The impact of dichotomization on network recovery.

Graphical models have 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, as opposed to analyzing the data at its original level of measurement, using a Bayesian analysis of the ordinal Markov random field model. This model is implemented in the R package bgms. Our analysis shows that dichotomization results in a loss of information, which affects the accuracy of network recovery. This is particularly true when considering the interplay between the dichotomization cutoffs used and the distribution of the ordinal categories. In addition, we demonstrate a difference in accuracy when using dichotomized data, depending on whether edges are included or excluded in the true network, which highlights the effectiveness of the ordinal model in recovering 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.