Page duration improved offline report prediction for 22 new ROAMM readers, with each test story excluded

We tested an offline model of page reports in ROAMM ds007629 version 1.3.0. The source has 44 readers and 2,200 pages. A fixed split used 22 readers for fitting and 22 for testing. Each test story was excluded from its model. Adding log page duration to a task and word-count model reduced test log loss from 0.6971 to 0.4792 nats per page. The decrease was 0.2180 nats per page, or 31.3 percent. A 5,000-draw reader bootstrap refitted the models. Its 95 percent interval was 0.1251 to 0.2997 nats per page. Shorter first-pass intervals predicted more reports. A report can end that interval. Thus this result measures an offline report association. It does not show online attention detection or a neural cause. The package preserves source files, hashes, code, predictions, models, resampling records, and replay checks.

mind wandering · naturalistic reading · offline prediction · page duration · self-report · ROAMM · logistic regression · reader bootstrap