A Cautionary Tail: A Framework and Case Study for Testing Predictive Model Validity

TL;DR

This paper introduces a framework for testing the validity of predictive models, highlighting the importance of field assessments to identify biases and ensure models perform well in real-world scenarios.

Contribution

It presents a novel field assessment framework for validating predictive models and demonstrates its application through a case study on rat infestation prediction in Washington, D.C.

Findings

Model performs well on new 311 data

Model fails to predict inspection outcomes accurately

Field assessments reveal biases not detected by traditional testing

Abstract

Data scientists frequently train predictive models on administrative data. However, the process that generates this data can bias predictive models, making it important to test models against their intended use. We provide a field assessment framework that we use to validate a model predicting rat infestations in Washington, D.C. The model was developed with data from the city's 311 service request system. Although the model performs well against new 311 data, we find that it does not perform well when predicting the outcomes of inspections in our field assessment. We recommend that data scientists expand the use of field assessments to test their models.