Health Care Expenditures Among Adults With Diabetes After Oregon’s Medicaid Expansion

The prevalence of diabetes in the population has increased substantially in the U.S. In 2017, more than 30 million people were diagnosed with this condition, more than twice as many as compared with 20 years ago (1–3). The economic costs associated with the condition are substantial, reaching around $327 billion in 2017 (4). Among people without health insurance coverage, diabetes is more likely to remain undiagnosed and untreated, and uninsured populations also have less health care access and lower utilization than insured populations (5–8). Therefore, gaining health insurance is likely to improve treatment and well-being among people with diabetes.

The Affordable Care Act (ACA) Medicaid expansion is the most significant policy change to increase access to health insurance for low-income Americans in the recent history of the U.S. It offered additional federal support for states that chose to increase Medicaid eligibility to U.S. citizens and qualified legal immigrants with an income at or below 138% of the federal poverty level (FPL) (9). After this historic expansion, millions of previously uninsured people became newly enrolled in Medicaid. As of June 2017, a total of 74 million people were insured by Medicaid, which represented an increase of more than 10 million enrollees compared with pre-expansion years (10,11). It is well documented that obtaining insurance due to the expansion was associated with increases in access and utilization relative to being uninsured (12–20). However, it is less clear as to whether the newly enrolled can be expected to have greater or lower expenditures than the population with existing Medicaid coverage (21,22). These concerns are heightened for population groups with high utilization and expenditures, such as those with diabetes.

Understanding expenditure variations between Medicaid recipients with different eligibility categories is important for several reasons. First, it can help U.S. state policy makers understand how expenditures may change as they make decisions regarding Medicaid expansions. Second, disaggregated expenditures can help identify differences in utilization that might explain differences in overall expenditures. Third, the assessment of expenditures by provider and service type can inform potential policies aimed to improve the quality of care for Medicaid enrollees with diabetes.

In this study, we examined Medicaid expenditures of a cohort of adult Oregon Medicaid recipients with diabetes who were either newly eligible due to the expansion or previously eligible. Oregon opened Medicaid eligibility in 2014 to adults with a household income at or below 138% of the FPL. By comparing overall expenditures for these two groups, we were able to assess whether newly eligible Medicaid beneficiaries with diabetes had higher expenditures than those who were previously eligible. Moreover, by defining 12 service categories, we were able to examine whether expenditures for these two groups were different across provider and service types.

This is a retrospective cohort study of adults with diabetes using Oregon Medicaid claims and enrollment data from January 2014 to September 2016.

We included Oregon Medicaid recipients between 19 and 64 years old (the population most affected by the ACA expansion), not pregnant (pregnant women have other public health insurance options and very different expenditures from those who are not pregnant), and not dually eligible for Medicaid and Medicare (we did not have information about Medicare expenditures). We then defined our study cohort as all Oregon Medicaid recipients who were enrolled all months of the study period (January 2014 to September 2016) and who had at least one claim with a positive diabetes diagnosis, defined below, during the first three quarters of 2014.

We used the first three quarters of 2014 to identify people with diabetes because 1) most of Oregon’s increase in Medicaid enrollment due to the expansion occurred during that time and 2) inspection of all newly eligible Medicaid recipients revealed a strong increase in diabetes prevalence during the first three quarters, likely because new Medicaid enrollees were diagnosed in claims during that time (see Supplementary Fig. 1). Diabetes diagnoses were based on the Healthcare Effectiveness Data and Information Set (HEDIS) comprehensive diabetes care measure (23). The HEDIS diabetes care measure identifies the population with type 1 and type 2 diabetes using prescriptions, ICD-9 (we only used claims before the ICD-10 introduction in October 2015 to identify people with diabetes), current procedural terminology, and uniform billing codes.

After defining the study cohort, we distinguished between newly eligible Medicaid recipients due to the expansion and, as a comparison group, previously eligible recipients. Ideally, we would have identified these groups based on state eligibility criteria before and after the expansion and individual information on income, assets, and family structure. Since our data did not contain individual eligibility information, we defined eligibility using claims before the Medicaid expansion. Specifically, individuals in our study cohort with at least 1 month of enrollment between 2009 and 2013 were considered previously eligible (i.e., eligible for Medicaid before the expansion), and individuals in our study cohort not enrolled between 2009 and 2013 were considered newly eligible (i.e., eligible only after the Medicaid expansion). This approach is similar to other articles examining the effect of the ACA for the newly insured (24,25).

