Reanalysis of “Historical redlining is associated with increasing geographical disparities in bird biodiversity sampling in the United States”

Martin Bulla , Peter Mikula

Ellis-Soto et al. (2023, Nature Human Behaviour) investigated whether the density and completeness of bird biodiversity sampling from citizen science observations across US cities covary with 1930s neighbourhood classifications based on perceived mortgage investment risk, a practice known as “redlining”. They claimed that worst-rated neighbourhoods were the most under-sampled urban areas for bird biodiversity and that such disparity in sampling increased between 2000 and 2020 by 35.6%.

We were initially unable to reproduce the data generation and analyses with the deposited code and data, but reproduced most reported findings after the authors provided missing data and we corrected several coding issues. However, the code underlying the spatial results (Fig. 1) and the key claim of a 35.6% temporal increase were absent. After correcting a major data-coding error (unintentional data multiplication), we recreated the temporal trends, including Fig. 4, in a manner consistent with the underlying true yearly data. We also demonstrate that Fig. 4 and the original claim of 35.6% increase arise from annual aggregates of sampling density that implicitly treat thousands of polygons as a single annual observation, precluding any modelling of spatial or temporal non-independence.

Despite model misspecification and annual data aggregation in the original analyses, our alternative analytical choices (a) reproduced the reported disparity among HOLC grades with effect sizes that differ in magnitude but are consistent in direction, and (b) revealed substantially more complex temporal trends. Specifically, the relative disparity between the worst- and best-rated neighbourhoods varied non-linearly over time and exceeded 350% by 2020. In contrast, absolute differences remain negligible until ~2010, rise to only ~5 observations per km² per year by ~2017, and reach at most ~25 observations by 2020 – small absolute magnitudes despite large relative changes.

To conclude, large relative disparities between the best- and worst-rated neighbourhoods do not imply large absolute differences in sampling density. Overall, disparity changes exceed ~200% between 2000 and 2020, which is inconsistent with the original authors’ claim of a 35.6% increase. The post-2010 rise coincides with the surge in smartphone-driven community science. The plateau around 2015 plausibly reflects broader smartphone accessibility. The sharp rise in 2020 aligns with COVID-19 restrictions, which markedly boosted urban green-space use and citizen science participation.