(ASPOT05) Untangling National Medicaid Data: 30 Data Quality Metrics to Support Maternal Health Studies in Two Common Data Models

Background: National U.S. Medicaid and Children’s Health Insurance Program (CHIP) data are an amalgamation of up to 53 jurisdictions with both fee-for-service and managed care plans that employ distinct benefits and coverage rules, making a highly heterogeneous data source. During 2014-2016, the U.S. migrated data collection and processing to the Transformed Medicaid Statistical Information System and these data are available for analytic use. Using this data source in pharmacoepidemiology studies can be challenging without extensive data quality assessment on a jurisdiction-year-plan type basis.

Objectives: To develop 30 data quality metrics to characterize jurisdiction-specific data that have been structured in the Sentinel and Observational Medical Outcome Partnership (OMOP) common data models (CDMs).

Methods: We recruited five technical experts with experience using Medicaid data and common data models. Experts prioritized metrics that would assess a jurisdiction’s fit-for-purpose qualities for maternal health studies. Following two rounds of discussion, prioritization, and voting, the panel selected 30 data quality metrics in five categories: demographics (9), enrollment (6), utilization (11), death (3), and data anomalies (1). These metrics were applied in two CDMs to take advantage of, and augment, those models’ standing data quality metrics and existing resources such as Medicaid’s DQ Atlas.

Results: While much of Medicaid’s data was reliable and usable for pharmacoepidemiology studies, in a few jurisdictions we found negative lengths-of-stay for inpatient stays, enrollment for infants that pre-dated conception, changes in the volume of utilization across years suggesting data missingness, and notable undercapture of emergency department visits. Many issues were addressable with jurisdiction-specific remedies. In maternal health, coverage of pregnant persons was highly variable by jurisdiction, reflective of jurisdiction-specific policies, with a median of 9 months enrollment before live birth delivery and 10 months enrollment postpartum. Jurisdictions that have not adopted Medicaid coverage via the Affordable Care Act have notably fewer adults represented. Application of the same data quality metrics to the two CDMs highlighted differences in measurement in the models, particularly in race and ethnicity capture and utilization stratified by care setting. The latter occurs because the two CDMs approach the concept of a “medical encounter” differently.

Conclusions: National U.S. Medicaid data is a heterogeneous but rich data source that requires jurisdiction-specific inclusion and exclusion criteria and cleaning to use it appropriately in pharmacoepidemiology studies.