We used allowed amounts recorded in claims to calculate individual expenditures for each quarter and then divided these quarterly expenditures by months enrolled during that quarter to obtain per-member per-month (PMPM) expenditures. Most claims recorded allowed amounts (i.e., reimbursement for an encounter); however, Oregon implemented alternative payment methodology for enrollees receiving care from select clinics. For these claims, information was available on diagnoses and procedures but not on allowed amounts. We followed previous literature in imputing these expenditures (26,27). Details about the imputation are contained in the Supplementary Data.

After imputing costs, we calculated total PMPM Medicaid expenditures as well as PMPM expenditures related to each of the following 12 categories: 1) inpatient stays, 2) emergency department (ED) visits, 3) primary care physician (PCP) visits, 4) specialist visits, 5) prescription drugs, 6) transportation services, 7) laboratory tests, 8) imaging and echography, 9) procedures, 10) durable medical equipment, 11) evaluation and management, and 12) other or unknown services. Details about the information used to create these categories are included in the Supplementary Data.

To create mutually exclusive expenditure groups, we considered ED visits to be only those ED visits that were not also inpatient stays, PCP visits to be only those PCP visits that were not inpatient stays or ED visits, and specialist visits to be all visits that were not inpatient, ED, or PCP visits. Expenditures for the 12 categories summed to total expenditures.

Patient characteristics derived from the Medicaid enrollment data used in our analysis included: age, race (white, black, Native American/Alaska Native, Asian or Pacific Islander, or other/unknown), ethnicity (Hispanic, non-Hispanic, or other/unknown), sex, and whether a person resided in an urban area. We also included the Charlson Comorbidity Index, which was developed to classify comorbidity conditions associated with mortality risk and includes diabetes as a comorbidity (28,29). We used the updated version of the index that reflects changes in the contribution of comorbidities to mortality (30) and the ICD-10 code–based construction of the index starting October 2015, when ICD-9 codes were changed to ICD-10 codes (31). In addition to the Charlson Comorbidity Index, we also included 11 mental health and substance abuse categories: 1) alcohol disorders, 2) substance use disorders, 3) schizophrenia and other nonmood disorders, 4) bipolar disorder, 5) major depression, 6) dysthymia, 7) disorders originating in childhood, 8) anxiety disorders, 9) personality disorders, 10) adjustment disorders, and 11) other behavioral conditions (32). The ICD codes used for these mental health and substance abuse categories are shown in Supplementary Table 1. The Charlson Comorbidity Index and the mental health and substance abuse categories were created for each study quarter separately and only used claims information from that quarter.

We propensity score matched previously to newly eligible Medicaid recipients to produce equivalent risk distributions because the newly eligible were likely to be in better health and therefore have lower expenditures. For each individual in our study cohort, we selected their first enrollment quarter to match previously and newly eligible Medicaid recipients. As a first step, we estimated a logistic regression with the dependent variable denoting whether a Medicaid recipient was newly or previously eligible and, as independent variables, the patient characteristics described in the previous section (10patient characteristics) as well as the number of months enrolled between January 2014 and September 2016 (range 1–33). In a second step, we calculated propensity scores and used these to perform a nearest neighbor match without replacement between the previously eligible and the newly eligible group with a 1:1 ratio and a caliper of 0.5. We estimated propensity scores using the twang (33) package in R (which uses a generalized boosted model) and performed matching using the Matching package in R (34). We used 1:1 matching to reduce the influence of possible outliers, i.e., people with high expenditures, who might otherwise have gotten matched multiple times. To assess matching quality, we calculated the absolute standardized mean difference (ASMD) between previously and newly eligible Medicaid recipients before and after matching. An ASMD of 0.1 or less was considered to indicate good balance (35).

In the matched sample, we examined quarterly trends in expenditures for the two groups using linear regression models. Specifically, we estimated one linear regression model for each calendar quarter separately, using PMPM expenditures as the outcome and a binary variable (equal to 1 if an enrollee was newly eligible and 0 if previously eligible) as explanatory variable. The coefficient of the intercept estimated average expenditures of previously eligible Medicaid recipients, and the coefficient of the intercept plus the coefficient of the newly eligible variable estimated average expenditures of the newly eligible. By combining estimates from each calendar quarter (description below), we obtained trends in average expenditures. We then repeated these regressions for each of the expenditure types. SEs were clustered at the matched pair level. We reported mean expenditure levels by eligibility group per quarter for each of the expenditure categories as well as differences and significance values.

We used this regression approach for two reasons. First, we did not have to make parametric assumptions about expenditure trends and assumptions about temporal correlation within individuals over time because we estimated separate regressions for each calendar quarter. Second, using regression models allowed us to add other variables used for matching as a sensitivity check.

