Department of
presents
Dr. Daniel McCaffrey
daniel_mccaffrey at rand dot org
ABSTRACT
The growing availability of longitudinal student test score data is increasing the interest among policy makers, researchers and educators in the use of this data to monitor the performance of teachers and schools for accountability and other purposes. One approach taken to using the data for the purpose of measuring teacher performance is so-called ``value-added modeling'' where complex multilevel statistical models are fit to the data and teacher performance is by measured fixed or random effects at the teacher level. However, the data lack a simple nesting structure because students are sequentially grouped over time in different classes. Models for such data that allow score to depend on both current and past teacher are commonly referred to as multi-membership models. A challenge of such models is the specification of the effects of teachers from earlier years on subsequent year student outcomes. Traditionally models have assumed that teacher effect persist undiminished and unchanged for all future years of testing. We propose an alternative model that allows teacher effects to dampen over time through persistence parameters that can be estimated from the data. We fit the model using the Bayesian framework to data from a large urban school district. We find that the persistence parameters are significantly less than one, the value assumed by the a models that forces teacher effects to be undiminished over time, our model better fits the data than the alternative, and we find that the model for the persistence of effects has an impact on inferences made about teachers. We will also discuss the potential bias that can result from misspecification of the persistence parameters.
Friday, September 22, 2006
3:35 - 4:35 pm
206 Cox Hall
Refreshments will be served on the second floor of Dabney Hall (left of Room 222) at 3:00 pm.