Multidimensional visualization of large-scale survey data

Abstract

Large-scale surveys often contain multiple related response dimensions whose joint structure is better understood using multidimensional visualization. Traditional one-dimensional summaries of large-scale social science surveys are difficult to interpret. We propose a visualization framework that combines model-based estimation and two-dimensional heatmaps to reveal multidimensional patterns in survey responses. The framework can be paired with a variety of statistical or machine learning models and is designed to facilitate interpretation of complex survey data through subgroup decomposition, temporal comparison, and exploration of joint response structures. We apply the framework to national survey data from 2014 to 2021 and show that the resulting heatmaps reveal multidimensional structure that is difficult to detect from one-dimensional summaries alone, including clustering, polarization, and differences in issue alignment across policy pairs and population groups. The framework is particularly valuable when associations between policy dimensions are weak and their relationships cannot readily be inferred from one-dimensional summaries. The framework can be implemented with both conventional regression models and flexible machine-learning approaches, with the latter generally providing closer in-sample fit to the observed policy scores. More generally, the proposed framework provides a scalable and interpretable tool for studying interconnected attitudes in multidimensional survey data.