Our choice of linear models was informed by several considerations. Linear models are relatively easy to estimate and have good validity in a large sample even with a nonnormally distributed outcome (36,37). Other more complicated methods, such as generalized linear models, may in some instances perform better but require making other assumptions about the distribution of the outcome. Finally, by using linear models, we followed other articles published in leading medical and health service journals (26,38–43).

Although linear models are often used for modeling health care expenditures, this model choice is less valid when outcomes are highly skewed. Moreover, one important drawback of linear regression is that it is sensitive to extreme values. These concerns are especially relevant for inpatient expenditures. To assess changes in the distribution of this expenditure category, we further estimated a logistic regression model with any PMPM inpatient expenditures as outcome and five quantile regression models for the five highest PMPM inpatient expenditure percentiles (where expenditures are positive) as a sensitivity check (44). We clustered SEs at the matched pair level for these regressions as well.

We performed several sensitivity checks. First, to assess how remaining imbalance in characteristics affected expenditure trends, we included all measures used for matching as additional independent variables in the linear regressions (45). And second, we examined expenditures for 1) the full, unmatched cohort sample and 2) the larger sample of matched individuals who were enrolled for at least 2 months during the first three quarters of 2014 but not necessarily for the full study period.

Before matching, the sample included 5,683 newly eligible Oregon Medicaid recipients with diabetes and 11,252 previously eligible Medicaid recipients with diabetes (Table 1). In the unmatched sample, newly eligible Medicaid enrollees tended to be older, male, and of other or unknown race, compared with previously eligible Medicaid enrollees. They were also more likely to report being Hispanic. Furthermore, they were more likely to have a lower comorbidity score and, with the exception of alcohol use disorders, less likely to have one of the mental health or substance use conditions. After matching, the sample consisted of 4,213 newly and 4,213 previously eligible Medicaid recipients. Matching reduced these differences and resulted in ASMD <0.1 for all characteristics.

Total PMPM expenditures of the newly eligible in the matched sample grew rapidly from $561 PMPM in the first postexpansion quarter to $889 PMPM in the third postexpansion quarter and then continued to grow more slowly afterward, to an average of $1,097 PMPM during the last three quarters of the study period (Fig. 1). Total PMPM expenditures of the previously eligible started higher, at $793 PMPM, and grew steadily throughout the study period, reaching an average of $1,249 PMPM during the last three quarters of the study period. PMPM expenditures for both groups grew at the same rate after the first three quarters. Total PMPM expenditures of the previously eligible were higher than total PMPM expenditures of the newly eligible for all study quarters. Differences were statistically significant for five of the eight last study quarters (see Supplementary Table 2 for statistical significance of expenditure differences).

PMPM inpatient expenditures were a large expenditure group, averaging $214 PMPM for the newly eligible and $239 PMPM for the previously eligible (Fig. 2). This PMPM expenditure category had high variance due to the skewed distribution, but PMPM expenditures of the newly eligible decreased slightly from an average of $236 PMPM during the first three quarters to an average of $214 PMPM during the last three quarters, whereas PMPM inpatient expenditures for the previously eligible increased from an average of $220 PMPM during the first three quarters to an average of $267 PMPM during the last three quarters. Differences between the two groups were not statistically significant, except for two study quarters (see Supplementary Table 2). Logistic regression coefficients were initially positive and not statistically significant but then turned negative and were statistically significant for five of the last eight study quarters (Supplementary Table 3). Quantile regression coefficients were in most cases not statistically significant.

PMPM prescription drug expenditures for the previously eligible started at $260 PMPM and increased steadily until the last three quarters of the study period, where these PMPM expenditures reached an average level of $432 PMPM (Fig. 3). PMPM prescription drug expenditures for the newly eligible started low, at $91 PMPM, increased sharply to $243 PMPM during the third quarter of 2014, and then continued to increase more slowly, reaching $407 PMPM during the last three study quarters. Differences were statistically significant until the last two study quarters (see Supplementary Table 2).

PMPM expenditures for ED, PCP, and specialist visits were comparably minor and none of them exceeded $100 PMPM for either group during the study period (see Supplementary Figs. 2–4). ED expenditures for the newly eligible averaged $39 PMPM over the study period and were stable. ED expenditures for the previously eligible averaged $49 PMPM and increased slightly over the course of the study period. Differences in these PMPM expenditures between the two groups were significant for the last eight study quarters (see Supplementary Table 2).

