The latest word interval did not reduce error for observed next intervals in a fixed linear model for 135 test people in ds004789

We tested whether the latest word interval adds prediction value beyond prior recall pace. We used the lowest numbered FR1 session for each of 273 people in OpenNeuro ds004789 version 3.1.0. A fixed hash rule separated training people from test people. All features used the available response history. For 135 test people, the fixed linear model had mean squared log-interval error of 0.782934 without the latest-interval feature and 0.783016 with it. The relative error reduction was -0.010 percent, with a 95 percent interval of [-0.673, 0.295] percent. The interval used 2,000 resamples of training and test people, with model refits and fixed training scales. A second model predicted a word within five seconds and included terminal responses. Its relative Brier error reduction was 0.344 percent, with interval [-0.053, 0.794] percent. Restricted lists without vocalizations or with only correct, unique words gave larger descriptive gains. The result applies to these fixed models and this participant split. The package includes 551 checked source files, code, provenance, prediction rows, model scores, bootstrap draws and reproduction instructions. All 17 analysis outputs matched exactly in complete reruns.

free recall · memory retrieval · interresponse intervals · prediction · ridge regression · OpenNeuro