Using 311 data to develop an algorithm to identify urban blight for public health improvement

The growth of administrative data made available publicly, often in near-real time, offers new opportunities for monitoring conditions that impact community health. Urban blight—manifestations of adverse social processes in the urban environment, including physical disorder, decay, and loss of anchor institutions—comprises many conditions considered to negatively affect the health of communities. However, measurement strategies for urban blight have been complicated by lack of uniform data, often requiring expensive street audits or the use of proxy measures that cannot represent the multifaceted nature of blight. This paper evaluates how publicly available data from New York City’s 311-call system can be used in a natural language processing approach to represent urban blight across the city with greater geographic and temporal precision. We found that our urban blight algorithm, which includes counts of keywords (‘tokens’), resulted in sensitivity ~90% and specificity between 55% and 76%, depending on other covariates in the model. The percent of 311 calls that were ‘blight related’ at the census tract level were correlated with the most common proxy measure for blight: short, medium, and long-term vacancy rates for commercial and residential buildings. We found the strongest association with long-term (>1 year) commercial vacancies (Pearson’s correlation coefficient = 0.16, p