The Impact of Dichotomization on Network Recovery

Network psychometrics has become an important method for studying the network structure of multivariate psychological data. Accurate recovery of the underlying network structure is paramount and requires that the models are appropriate for the data at hand. Traditionally, Gaussian graphical models for continuous data and Ising models for binary data have dominated the literature. However, psychological research often relies on ordinal data from Likert scale items, creating a model-data mismatch. This paper examines the effect of dichotomizing ordinal variables on network recovery using Bayesian analysis of an ordinal Markov random field model implemented in the R package bgms. Our analysis shows that dichotomization can lead to a loss of information that affects the accuracy of network recovery, especially when considering the interplay between the binarization cutoffs used and the distribution of the ordinal categories. In addition, we show that there is a difference in accuracy when using dichotomized data depending on whether edges are present or absent in the true network, highlighting the effectiveness of using the ordinal model in revealing conditional independence relationships. These findings underscore the importance of using models that deal directly with ordinal data to ensure more reliable and valid inferred network structures in psychological research.