Application of multilevel IRT modeling to the study of self-esteem in adolescents

Abstract

The relationships between predictor variables and item response theory (IRT) latent trait estimates has traditionally been explored using a two-step approach of first, estimating IRT parameters and second, using the resulting latent trait estimates as dependent variables in a regression analysis. Recently, it has been shown that multilevel IRT modeling can be used to estimate the relationship between latent traits and predictor variables more accurately than the two-step approach (Adams, Wilson, & Wu, 1997). Kamata’s (1999; 2001) three-level IRT model expands the traditional two-level IRT model to situations in which person are nested within some setting, such as a classroom or school. Formulation of the model in this manner allows for the study of complex relationships between latent trait estimates and predictor variables in a multilevel context. While in the past ten years there has been a large increase in the use of multilevel models in psychological and educational research, multilevel IRT models have rarely been utilized in applied research. The lack of applied research with multilevel IRT modeling motivated the use of Kamata’s multilevel IRT model (1999; 2001) to study self-esteem in adolescents. The present study used data collected during the norming of the CFSEI-3 (Battle, 2002) from 905 adolescents in 13 different sites. Specifically, this study examined the relationships between academic self-esteem, demographic characteristics of the adolescent and characteristics that describe the social context of the adolescent. The same relationships were examined using general selfesteem. The results support gender and age differences in academic self-esteem and a gender by age interaction in general self-esteem. The variation in the nature of such relationships across the 13 sites encourages the use of multilevel modeling in the study of adolescent self-esteem. As well, the differences in the relationships of demographic variables with academic and general self-esteem encourages the study of both the global and specific facets of self-esteem. Estimation problems were encountered with the more complex models and those models studying ethnic differences in self-esteem. The causes of such estimation problems are discussed. As well, the advantages and potential problems in using multilevel IRT models in applied research are discussed.