"Predicting" after peeking into the future: Correcting a fundamental flaw in the SAOM -- TERGM comparison of Leifeld and Cranmer (2019)

TL;DR

This paper critiques a previous comparison of SAOMs and TERGMs, revealing that flawed model specifications led to invalid conclusions, emphasizing the importance of correct model specification in network analysis.

Contribution

It identifies and corrects a fundamental flaw in the prior comparison, clarifying the proper use of covariates and model specifications in TERGMs and SAOMs.

Findings

Previous comparison used circular covariates, invalidating results.

Corrected model specifications show different conclusions.

Researchers should disregard flawed recommendations from prior study.

Abstract

We review the empirical comparison of SAOMs and TERGMs by Leifeld and Cranmer (2019) in Network Science. We note that their model specification uses nodal covariates calculated from observed degrees instead of using structural effects, thus turning endogeneity into circularity. In consequence, their out-of-sample predictions using TERGMs are based on out-of-sample information and thereby predict the future using observations from the future. We conclude that their analysis rest on erroneous model specifications that render the article's conclusions meaningless. Consequently, researchers should disregard recommendations from the criticized paper when making informed modelling choices.