Developing and Validating a Measure to Estimate Poverty in Medicare Administrative Data

To develop and validate a measure that estimates individual level poverty in Medicare administrative data that can be used in studies of Medicare claims.

A 2008 to 2013 Medicare Current Beneficiary Survey linked to 2008 to 2013 Medicare fee-for-service beneficiary summary file and census data.

We used the Medicare Current Beneficiary Survey to define individual level poverty status and linked to Medicare administrative data (N=38,053). We partitioned data into a measure derivation dataset and a validation dataset. In the derivation data, we used a logistic model to regress poverty status on measures of dual eligible status, part D low-income subsidy, and demographic and administrative data, and modeled with and without linked census and nursing home data. Each beneficiary receives a predicted poverty score from the model. Performance was evaluated in derivation and validation data and compared with other measures used in the literature. We present a measure for income-only poverty as well as one for income and asset poverty.

A score (predicted probability of income poverty) >0.5 yielded 58% sensitivity, 94% specificity, and 84% positive predictive value in the derivation data; our score yielded very similar results in the validation data. The model's c-statistic was 0.84. Our poverty score performed better than Medicaid enrollment, high zip code poverty, and zip code median income. The income and asset version performed similarly well.

A poverty score can be calculated using Medicare administrative data for use as a continuous or binary measure. This measure can improve researchers' ability to identify poverty in Medicare administrative data.