The Low Response Score (Lrs) A Metric To Locate, Predict, And Manage Hard-To-Survey Populations

In 2012, the US Census Bureau posed a challenge under the America COMPETES Act, an act designed to improve the competitiveness of the United States by investing in innovation through research and development. The Census Bureau contracted Kaggle. com to host and manage a worldwide competition to develop the best statistical model to predict 2010 Census mail return rates. The Census Bureau provided competitors with a block group-level database consisting of housing, demographic, and socioeconomic variables derived from the 2010 Census, five-year American Community Survey estimates, and 2010 Census operational data. The Census Bureau then challenged teams to use these data (and other publicly available data) to construct the models. One goal of the challenge was to leverage winning models as inputs to a new model-based hard-to-count (HTC) score, a metric to stratify and target geographic areas according to propensity to self-respond in sample surveys and censuses. All contest winners employed data-mining and machine-learning techniques to predict mail-return rates. This made the models relatively hard to interpret (when compared with the Census Bureau's original HTC score) and impossible to directly translate to a new HTC score. Nonetheless, the winning models contained insights toward building a new model-based score using variables from the database. This paper describes the original algorithm-based HTC score, insights gained from the Census Return Rate Challenge, and the model underlying a new HTC score.