Content area
Full text
Abstract: Multilevel modeling is a flexible alternative to the traditional factorial ANOVA approach in the analysis of experimental data with repeated measures. This article describes a psycholinguistic experiment and provides a detailed account of the data analysis, demonstrating the use of multilevel models to include a continuous predictor and complex assumptions about error variance. The experiment investigated the effects of structural priming on reaction times in a word monitoring task. Pairs of sentences with identical or different syntactic structures were presented to 4- and 5-year-old children, whose task was to respond to a word presented in the second sentence. Multilevel modeling analysis revealed an interaction between the experimental condition and position of the trial within the experiment: the reaction times in the same-structure condition decreased over the course of the experiment, while they increased in the different-structure condition. The analysis demonstrates how multilevel models can be used to detect change in responses over the course of an experimental session.
Key words: syntactic priming, word monitoring, multilevel models, reaction time
(ProQuest: ... denotes formulae omitted.)
Psycholinguistic experiments that use reaction times as the dependent variable usually compare mean reaction times in different experimental conditions using repeated-measures ANOVA. This method, based on the least squares estimation, has a number of limiting assumptions. One particular limitation concerns the extension of ANOVA that includes continuous predictors, i.e. ANCOVA. This method can only provide meaningful results if the effects of the continuous covariate are the same across the categorical conditions. This is one of the reasons why the use of continuous independent variables is infrequent in the analysis of experimental data and is often replaced by converting the continuous variable to a categorical factor with levels such as high vs. low score on the continuous variable. This type of conversion, however, leads to a loss of information and is strongly discouraged on methodological grounds (Cohen, 1983; MacCallum et al., 2002). An alternative method of data analysis is the use of mixed models, also known as hierarchical linear models or random coefficient models (Bryk, Raudenbush, 1992; Pinheiro, Bates, 2000). This article presents a detailed account of a multilevel analysis of an experimental study with reaction time as the dependent variable. Besides the inclusion of a continuous independent variable in...