Code is available on GitHub:
https://github.com/TheGoldLab/ms_2026_mcgaughey_gold_elife

See code/README.m for the full project overview (dependencies, figure scripts, code/utilities/ reference) and code/builds/README.m for how the derived/cached files below are regenerated. This file documents only the contents of data/ itself, at the column level.

Contents

data/
  mergedTable_proc.mat             - canonical trial-wise table, all monkeys (9.5 GB loaded)
  mergedTable_proc_core.mat        - derived tier: behavioral columns only (~136 MB)
  mergedTable_proc_neural.mat      - derived tier: behavioral + neural, no pupil traces (~220 MB)
  normalizationTerm.mat            - per-unit firing-rate normalization terms
  sensitivity_diff_labeled_BP.mat  - per-session slope diff, behavior+pupil subset
  sensitivity_diff_labeled_N.mat   - per-session/unit slope diff, neural subset
  sensitivity_diff_labeled_NP.mat  - per-session/unit slope diff, neural+pupil subset
  behaviorFits/                    - pre-computed logistic-regression fits (see below)

mergedTable_proc.mat / mergedTable_proc_core.mat / mergedTable_proc_neural.mat

All three store the same MATLAB table under the variable name mergedTable_proc, one row per trial, pooled across monkeys An, Ch, and Mi. mergedTable_proc.mat is the canonical 80,314 x 32 source table; the other two are lighter derived tiers (dropping "heavy" columns) produced from it by code/builds/buildMergedTableTiers.m so that scripts that never touch pupil traces or spike times don't pay to load them. See code/README.m "data/" section for which figure script loads which tier, and re-run buildMergedTableTiers.m if mergedTable_proc.mat is ever regenerated — the two tiers are not meant to be hand-edited.

Each session/unit's row-grouping key (ses_ID) encodes which data types are available for it via a trailing code — B (behavioral), N (neural), P (pupil), e.g. "AnDR_20210614_Unit1_BNP" — and createDatSubset.m (code/utilities/) filters rows by that code (e.g. 'N', 'NP', 'BP').

Columns of the full table, in order (indices below match mergedTable_proc.mat; the core/neural tiers renumber after dropping columns, but variable names are unchanged):

Columns 24–26 (continuous pupil traces) account for 98.7% of the full table's in-memory size — that's why only the pupil-analysis scripts (Fig6_FigSupp1.m, Fig6_FigSupp2AB.m, Fig6_FigSupp2CD.m, behaviorLogisticFitsPupilTerm.m, processPupilBaseline.m, processPupilFull.m) load mergedTable_proc.mat directly; everything else uses one of the two lighter tiers.

normalizationTerm.mat

normalizationTerm — [numUnits x 1] double

Per-unit max-firing-rate normalization term computed from correct trials during initial neural processing (see processNeuralData.m / behaviorLogisticFitsNeuralTerm.m, where it's populated as normalizationTerm(uu,1) = max(maxRate)). Passed back into processNeuralData.m / processNeuralDataErrorTrials.m as normalizationTerm_input so that early-vs-late block-split analyses (Figure 4-Figure Supplement 1B) and error-trial analyses normalize firing rates consistently rather than recomputing the max per subset.

sensitivity_diff_labeled_{BP,N,NP}.mat

sensitivity_diff_labeled_<code> — [numRows x 2] cell array

• column 1: session/unit ID string (e.g. "AnDR_20210614_Unit1_BNP")
• column 2: double, difference in psychometric slope/sensitivity (β₂, time-dependent term) between conditions for that session/unit — extracted via cell2mat(...(:,2)) wherever these are loaded (e.g. Fig5_FigSupp3.m, Fig6_FigSupp2AB.m)

Suffix denotes which data-type subset (see createDatSubset.m codes) the rows are restricted to:

behaviorFits/

Pre-computed logistic-regression fits, generated by the four code/builds/behaviorLogisticFits*.m scripts (fitting is time-intensive, so these outputs are checked in rather than regenerated on each run). See code/builds/README.m for build provenance and validation notes. On Dryad, these files are in the top-level data directory (file names starting with Logistics..).

Behavior-only model (behaviorLogisticFits.m):

• LogisticFits.mat → fits — [161 sessions x 4 params x 2 hazards] double: β₀ switch/stay bias, β₁ right/left directional bias, β₂ sensitivity to viewing duration (time slope), β₃ lapse rate
• LogisticFits_Rsq.mat → R_sq — [161 sessions x 1 x 2 hazards] double, Tjur's pseudo-R²

Behavior-only, shuffled null (behaviorLogisticFitsShuffle.m; 100 shuffled refits per session, stacked as 100 repeats of the per-session block — tile any per-session mask with repmat(mask, 100, 1) to index):

• LogisticFits_Shuffle.mat → fits_shuffle — [16100 x 4 x 2]
• LogisticFits_Rsq_Shuffle.mat → R_sq_shuffle — [16100 x 1 x 2]

Behavior + trial-wise MT neural term (behaviorLogisticFitsNeuralTerm.m):

• LogisticFits_NeuralTerm.mat → fits_neuralTerm — [153 units x 1 x 2 hazards] double, the neural x signed-time interaction coefficient (β₂) only, i.e. fits(uu,3,:) from the underlying 5-parameter fit — not the full parameter vector
• LogisticFits_NeuralTerm_Rsq.mat → R_sq_neuralTerm — [153 x 1 x 2]
• LogisticFits_NeuralTerm_control.mat → fits_neuralTerm_control — [153 x 1 x 2] (PREF-direction-only, behavior-only control fit's signed-time term)
• LogisticFits_NeuralTerm_Rsq_control.mat → R_sq_neuralTerm_control — [153 x 1x 2]

Behavior + trial-wise evoked-pupil term (behaviorLogisticFitsPupilTerm.m):

• LogisticFits_PupilTerm.mat → fits_pupilTerm — [153 units x 1 x 2 hazards] double, the pupil x signed-time interaction coefficient (β₂) only, same convention as fits_neuralTerm above
• LogisticFits_PupilTerm_Rsq.mat → R_sq_pupilTerm — [153 x 1x 2]
• LogisticFits_PupilTerm_control.mat → fits_pupilTerm_control — [153 x 1 x 2]
• LogisticFits_PupilTerm_Rsq_control.mat → R_sq_pupilTerm_control — [153 x 1 x 2]

All hazard/frequency dimensions are indexed [LSF, HSF] (cfg.hazard.codes = [2, 6] from projectDefaults.m); all *_control files are the matched behavior-only control fit for that model's units, not a separate experiment.