Bayesian model averaging for mathematics achievement growth rate trends

Abstract

In this study, we investigated the use of Bayesian model averaging (BMA) for latent growth curve models. We used the Trends in International Mathematics and Science Study (TIMSS) to predict growth rates in 8th-grade students' mathematics achievement. The dataset on male and female students' mathematics achievement contained 6 predictors, meaning that 64 model combinations were generated. Results highlighted science achievement score and teaching years as the most important predictors of both male and female students' growth in mathematics achievement. In this study, the growth rate of mathematical achievement for each country was compared with the predicted density and the density based on actual data. Most countries did not differ significantly in observed and predicted growth rates for male and female groups. For sensitivity analysis, the model prior had the smallest log-predictive score (LPS) value when specified as a binomial model with m = 4 for both the male and female data groups, regardless of the parameter prior. When comparing the fixed prior and flexible prior for parameters, the LPS value was relatively small when the fixed prior for parameters was set regardless of the model prior.