Exposing Undercounts in the Census through Regression Modelling

Although many community leaders have proposed that language barriers pose significant obstacles to Census outreach, this paper explores the viability of using predictive models to quantify the extent the role language plays.

Undercounts in the Census are notoriously difficult to measure, but are necessary to understanding the extent of systemic divides within America. Although many community leaders have proposed that language barriers pose significant obstacles to Census outreach, this paper explores the viability of using predictive models to quantify the extent the role language plays. By using a multivariable regression model trained on student data from the Civil Rights Data Collection, we concluded that Limited English Proficient (LEP) Asian populations were undercounted by over 98,000 people, LEP Native Hawaiian/Pacific Islander populations were undercounted by over 73,000 people, and LEP White populations were undercounted by over 164,000 people. However, biases in the dataset made the results for other ethnic groups unreliable, indicating that regression modeling should be used as a tool for identifying areas of improvement in the Census rather than producing an exact estimate of the population. Our data suggests a strong correlation between language proficiency and Census undercounts, and the Census Bureau could ideally use the results to create targeted language outreach programs for specific states.

Explore More!

Research Paper

Source Code