experiments-database/analysis/matlab

tDCS reaching study — MATLAB GLM port

This directory is a MATLAB port of the Python tDCS GLM analysis (analysis/tdcs_glm.py, documented in analysis/METHODS.md), driven by raw app-export matrices rather than the pre-joined tdcs_reach_data.csv. It reproduces the same three models (count level, rate level, learning rate) and the same H2 anchor check / unknown-condition classification, but replaces the Python's population-average GEE with a fitglme GLMM (Poisson/Binomial, subject random intercept) — see "GLMM vs GEE" below.

Requirements

• MATLAB R2025b + Statistics and Machine Learning Toolbox.

• A license activated at ~/matlab/licenses/license.dat (already in place in this environment). If MATLAB ever reports it is unlicensed, activate with:

  ~/matlab/bin/activate_matlab.sh
  

Layout

• tdcs_config.m — curated 12-animal group map, reference group, merge maps.
• tdcs_load_data.m — parses raw/raw_success.csv and raw/raw_total.csv (app-export matrix format), joins them, and applies the curated group map.
• tdcs_scenario_data.m — builds one of the 6 scenarios (merge map + day window), and saves the scenario table to derived/<scenario>.csv.
• tdcs_models.m — fits the count/rate/learning GLMEs, the per-animal Welch/Mann-Whitney tests, the H2 anchor check, and (unmerged scenarios only) the unknown-condition classification.
• tdcs_report.m — prints descriptives, the raw GLME fixed-effects tables, and an interpretation section (H2 verdict, per-animal p-values, classification, caveats); also writes results/<scenario>.txt.
• tdcs_glm.m — the entry point: tdcs_glm(scenario) runs tdcs_scenario_data → tdcs_models → tdcs_report for one named scenario, via an explicit switch/case over all 6 scenario names.
• run_all.m — runs all 6 scenarios in one script.
• run_tests.m — runs the matlab.unittest suite in tests/ and exits nonzero on any failure.
• tests/ — matlab.unittest test classes (tLoadData, tScenarioData, tModels) with golden values cross-checked against the Python.
• raw/ — pre-generated app-export CSVs (input; not modified by this code).
• derived/ — scenario tables written by tdcs_scenario_data (generated).
• results/ — report text files written by tdcs_report/tdcs_glm (generated).

Running

Run one scenario (valid names: unmerged_full, unmerged_d0_10, mergeA2_full, mergeA2_d0_10, mergeB2_full, mergeB2_d0_10):

cd analysis/matlab
~/matlab/bin/matlab -batch "addpath(pwd); tdcs_glm('mergeB2_full')"

This prints the full report to the console and writes it to results/mergeB2_full.txt.

Run all 6 scenarios:

~/matlab/bin/matlab -batch "addpath(pwd); run_all"

Run the test suite:

~/matlab/bin/matlab -batch "addpath(pwd); run_tests"

(run_tests calls error(...) on any failure, so the matlab -batch process exits with a nonzero status — suitable for CI.)

Validation against the Python

Per-animal statistics (per-subject Welch t-test / one-sided Mann-Whitney U, Box-B2 vs Box-A2) are pure stats independent of the GLM choice, and were matched to the Python's golden values to 3 decimals across all 6 scenarios (see docs/superpowers/plans/2026-07-20-matlab-tdcs-analysis.md, Task 3).

GLMM vs GEE (read this before comparing numbers to the Python)

The Python reference uses GEE (population-average effects, exchangeable working correlation, robust/sandwich SEs) for the count and learning-rate models, and a cluster-robust Binomial GLM for the rate model. This MATLAB port instead uses GLMM (fitglme, subject-specific effects via a (1|subject) random intercept, Laplace/PL approximation) for all three models, because MATLAB's Statistics and Machine Learning Toolbox has no GEE implementation.

