion for parents, parents in the state up to 2 years before and up to 2 years after the expansion were included in the table, and similarly for childless adults and large expansions. There were no significant changes in access and utilization following the medium eligibility expansions. Large eligibility expansions, on the other hand, were followed by increased access to care and increased utilization of preventive services. In particular, the percentage of adults with unmet needs for medical care decreased from 8.9% to 5.9%, the percentage of adults with unmet needs for prescription drugs decreased from 6.9% to 4.1%, and the percentage of adults with delays getting prescription drugs decreased from 6.6% to 4.1%. The multivariate analyses leverage all changes in the Medicaid eligibility, including decreases and smaller increases, and hence a larger sample.

Table 1.

Pre‐post differences in access and preventive services, by size of Medicaid eligibility expansions

	Medium eligibility expansions a		Large eligibility expansions b
	Pre	Post	Pre	Post
Has usual source of care (%)	68.8	67.9	59.8	62.8
In the past year, percent with
Unmet needs for medical care	5.2	4.6	8.9	5.9**
Delays getting medical care	4.6	4.2	7.4	6.5
Unmet needs for prescription drugs	3.2	2.9	6.9	4.1**
Delays getting prescription drugs	3.0	3.0	6.6	4.1**
Checkup in past year	61.1	58.7	49.9	53.6
Flu shot in past year	26.5	27.8	24.0	26.9
Blood pressure check in past year	77.7	74.9	70.4	72.5
Sample size	1801	1880	2778	2823
Cholesterol screening c
In past year	63.8	59.6	57.0	61.9
In past 5 y	88.6	83.5	80.7	81.6
Sample size	549	565	1005	1008
Mammogram d
In past year	56.3	52.1	48.9	51.8
In past 2 y	69.0	69.6	65.0	69.6
Sample size	154	172	365	377
Colorectal cancer screening e
In past year	11.6	19.9	16.9	19.9
In past 5 y	39.9	45.4	38.9	40.8
Sample size	284	294	632	670

^aFor parents, the proportion of the nation's parents that would be eligible for Medicaid under the state's rules rose by 20‐50 percentage points. For childless adults, the proportion of the nation's childless adults that would be eligible for Medicaid under the state's rules rose by 20‐50 percentage points.

^bFor parents, the proportion of the nation's parents that would be eligible for Medicaid under the state's rules rose by more than 50 percentage points. For childless adults, the proportion of the nation's childless adults that would be eligible for Medicaid under the state's rules rose by more than 50 percentage points.

^cSample restricted to men aged 35 or older and women aged 45 or older.

^dSample restricted to women aged 50 or older.

^eColorectal cancer screening in past year was defined as any blood stool test, any sigmoidoscopy, or any colonoscopy in the last year. Colorectal cancer screening in past 5 y was defined as any blood stool test in past year, or any sigmoidoscopy/colonoscopy in the past 5 y. Sample restricted to adults aged 50 or older.

^*Pre‐post difference statistically significant at 0.05 level.

^**Pre‐post difference statistically significant at 0.01 level.

Source: Authors’ calculations from the Medical Expenditure Panel Survey, 2005‐2015. Nonelderly adult citizens with incomes ≤138% FPL and not Medicare or Supplemental Security Income beneficiaries. Sample limited to adults in states up to 2 y pre‐ and postexpansion.

The mean effects of changes in the simulated Medicaid eligibility rate on access and preventive services were statistically significant for some of our access measures and for flu shots (Table 2). In particular, a 1 percentage point increase in the simulated Medicaid eligibility rate increased the probability of having a usual source of care by 0.043 percentage points. This means that, for example, for states that did not cover any childless adults and then began covering all of them, having a usual source of care increased by 4.3 percentage points, because we assumed a linear relationship between the eligibility rate and outcomes. Similarly, a 1 percentage point increase in the simulated Medicaid eligibility rate decreased the probability of having any unmet need for medical care by 0.024 percentage points, decreased the probability of experiencing any delay in getting prescription drugs by 0.022 percentage points, and increased the probability of having a flu shot by 0.044 percentage points. The 2.4 percentage point decrease in unmet needs for medical care resulting switching from covering none to all is 27% of the pre‐expansion baseline in Table 1 for states with large expansions, 8.9 percent. The relative effects are also large for delays getting prescription drugs (33%) and having a flu shot (19%).

