Describing and assessing a new method of approximating categorical individual‐level income using community‐level income from the census (weighting by income probabilities)

OBJECTIVE: To assess a new approach (weighting by “income probabilities [IP]”) that uses US Census data from the patients' communities to approximate individual‐level income, an important but often missing variable in health services research. DATA SOURCES: Community (census tract level) income data came from the 2017 5‐year American Community Survey (ACS). The patient data included those diagnosed with cancer in 2017 in Ohio (n = 65,759). The reference population was the 2017 5‐year ACS Public Use Microdata Sample (n = 564,357 generalizing to 11,288,350 Ohioans). STUDY DESIGN/METHODS: We applied the traditional approach of income approximation using median census tract income along with two IP based approaches to estimate the proportions in the patient data with incomes of 0%–149%, 150%–299%, 300%–499%, and 500%+ of the federal poverty level (FPL) (“class‐relevant income grouping”) or 0%–138%, 139%–249%, 250%–399%, and 400%+ FPL (“policy‐relevant income grouping”). These estimated income distributions were then compared with the known income distributions of the reference population. DATA COLLECTION/EXTRACTION METHODS: The patient data came from Ohio's cancer registry. The other data were publicly available. PRINCIPAL FINDINGS: Both IP based approaches consistently outperformed the traditional approach overall and in subgroup analyses, as measured by the weighted average absolute percentage point differences between the proportions of each of the income categories of the reference population and the estimated proportions generated by the income approximation approaches (“average percent difference,” or APD). The smallest APD for an IP based method, 0.5%, was seen in non‐Hispanic White females in the class‐relevant income grouping (compared with 16.5% for the conventional method), while the largest APD, 7.1%, was seen in non‐Hispanic Black females in the policy‐relevant income grouping (compared with 18.0% for the conventional method). CONCLUSIONS: Weighting by IP substantially outperformed the conventional approach of estimating the distribution of incomes in patient data.