A recall lag 2 GloVe term increased mean test loss in 68 selected ds004789 people at fixed ridge 0.001

We tested a word measure from the second most recent correct recall. We asked whether it improved the next correct recall prediction. We used public events from OpenNeuro ds004789 version 3.1.0. We fit the models in 67 people and tested them in 68 other people. Both models had terms for study lag from both prior words. They also had terms for study position and word identity. Both used similarity from the most recent word. Adding the earlier similarity term gave a paired gain of -0.005376 nats. A positive gain means lower loss. The approximate 95% interval from 2,000 resamples with new fits was [-0.013447, 0.000062] nats. The reference model had lower mean test loss. The interval included zero. This result applies to the fixed word pool, selected people, model terms, and ridge penalty. The package contains source records, every resample, predictions, code, and checks. It does not identify a brain mechanism.

free recall · memory · GloVe · cosine similarity · prediction · ridge regression · OpenNeuro