# Provider Attribution in Medicare: Challenges and Solutions

## TL;DR

This paper improves Medicare data by identifying clinicians for patient encounters and comparing attribution methods across Medicare Advantage and Traditional Medicare.

## Contribution

The study introduces a hierarchical method for provider attribution and improves NPI identification in Medicare encounter data.

## Key findings

- NPI identification rates improved from 83% to 89% in Medicare Advantage from 2016 to 2022.
- Over 99% of medical encounters in both Medicare Advantage and Traditional Medicare had NPIs and specialty codes identified.
- A hierarchical attribution method showed the highest and most consistent performance across years.

## Abstract

To enhance National Provider Identifier (NPI) and specialty information available in Medicare Advantage (MA) encounter data and use the enhanced data to evaluate methods for retrospective attribution of the patient's usual clinician, comparing results across MA and Traditional Medicare (TM) populations.

We fill in missing clinician identifiers and specialty codes in MA encounter data using Centers for Medicare and Medicaid Services (CMS) and publicly available provider datasets. We attributed patients to the usual clinician using 16 methodological options, comparing the performance of these attribution methods in MA and TM.

We used a 20% sample of MA encounter data and TM claims data for 2016–2022, incorporating information from CMS's Medicare Data on Provider Practice and Specialty, archived data from the National Plan and Provider Enumeration System, and specialty‐taxonomy crosswalks derived from CMS publications.

For MA, we identified individual NPIs for 83% of medical claims in 2016, improving to 89% in 2022. Among MA medical claims billed by physicians and advanced practice providers, 95% of NPIs were for individual clinicians by 2022. In total, we identified individual or organization NPIs and specialty codes for over 99% of medical encounters in both TM and MA in all years. Rates of patient attribution were stable over time, and most methods had similar performance in MA and TM. We recommend a hierarchical attribution method that resulted in the highest fraction attributed with good consistency of attributed clinician year over year. Published reference files and SAS code make these NPI identification and patient attribution methods accessible.

Our methods allow researchers to identify provider NPIs that can be matched to external clinician data, used to attribute patients to a usual source of care, or to fit clinician fixed effects in studies of MA and TM.