<p>Abstract: Although the Self-Referent Encoding Task (SRET) is commonly used to measure self-referent cognition in depression, many different SRET metrics can be obtained. The current study used best subsets regression with cross-validation and independent test samples to identify the SRET metrics most reliably associated with depression symptoms in three large samples: a college student sample (<em>n</em> = 572), a sample of adults from Amazon Mechanical Turk (<em>n</em> = 293), and an adolescent sample from a school field study (<em>n</em> = 408). Across all three samples, SRET metrics associated most strongly with depression severity included number of words endorsed as self-descriptive and rate of accumulation of information required to decide whether adjectives were self-descriptive (i.e., drift rate). These metrics had strong intra-task and split-half reliability and high test-retest reliability across a 1-week period. Recall of SRET stimuli and traditional reaction time metrics were not robustly associated with depression severity.</p><p>This dataverse includes all data and code used for the paper. HTML files showing analyses (but no code) can be viewed on <a href="https://jdbest.github.io/sretmodels/" title="MDL gitpages website for SRET modeling">github at https://jdbest.github.io/sretmodels/</a>.</p>