Improving Context Scale Interpretation Using Latent Class Analysis for Cut Scores

Abstract

This paper introduces an approach that uses latent class analysis to identify cut scores (LCA-CS) and categorize respondents based on context scales derived from large-scale assessments like PIRLS, TIMSS, and NAEP. Context scales use Likert scale items to measure latent constructs of interest and classify respondents into meaningful ordered categories based on their response data. Unlike conventional methods reliant on human judgments to define cut points based on item content, model-based approaches such as LCA find statistically optimal groups, a categorical latent variable, that explains item score differences based on score distribution differences between latent classes. Cut scores for these classes are determined by conditional probability calculations that relate class membership to observed scores, finding the intersection point of adjacent smoothed probability distributions and connecting it to the construct. Demonstrated through application to PIRLS 2021 data, this is useful to validate existing categorizations of the context scale by human experts, and can also help to enhance classification accuracy, particularly for scales exhibiting highly skewed distributions across diverse countries. Recommendations for researchers to adopt this LCA-CS approach are provided, demonstrating its efficiency and objectivity compared to judgment-based methods.