The Low Response Score (LRS): A Metric to Locate, Predict, and Manage Hard-to-Survey Populations
The Low Response Score (LRS): A Metric to Locate, Predict, and Manage Hard-to-Survey Populations (DOI: https://doi.org/10.1093/poq/nfw040) was written by Chandra Erdman and Nancy Bates in 2017. It was published in Public Opinion Quarterly (vol. 81, no. 1).
This paper introduces the LRS, a model-based score that is predictive of difficulty to survey. This is an improvement upon the HTC (hard-to-count) score created in 2001.
The LRS comes from a crowdsourced challenge administered by Kaggle.com. Model submissions are evaluated with respect to population-weighted MSE and ranked continuously.
- 244 participants
top 3 used machine learning methods, specifically gradient boosting or random forests
were able to use non-PDB covariates, but the most predictive ones are available for the Census Bureau to use
- "the single most influential predictor in the winning model was the percentage of renter households in a block group"
The most influential applicable covariates identified through the competition were then used as predictors in a regression model. The LRS is the model-based probability of nonresponse at the block group and tract level.
The authors then examine the LRS in light of three comparable D.C. block groups: Columbia Heights, Trinidad, and Anacostia.