Table 2.

Effects of a 1 percentage point increase in the Medicaid eligibility rate on access and preventive service use

Outcome measures	Mean effects
Percent with usual source of care	0.043*
	(0.020)
In the past year, percent with
Unmet needs for medical care	−0.024**
	(0.009)
Delays getting medical care	−0.012
	(0.009)
Unmet needs for prescription drugs	−0.013
	(0.007)
Delays getting prescription drugs	−0.022**
	(0.006)
Checkup in past year	0.032
	(0.019)
Flu shot in past year	0.046**
	(0.014)
Blood pressure check in past year	0.017
	(0.021)
Sample size	41 617
Cholesterol test a
In past year	0.039
	(0.034)
In past 5 y	0.022
	(0.027)
Sample size	14 513
Mammogram b
In past year	0.021
	(0.041)
In past 2 y	0.021
	(0.042)
Sample size	4697
Colorectal cancer screening c
In past year	0.008
	(0.022)
In past 5 y	−0.002
	(0.035)
Sample size	8487

Standard errors in parentheses. Regression includes state, year, state × childless adult, and year × childless adult fixed effects.

^aSample restricted to men aged 35 or older and women aged 45 or older.

^bSample restricted to women aged 50 or older.

^cColorectal cancer screening is defined as any blood stool test in the past year or any sigmoidoscopy or colonoscopy in the last 5 y. Sample restricted to adults aged 50 or older.

^* P < .05.

^** P < .01.

Source: Authors’ calculations from the Medical Expenditure Panel Survey, 2005‐2015. Nonelderly adult citizens with incomes ≤138% FPL and not Medicare or Supplemental Security Income beneficiaries.

Table 3 presents the results of our additional analysis, which shows that effects of Medicaid eligibility remain statistically insignificant when we replace the Medicaid Eligibility Rate with cumulative eligibility over the recommended time frame, for preventive services that have recommended time frames of more than a year. Note that the samples are smaller, because simulated Medicaid eligibility rates are first available for 2005. For example, for cholesterol screening test over the past five years, we were only able to use data from 2009 and later.

Table 3.

Effects of years of simulated Medicaid eligibility during the recommended time frame on percent adhering to preventive service recommendation

Outcome measures	Mean effects
Cholesterol test in past 5 y a	1.8
	(1.0)
Sample size	10 040
Mammogram in past 2 y b	−0.4
	(2.5)
Sample size	4399
Colorectal cancer screening c	1.3
	(1.3)
Sample size	5898

Standard errors in parentheses. 2SLS = two‐stage least squares. Regression includes state, year, state × childless adult, and year × childless adult fixed effects.

^aSample restricted to men aged 35 or older and women aged 45 or older in the MEPS 2009‐2015.

^bSample restricted to women aged 50 or older in the MEPS 2006‐2015.

^cColorectal cancer screening is defined as any blood stool test in the past year or any sigmoidoscopy or colonoscopy in the last 5 y. Sample restricted to adults aged 50 or older in the MEPS 2009‐2015.

^* P < .05.

^** P < .01.

Source: Authors’ calculations from the Medical Expenditure Panel Survey (MEPS), 2006‐2015. Nonelderly adult citizens with incomes ≤138% FPL and not Medicare or Supplemental Security Income beneficiaries.

Turning to the 2‐stage model, we first present the partial observability bivariate probit results in Table 4. We used the coefficients to estimate mean marginal effects. Each percentage point increase the in the simulated Medicaid eligibility rate increased the probability of enrolling in Medicaid by 0.236 percentage points. The F test on this variable was 23, supporting its use as an instrument. Women and non‐Hispanic black adults were more likely to be enrolled, whereas married people, older adults, and those with more education were less likely to be enrolled. Controlling for state and year fixed effects, longer recall periods were associated with respondents less accurately reporting Medicaid enrollment. Medicaid enrollment was reported more accurately for non‐HIspanic black adults, older adults, and married people reported, and less accurately for adults with more than high school educations. Childless adults reported less accurately, consistent with recent research. ²⁵ The negative correlation in the errors for true Medicaid and reporting accuracy suggests that additional unobserved factors reduce accurate reporting among adults enrolled in Medicaid.

Table 4.

