Model Estimation when Observations are not Independent:

Abstract

Parameter estimation and the computation of standard errors in social science models often require the assumption that observations are independent. This assumption is frequently violated with pooled cross-section and time-series data and household survey data. A recent article by Liang and Zeger (1986) shows that classical estimation methods retain good statistical properties in a wide variety of analyses where observations are not independent, and that correct standard errors of estimated model parameters are not difficult to compute. This article describes one of Liang and Zeger's results and presents its applications to the estimation of linear and logistic regression models from pooled cross-section and time-series data.

Rate of forgetting is independent from initial degree of learning across different age groups

“…However, at the item level, their dependent variable is a binomial outcome (1 correct, 0 incorrect). Generalised mixed effects models are recommended when dealing with binomial data ( Bye & Riley, 1989 ) as they can account for the multi-level structure of the data ( Quené & van den Bergh, 2004 ). We followed this recommendation for the analysis reported here.…”