Adding the previous first interval reduced mean squared log error by 2.55% in fixed linear models that omitted each run for 23 adults from ds006720

We tested whether the previous first response interval improved prediction of the current first interval in OpenNeuro ds006720 version 1.0.0. Fixed linear models used current and previous condition labels and linear progress terms. Each test omitted one run from model fitting. The analysis used 6,236 adjacent successful trial pairs from 23 adults. Each person had eight test runs. Mean squared error for logged durations fell from 0.085581 to 0.083399. The reduction was 2.55%. Mean person gain was 0.002181, with a 95% person bootstrap interval of 0.001134 to 0.003409. Nineteen people had positive gain. The primary models used runs before and after each test run. A planned test that trained on runs 1 to 4 and predicted runs 5 to 8 had uncertain gain. Thus, the result does not establish useful prediction of later runs. Residual correlations also depended on trend adjustment. An added simulation showed negative residual correlation from overlapping pairs and nuisance fitting under independent Gaussian trial noise. This check was selected after the primary results. It was not used to correct the real data. The package preserves the local plan, 189 source files, hashes, all trial and pair decisions, coefficients, predictions, random draws, simulations, figures, and code. A fresh source download passed all hashes. Replay matched all 43 numerical output files exactly. The study does not identify a causal history effect or a brain mechanism.

serial timing · response history · duration reproduction · prediction error · cross-validation · residual correlation · person bootstrap · OpenNeuro