First‐stage results: mean marginal effects on probability of Medicaid enrollment

	Mean effect from partial observability bivariate probit regression
	True enrollment	Reporting accuracy
Medicaid eligibility rate	0.236**
	(0.050)
Recall		−0.011*
		(0.004)
Women	0.112**	0.054
	(0.042)	(0.058)
Race/ethnicity
Non‐Hispanic black	0.066**	0.081**
	(0.017)	(0.031)
Hispanic	0.021	−0.005
	(0.023)	(0.024)
Non‐Hispanic other	−0.005	0.035
	(0.016)	(0.030)
Age
21‐29	−0.152**	0.037
	(0.043)	(0.084)
30‐39	−0.308**	0.276**
	(0.094)	(0.088)
40‐49	−0.377**	0.442**
	(0.116)	(0.091)
50‐59	−0.367**	0.406**
	(0.139)	(0.122)
60+	−0.418**	0.479**
	(0.119)	(0.072)
Education
High school or GED	−0.075**	−0.024
	(0.011)	(0.013)
More than high school	−0.120**	−0.099**
	(0.015)	(0.034)
Married	−0.133**	0.057*
	(0.022)	(0.024)
Basis of potential eligibility
Childless	−0.069	−0.281*
	(0.042)	(0.130)
Correlation	−0.155
	0.469

Standard errors in parentheses. Regression includes state, year, state × childless adult, and year × childless adult fixed effects for true enrollment, and state and year fixed effects for reported enrollment.

^* P < .05.

^** P < .01.

Source: Authors’ calculations from the Medical Expenditure Panel Survey, 2005‐2015. Nonelderly adult citizens with incomes ≤138% FPL, not Medicare or Supplemental Security Income beneficiaries.

Table 5 shows the mean effects (local average treatment effects) from the two‐stage regressions. To highlight the relative magnitude of the effects, we also present the mean outcome measures for adults never enrolled in Medicaid. The effects of Medicaid enrollment were statistically significant for three access measures. In particular, Medicaid coverage increased the probability of having a usual source of care by 16.5 percentage points, decreased the probability of having any unmet need for medical care by 7.5 percentage points, and decreased the probability of experiencing any delay in getting prescription drugs by 7.7 percentage points.

Table 5.

Effects of Medicaid enrollment on access and preventive service use

Outcome measures	Baseline: Adults never enrolled in Medicaid	Mean effects from two‐stage regressions
Percent with usual source of care	57.2	16.5*
		(7.7)
In the past year, percent with
Unmet needs for medical care	8.1	−7.5*
		(3.4)
Delays getting medical care	7.0	−4.0
		(2.7)
Unmet needs for prescription drugs	5.3	−6.3
		(3.5)
Delays getting prescription drugs	4.9	−7.7*
		(3.1)
Checkup in past year	46.5	17.1*
		(8.2)
Flu shot in past year	21.1	12.6**
		(4.2)
Blood pressure check in past year	67.3	11.9
		(6.7)
Cholesterol test a
In past year	53.7	7.5
		(9.4)
In past 5 y	77.4	−1.6
		(8.0)
Mammogram b
In past year	46.4	13.3
		(15.8)
In past 2 y	63.8	4.9
		(14.2)
Colorectal cancer screening c
In past year	15.3	3.8
		(7.3)
In past 5 y	49.6	0.2
		(12.0)

Standard errors in parentheses.

^aSample restricted to men aged 35 or older and women aged 45 or older.

^bSample restricted to women aged 50 or older.

^cColorectal cancer screening in past year was defined as any blood stool test, any sigmoidoscopy, or any colonoscopy in the last year. Colorectal cancer screening in past 5 y was defined as any blood stool test in past year, or any sigmoidoscopy/colonoscopy in the past 5 y. Sample restricted to adults aged 50 or older.

^* P < .05.

^** P < .01.

Source: Authors’ calculations from the Medical Expenditure Panel Survey, 2005‐2015. Nonelderly adult citizens with incomes ≤138% FPL, not Medicare or Supplemental Security Income beneficiaries.

The effects of Medicaid enrollment were statistically significant for two preventive services measures. Medicaid coverage increased the probability of having a checkup by 17.1 percentage points and the probability of having a flu shot by 12.6 percentage points. The effects for both access and these preventive services are large relative to the means among adults who never reported having Medicaid.