During the first four quarters, PMPM PCP expenditures for the newly eligible were significantly higher than PMPM PCP expenditures of the previously eligible. Afterward, PMPM PCP expenditures for both groups were almost identical, not statistically different from each other, and stable (mean values $51 PMPM and $49 PMPM for newly and previously eligible, respectively). PMPM expenditures for specialist visits started lower for the newly insured compared with the previously insured but then were slightly higher throughout most of the remaining study quarters. Average PMPM specialist expenditures were $61 PMPM and $55 PMPM for the newly and previously eligible, respectively.

PMPM expenditures for the other categories were minor to moderate and fairly similar for the two eligibility groups (Supplementary Figs. 5–11).

PMPM expenditures for transportation, durable medical equipment, procedures, and other and unknown services were small to moderate and trended slightly upwards for both groups. Average PMPM expenditures for the newly and previously eligible were as follows: $20 PMPM and $31 PMPM (transportation), $37 PMPM and $44 PMPM (durable medical equipment), $86 PMPM and $109 PMPM (procedures), and $44 PMPM and $56 PMPM (other/unknown services). Differences for the expenditure categories tended to be not statistically significant, with the exception of expenditures related to transportation (see Supplementary Table 2).

PMPM expenditures for laboratory tests, imaging and echography, and evaluation and management were small and similar. Expenditure differences between the two groups were mostly insignificant, with the exception of evaluation and management expenditures (see Supplementary Table 2). PMPM expenditures for laboratory tests among the previously eligible were very stable. PMPM expenditures for laboratory tests among the newly eligible increased slightly during the first three quarters and then had the same level as PMPM expenditures for the previously eligible. PMPM expenditures for imaging and echography were stable over time for both groups. PMPM expenditures for evaluation and management were small but increased slightly for both groups over time.

During the first three postexpansion quarters, total PMPM Medicaid expenditures of the newly eligible increased by $137 PMPM relative to the increase in total PMPM Medicaid expenditures of the previously eligible (Supplementary Table 4). Afterward, changes in total PMPM expenditures for the two groups were almost identical. Decomposing the change in total PMPM expenditure during the first three postexpansion quarters into PMPM expenditure changes for each of the 12 categories shows that for all but two categories, PMPM expenditures of the newly eligible increased relative to PMPM expenditures of the previously eligible. By far the largest relative increase was for PMPM prescription drug expenditures. The two categories where PMPM expenditures of the newly eligible decreased relative to PMPM expenditures of the previously eligible during these first three postexpansion quarters were PMPM inpatient stays and PMPM ED visits, with PMPM inpatient stays having by far the largest relative decline. Between the fourth and last study quarter, PMPM prescription drug expenditures of the newly eligible continued to increase relative to PMPM prescription drug expenditures of the previously eligible, whereas PMPM inpatient expenditures of the newly eligible continued to decrease relative to PMPM inpatient expenditures of the previously eligible.

Adjusting for the remaining imbalance in individual characteristics changed coefficients only slightly (Supplementary Table 5). PMPM expenditures for the unmatched sample (Supplementary Figs. 12–24) and for individuals with some enrollment during the study period (Supplementary Figs. 25–37) followed patterns similar to those of the main sample. PMPM inpatient, ED, and PCP expenditures for the previously eligible were higher compared with the newly eligible in the unmatched sample.

Our analysis of Oregon Medicaid claims data after the ACA expansion for enrollees with diabetes highlights important insights into how expenditures and utilization changed for those who became eligible for Medicaid due to the expansion. We found a strong and ongoing increase in PMPM expenditures for this group throughout our study period. Most of that increase occurred during the first expansion year, but PMPM expenditures continued to grow during the 2nd and 3rd year as well. Importantly, the increase was unevenly distributed across expenditure categories. Most notably, PMPM prescription drug expenditures of the newly eligible started at a low level but increased strongly throughout the study period. By contrast, PMPM expenditures for other categories such as ED use or inpatient stays remained fairly stable or even slightly decreased.

The overall increase in PMPM Medicaid expenditures among newly eligible Medicaid enrollees with diabetes suggests that they received increased treatment after gaining health insurance. The initial increase in total PMPM expenditures during the first three postexpansion quarters could also be related to the newly eligible receiving a diabetes diagnosis after Medicaid enrollment because being uninsured is associated with lower diabetes detection (6). Our results of increasing PMPM expenditures for the newly insured after the expansion are also consistent with previous research on the Medicaid expansion that documented an increase in utilization among this population due to the reform (16,20).

