Medicaid Enrollment Churn and Implications for Continuous Coverage Policies
Data Source and Linkage
Our analysis is based on the Transformed Medicaid Statistical Information System (T-MSIS) Analytic Files (TAF) Research Identifiable Files (RIF). We include beneficiaries who were enrolled at any point in 2018. We use the 2018 Demographic Eligibility (DE) Base file to determine eligibility pathway based on last-best eligibility data. We draw enrollment dates from the 2017-2019 DE Dates files. Our analysis uses enrollment dates from 2017 and 2019 for look-back and look-ahead years, respectively, so that we can examine what happens to enrollees a full year before and after an enrollment start date or disenrollment date in 2018. Data for 2019 are from the preliminary version of the T-MSIS RIF. All other DE Base and DE Dates files are final versions (Release 1).
State Exclusion Criteria
We use 41 states in our analysis. We exclude 10 states (FL, IN, KY, ME, MS, NE, OK, OR, UT, and WY) due to missing or inconsistent data based on state-level information available at the DQ Atlas as well as our own analysis. We relied on data quality assessments from DQ Atlas to exclude states that have a “medium concern”, “high concern”, or “unusable” data quality assessment for: (1) the average monthly Medicaid/CHIP enrollment compared to an external enrollment benchmark (Medicaid/CHIP Performance Indicator Data) (five states (IN, KY, ME, MS, NE)); (2) the average length of enrollment gaps, and (3) the percent of beneficiaries with overlapping Medicaid and S-CHIP enrollment spans (no states were excluded based on measures 2 or 3). We further excluded states based on: (4) the percent of beneficiaries missing an eligibility group code (threshold of >=10%, excluded OR); (5) the percent of beneficiaries with only one enrollment span (threshold of >=99.5, excluded FL, KY, and WY); and (6) the percent of beneficiaries with three or more enrollment spans in a year (threshold of >=5%, excluded OK). Notably, our exclusion criteria for related to the number of enrollment spans in a year are less restrictive than that in the DQ Atlas because we did not want to make assumptions about the number of enrollment spans in our analysis, but did want to remove extreme cases that are the most likely to represent inaccurate enrollment dates.
For analysis that includes restricted benefit enrollees, we further excluded two states (CA and ND) due to large amounts of missing data among restricted benefit enrollees. CA was excluded because roughly 75% of observations coded with restricted benefits in the DE Base file were missing a federal identifier (BENE_ID). ND was excluded due to large amounts of missing enrollment data in the DE Dates files and, after merging the DE Dates and DE Base files, more than half (53%) of restricted benefit enrollees in the DE Base file did not have a matching BENE_ID in the DE Dates file and were dropped from the analysis.
Beneficiary Linking, Eligibility Classification, and Exclusion
We linked individuals across years using BENE_ID, which are unique enrollee identifiers created by the Chronic Conditions Warehouse and are recommended for use when combining data for multiple years. We also use BENE_ID to link the DE Base and DE Dates files. We assigned restricted benefit status and eligibility group code using a last-best approach for 2018, which assigns eligibility based on the most recent eligibility code in 2018. We classified eligibility groups using a hierarchy that first checked if the eligibility group code was missing, then for medically needy eligibility, disability (under age 65), and expansion adult. Any enrollees in the DE Base file that had a non-missing eligibility code and had not been assigned an eligibility group through this hierarchy were then were assigned by age to children (ages 20 and under), adults (ages 21 to 64), and aged (ages 65 and over).
The 2018 DE Base files for 41 states in our analysis contained 82.9 million observations after removing a small number of “dummy” records that represent enrollees who have claims data but no eligibility data provided by the states. We removed observations missing eligibility codes or restricted benefit status codes (624,000 observations), people qualifying through a medically needy pathway (1.7 million observations), people missing a BENE_ID for linking files (3.1 million observations), and people with duplicated BENE_ID (262,000 observations). After merging with the DE Dates files, there were less than 5,000 enrollees who did not have a matching BENE_ID in the DE Dates and DE Base files and were dropped from the final sample. Our final sample included 77.2 million unduplicated enrollees. Of these, 71.3 million were full-benefit enrollees. Our analysis of restricted benefit enrollees included 5.4 million people in the 39 states included in that component of the analysis.
The DE Dates files provide a start and end date for every enrollment span in our time period. As noted above, our analysis uses enrollment dates from 2017 and 2019 for look-back and look-ahead years for a more complete picture of churning. Before calculating churn rates, we first merged all overlapping and contiguous enrollment spans for enrollees, which we defined as enrollment spans that are separated by one day or less. For example, if a person has two enrollment spans with an end date and a start date that are one day apart (i.e., the person disenrolled and re-enrolled the next day), we considered these spans to be contiguous and merged them into one enrollment span. After merging overlapping and contiguous spans, we identified enrollment gaps, which we defined as the number of days between two enrollment spans for an enrollee. To be counted as “churn” in our analysis, a person would have had (1) an enrollment gap of 365 days or less and (2) and enrollment gap that started or ended in 2018. For example, a person that disenrolled in 2017 but then re-enrolled within 365 days in 2018 would be included in our churn estimates. Similarly, a person who disenrolled in 2018 and then re-enrolled in less than 365 days would also be counted in our churn estimates. Sensitivity analyses that only included gaps longer than 31 days showed a marginal decrease in churn rates (approximately 1 percentage point); while shorter gaps could reflect data reporting errors, they also could be true disenrollments followed by relatively quick re-enrollment.
Previous studies have estimated Medicaid churn rates at around 10%. There is some variation across studies due to use of different data sources, national versus state-specific estimates, and the focus on different populations (e.g., children versus adults). A recent analysis from MACPAC, which used the same data source but different methods as our analysis found that 8% of Medicaid and CHIP enrollees re-enrolled in coverage within one year of disenrolling. The most significant difference in our methods from MACPAC analysis was that the MACPAC analysis only utilized a look-ahead year, while our analysis includes a look-ahead year and a look-back year. There are other differences in methods from MACPAC’s analysis, but those likely have a smaller impact than the use of a look-back year. Our estimate of a 10% churn rate overall is also similar to estimates of 9% and 11% from studies of churn in specific states.