5. DISCUSSION

Using changes in eligibility to generate local average treatment effects, we found that Medicaid enrollment was associated with improved access to care and increased utilization of two preventive services, flu shots and checkups. These results are consistent with those of the Oregon Health Insurance Experiment (OHIE), which found larger increases in access to care and use of some preventive services. ⁴ , ⁵ , ⁶ The OHIE enrolled only people who lacked insurance, whereas other Medicaid expansions have not imposed this requirement. Given our relatively more advantaged treatment group, it is not surprising the local average treatment effects are smaller in our study. The fact that adults in nonexpansion states with income over 100% of FPL had access to Marketplace subsidies starting in 2014 may also have contributed to dampening the effect of Medicaid. However, our estimated effects of Medicaid on access are somewhat larger than those found by Winkelman et al, ¹⁴ but their study did not account for underreporting and unobserved factors associated with gaining Medicaid coverage.

Obtaining Medicaid may have an immediate impact on the use of preventive services. On the other hand, adults churn on and off Medicaid, ¹⁵ and adults who were previously enrolled in Medicaid had access to preventive services in the past, so impacts on use during recommended time frames may be attenuated. In our analysis, however, we did not find evidence of any impact of Medicaid enrollment on cholesterol screening, mammogram, or colorectal cancer screening, whether we examine utilization in past year or utilization within the recommended time frames for these services, and whether we used simulated eligibility rate of the current year or that of the recommended time frame. The sample sizes for these variables, however, were smaller than the sample sizes for other variables, as the subpopulations of adults for whom these services were recommended were small. Additional research, perhaps with larger sample size, might be needed to examine the impact of Medicaid enrollment on preventive services utilization of these subpopulations.

5.1. Limitations

The recommended time frames for preventive services incorporated in this analysis are generally applicable only to people who have no signs or symptoms of the specific disease or condition under evaluation. ³⁶ For higher risk individuals, the recommended frequencies are generally higher. We, however, do not have enough information in our data to determine which adults should be considered high risk for specific diseases.

The two‐stage method used in this analysis has limitations. We estimated local average treatment effects on those who enrolled because of changes in eligibility rates, and the effects may differ for other marginal enrollees. In addition, the results may not generalize beyond nondisabled, nonelderly adults enrolled in Medicaid for a full calendar year during 2005‐2015.

We assumed that changes in Medicaid eligibility affect outcomes only through changes in Medicaid enrollment. While we believe that this is a reasonable assumption, there are a few potential ways in which this assumption might be violated. First, this assumption would be violated if providers lacked capacity to serve new Medicaid enrollees and constrained capacity reduced access. However, other research suggests there is no consistent evidence of spillovers to people who already had insurance. ³⁷ Second, other program characteristics may be correlated with Medicaid eligibility, and state × eligibility group and year × eligibility group fixed effects may not fully account for those factors. Third, there may be income effects within families with mixed eligibility and, for practices with larger caseloads of low‐income people, income effects on provider behavior.

Our model also assumes that misreported Medicaid is entirely due to underreporting. This is consistent with 17.5% of MEPS sample members with Medicaid reporting another source of coverage or no insurance throughout the year, ³ a level of underreporting found in a variety of studies. ² (Note that, in the MEPS, some reports of other state or local public coverage are edited to be Medicaid, so the impact of underreporting is somewhat less severe.) Nonetheless, it is likely that some people without Medicaid report it. In their analysis of the sensitivity of their two‐stage model to misspecification, Nguimkeu et al ¹ found that the estimated effects exhibit little bias with low rates of false positives.

Lastly, we used a linear probability model for binary outcomes, and estimating a trivariate probit model with partial observability may yield different results.

6. CONCLUSION

Our estimates suggest that, among nonelderly, nondisabled adults, Medicaid enrollment promotes improved access to care and increased use of some preventive services. Additional research is needed on impacts for subgroups, such as parents, childless adults, and the smaller and generally older populations for whom screening tests are recommended.

DISCLAIMER

The authors are employees of the Agency for Healthcare Research and Quality. The views and opinions expressed in this article are those of the authors, and no official endorsement by the Department of Health and Human Services or the Agency for Healthcare Research and Quality is intended or should be inferred.

Supporting information

AppendixS1

Hill SC, Abdus S. The effects of Medicaid on access to care and adherence to recommended preventive services. Health Serv Res.2021;56:84–94. 10.1111/1475-6773.13603