The accuracy of harmonized tract time series by model, year and characteristic

Keywords: census tracts, areal interpolation, spatio-temporal analysis, neighborhood change
Abstract Type: Paper Abstract

Abstract

Studies of neighborhood change often use census tract data, but recurring updates to tract definitions can greatly complicate the measurement of population change within tracts. This has led secondary data providers to produce harmonized tract time series, which describe characteristics from multiple years for a single year’s tracts (e.g., 1990–2020 data for 2010 tracts). These products are now widely used, but their accuracy has remained largely undetermined. To achieve a broad assessment, we first computed benchmark statistics for 1990 and 2000 characteristics of 2010 census tracts by tabulating from geocoded restricted-access census microdata. We then used the benchmarks to compute errors in public estimates of fourteen characteristics (eight counts and six proportions) from five distinct estimation models, including those used by the Neighborhood Change Database (NCDB), Longitudinal Tract Database (LTDB), and National Historical Geographic Information System (NHGIS). We find many variations in accuracy across models, years, and characteristics. The largest differences correspond to the model basis: models based on block data (as used by NHGIS) are exceptionally more accurate than models based on tract data (as used by NCDB and LTDB). The differences in accuracy between two tested block-based models are much smaller, but the NHGIS model is still generally more accurate than a simpler areal-weighting approach. We recommend using block-based series when possible and using caution otherwise.