Identifying Medicare Beneficiaries with Disabilities: Improving on Claims-Based Algorithms - Executive Summary

Yonatan Ben-Shalom and David Stapleton

Mathematica Policy Research

Objectives: (1) To assess the feasibility of using existing claims-based algorithms to identify community-dwelling Medicare beneficiaries with disability based solely on the conditions for which they are being treated; and (2) to improve upon these algorithms by combining them in predictive models. This capability is important for helping assess whether care coordination interventions and other programs designed to improve outcomes for a broad class of individuals are effective for the subset who have disabilities. Such information is typically unavailable in claims-based analyses.

Data Source: Data from the Medicare Current Beneficiary Survey (MCBS) and matched Medicare claims on 12,415 community-dwelling fee-for-service Medicare beneficiaries who completed the MCBS baseline questionnaire in 2003-2006.

Study Design: Using Medicare claims data, we created six potential indicators of disability and used the indicators in logistic regression models to predict three indicators of disability based on self-reports: limitation in at least three activities of daily living (ADLs), limitation in at least one ADL, and limitation in at least one ADL or instrumental activity of daily living. Using receiver operating characteristic curves, we compared the true positive (sensitivity) and true negative (specificity) rates of the individual indicators to those of the regression-based predictive models.

Principal Findings: The predictive performance of the regression-based models is better than that of the individual claims-based indicators, providing better sensitivity for any level of specificity and vice versa. For community-dwelling Medicare beneficiaries who are age 65 or older, we use a decision rule that classifies an individual as having a disability if his or her predicted probability of having a disability exceeds a specified value. This threshold value is that which maximizes the sum of sensitivity and specificity. Using this threshold, our model prediction is congruent with self-reports of disability for 72 percent of those who reported they had a limitation in at least one ADL and 65 percent of those who did not report having any ADL limitations. Examination of the incongruent cases showed that obese individuals who have a disability were especially likely to be missed by the model. For individuals ages 18-64, the model fails to identify nearly half who reported limitation in at least one ADL (sensitivity=0.54). For this group, dual-eligible beneficiaries with disabilities were especially likely to be missed by the model. In both age groups, the individuals without self-reported disabilities who are most likely to be misclassified as having disabilities were those in relatively good health (according to their self-reports) who had relatively high service use.

Conclusions: Predictive models, which may be tailored according to beneficiary subgroup and self-reported disability measure, provide a better sensitivity-specificity trade-off than individual claims-based disability flags and therefore provide an improved tool for researchers seeking to identify people with disabilities in claims data. Such models may be improved by incorporating data on service use (such as home health care and skilled nursing facilities) and prescriptions in general and Medicaid data for dual-eligible beneficiaries in particular (to capture use of long-term support services).