The 95 percent resampling interval included zero for gain from position terms in 12 ds008495 v1.0.2 test people under fixed prediction models

We tested whether option position improved prediction of social relationship judgments after video identity. We used OpenNeuro ds008495 v1.0.2. Thirteen of the 25 people included by the source trained two fixed models. Twelve other people tested them on the same 50 videos. The model with position terms had 0.00352 nats more mean log loss per valid judgment. Gain means baseline loss minus added model loss. Its value was -0.00352 nats. We resampled whole people 2,000 times and fitted the models again. The 95 percent interval was [-0.01930, 0.00350] nats. The interval assumes independent people and holds the video set fixed. An exact integer proof shows that the design can separate both position terms. This proof permits free video and person terms in the tested additive utility model. The package provides a reusable design audit tool. It includes all 36 selected source files, all fitted parameters, results, and checks. The prediction result does not establish a causal position effect or a zero effect.

social judgment · response position · multinomial logit · prediction · log loss · resampling · design identifiability · OpenNeuro