Comparing results from the unmatched sample to results from the matched sample also provides additional insights. First, we found that total PMPM expenditures of newly eligible Oregon Medicaid enrollees with diabetes were much lower than total PMPM expenditures of all (matched and unmatched) previously eligible Medicaid enrollees with diabetes. This finding adds to a growing literature demonstrating that those who became eligible due to the Medicaid expansion have, on average, lower expenses than those who were eligible before the ACA expansion (21,25). And second, matching strongly reduced this difference, suggesting that differences in patient characteristics between the previously and newly eligible in the unmatched sample explain most of their expenditure differences. These differences in patient characteristics reflect changing eligibility due to the expansion. Prior to the expansion, Medicaid was only accessible to children living in low-income households or adults with low income who had special medical needs (e.g., due to disability or pregnancy) (46). The expansion increased eligibility to all Oregon adults with incomes at or below 138% FPL, thereby offering Medicaid insurance to an adult population with, on average, less severe medical conditions (27).

One possible contextual factor that might explain the similarity in spending for newly and previously eligible Medicaid enrollees with diabetes is the creation of Oregon coordinated care organizations (26,47). The coordinated care organizations are a type of managed care organization that combines a global budget with quality measures, some of which are tied to financial payments. This organizational structure may have facilitated access for the newly eligible and thereby contributed to similar expenditure patterns among the newly and previously eligible.

Our study has several limitations. First, claims-based identification of diabetes cases is imperfect. For instance, some patients with diabetes might not have accessed a PCP during the first three postexpansion quarters that we used to identify diabetes. These patients were excluded from our analysis. Similarly, some Medicaid patients with diabetes might not have been diagnosed with the condition even though they visited a physician. Conversely, some Medicaid patients might have been diagnosed with diabetes based on the HEDIS measure and thus were included in our study even though they did not have the condition. Including such false positives reduces mean expenditures, but it is unclear whether such misdiagnosis is more pronounced for one of our expansion groups and, thus, also affected differences between the two groups. Second, we did not have exact eligibility information (i.e., income, assets, and family structure) and used an imperfect approach to identifying these two groups. As a result, some of the patients classified as newly eligible using our approach might have been eligible under pre-expansion eligibility rules (e.g., if they became disabled), while some of the patients classified as previously eligible using our approach might not have remained eligible under pre-expansion eligibility rules (e.g., their income or health improved). In the unmatched sample, such misclassification would have led to an overestimation of expenditure trends for the newly eligible (because those misclassified likely had worse health and, correspondingly, higher expenditures) and underestimation of expenditure trends of the previously eligible (because those misclassified likely had better health or higher income and, correspondingly, lower expenditures). However, matching accounted for differences in health and thus likely addressed such misclassification issues. Third, our comparison of the unmatched and matched sample showed that matching eliminated most expenditure differences between the two eligibility groups, but we cannot rule out that PMPM expenditure differences in the matched sample were due to differences in characteristics unobserved to us. However, a difference-in-differences research design that compares pre- and postexpansion expenditures of the newly eligible to either the previously eligible or the uninsured and that could eliminate differences due to unobserved factors was not feasible because claims records only exist for those currently enrolled in Medicaid. Last, our study is limited to one state (Oregon); thus, some findings presented here might not generalize to other states.

Our analysis of Oregon Medicaid enrollees with diabetes yields several important implications for health policy. First, our results could help states make expansion decisions by showing lower expenditures for newly eligible Medicaid enrollees with diabetes compared with previously eligible Medicaid enrollees with diabetes. And second, our study demonstrates the usefulness of examining detailed expenditure categories. Specifically, we showed that prescription drug expenditures increased strongly among the newly eligible, while inpatient expenditures of the newly eligible decreased somewhat relative to such expenditures of the previously eligible.

Acknowledgments. The authors acknowledge the significant contributions to this study that were provided by collaborating investigators in the NEXT-D2 (Natural Experiments for Translation in Diabetes 2.0).

Funding. This study was funded by a joint grant from the Centers for Disease Control and Prevention, the National Institute of Diabetes and Digestive and Kidney Diseases, and the Patient-Centered Outcomes Research Institute (CDC U18DP006116).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. S.R.L. conducted the analysis and led manuscript development. M.M., J.O., and R.S. provided methodological input and revised the manuscript. H.A., S.R.B., M.H., K.J.M., and J.D. provided conceptual input and revised the manuscript. N.H. provided methodological and conceptual input and revised the manuscript. S.R.L. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented at the 46th Annual Meeting of the North American Primary Care Research Group, Chicago, IL, 9–13 November 2018.