Understanding the structure of data when planning for analysis: application of hierarchical linear model

Abstract

A study in which achievement test scores are collected from a sample of learners nested within classrooms that are in turn nested within schools has data structure that is hierarchical. This is because each learner belongs to one and only one classroom and each classroom belongs to one and only one school. There are three levels of random variation in such data: variation among learners within classrooms, variation among classrooms within schools, and variation among schools. Despite the prevalence of hierarchical structures in behavioural and social research, social scientists often fail to address them adequately in the data analysis phase. This neglect has reflected limitations in conventional statistical techniques for the estimation of linear models with nested structures. In social sciences research, these limitations have generated concerns about aggregation bias, estimation precision, and the 'unit of analysis' problem.