UCLA

Many educational websites offer tools that predict students' chances of admissions to college using models estimated on crowd-sourced data. Unfortunately, these data are generally not representative of the actual population of college applicants, so admission probability calculators (APCs) which use them offer biased estimates of admission probability, Pr(A). As a solution, I propose a simulation-based framework for estimating Pr(A) that is not modeled on crowd-sourced data. I model the joint distribution of enrollee characteristics at a particular university using a Gaussian copula and official margin data published by the College Board and the university itself. I fit a model for Pr(A) to this data based on several reasonable assumptions about the character of the applicant pool. I show that the estimates of Pr(A) yielded by my framework are excellent predictors of admission outcome for a sample of student data collected from the Internet.