266 – SPEED: Statistical Computing and Sports, Part 2A
Effects of Missing Data on Student Growth Estimates
Katherine M. Wright
Northwestern University
John L. Gatta
Northwestern University
Terri D. Pigott
Loyola University Chicago
One property of student growth data that is often overlooked despite widespread prevalence is incomplete or missing observations. As students migrate in and out of school districts, opt out of standardized testing, or are absent on test days, there are many reasons student records are fractured. Missing data in growth models can bias model estimates and growth inferences. This paper presents empirical explorations of how well missing data methodologies recover attributes of would-be complete student data used for teacher evaluation. Missing data methods are compared in the context of a Student Growth Percentiles (SGP) model used by several school systems for accountability purposes. Using a real longitudinal dataset, we evaluate the sensitivity of growth estimates to missing data and compare the following missing data methods: listwise deletion, likelihood-based imputation using an expectation-maximization algorithm, multiple imputation using a Markov Chain Monte Carlo method, multiple imputation using a predictive mean matching method, and inverse probability weighting. Methodological and practical consequences of missing data are discussed.