GLMM and GEE estimate related but distinct quantities (subject-specific vs. population-average effects), so exact coefficients, ratios, and p-values will differ between the two implementations — sometimes by a fair amount for the small, unbalanced groups in this study. What should agree is the interpretation: the direction of the H2 anchor effect, which unknown condition classifies as A2-like/B2-like/ambiguous, and whether the learning rates differ across groups. Treat the MATLAB numbers as a sensitivity check on the Python's conclusions, not as a numerically-identical reproduction.

Scope vs. the Python

The MATLAB port reproduces the Python's three core models (count, rate, learning-rate), per-animal tests, unknown-group classification, and the H2 verdict (read off the main model's A2-vs-B2 term). The Python's two auxiliary sensitivity sections — report_anchor_only (a two-anchor-only refit) and report_mixed (a Bayesian MAP random-intercept model) — are intentionally not ported: the anchor-only refit is closely tracked by the main-model H2 term, and MATLAB's native fitglme already is the random-intercept model.

Phased learning analysis and power

• Phased days × tDCS LME — the same success ~ day*tDCS + (1|subject) model refit within learning phases (0–5, 6–10, 6–13), Box-B2 vs Box-A2, with per-phase slopes and the interaction 95% CI:

  matlab -batch "tdcs_glm('phase_mergeA2')"   # or phase_unmerged / phase_mergeB2
  

  The Box-B2 faster-acquisition signal sits in the early (0–5) phase; late phases converge. The interaction p/CI use fitlme observation-level DF and are anticonservative at these small n (see power analysis).

• Monte-Carlo power — power to detect the early-phase interaction, scored with the cluster-honest per-animal test and the LME test, across subjects-per-group and effect sizes:

  matlab -batch "run_power"          # default: mergeA2, phase 0–5
  

  Result: at our n (3–5/group) honest power is ~0.3–0.7 even at the observed effect; ~8/group reaches ~80–90% if the effect is as large as observed, ~20/group if it is half that. The LME test is anticonservative only at small n and converges to the honest test by n≈12.

Figure and data export

• Figure — two-panel learning figure (curves + per-phase slopes, Box-B2 vs Box-A2), written to results/figure_learning.png:

  matlab -batch "make_figure"
  

• Export all data variations to CSV — organized under export/:

  matlab -batch "run_export"
  

  Produces curated_all.csv, scenarios/<merge>_<window>.csv (9), anchors/<merge>_B2vsA2.csv (3, with the binary tDCS factor), and phases/<merge>_phase_<lo>-<hi>.csv (9), plus a MANIFEST.

Paper replication (verbatim formula)

The paper's model, fitlme(tbl, 'behavior ~ stim + day + stim:day + (1|rat)'), is replicated verbatim (behavior = successful reaches, stim = tDCS [Box-B2=1 vs Box-A2=0], rat = subject) over the full training range:

matlab -batch "tdcs_glm('paper_mergeA2')"   # or paper_unmerged / paper_mergeB2

paper_mergeA2 reproduces the paper's interaction (F(1)=7.09 vs 7.12, p=0.009 vs 0.008); it also warns that the full-range interaction is confounded by unequal day coverage (Box-A2 ends ~day 13) — see the _d0_13 and phased analyses. The Monte-Carlo power analysis (run_power, tdcs_power_sim) uses this same behavior ~ stim + day + stim:day + (1|rat) model.

mergeNaive: pooled control vs tDCS

mergeNaive pools Naive into the control side (control = Box-A2 + Naive, n=7) and Right-Electrode into the tDCS side (tDCS = Box-B2 + Right-Electrode, n=4); Box-A is left out of the anchor contrast. Run the GLMMs, the paper LME, and power:

matlab -batch "tdcs_glm('mergeNaive_full')"    # GLMM (rate/count) tDCS vs control
matlab -batch "tdcs_glm('paper_mergeNaive')"   # paper LME: behavior ~ stim+day+stim:day+(1|rat)
matlab -batch "tdcs_power_sim('mergeNaive',[0 5])"

With the larger pooled control this is the closest replication of the paper's full pattern: a significant days x tDCS interaction (F(1)=7.30, p=0.008 vs paper 7.12/0.008), an equal-on-Day-1 stim main effect (p=0.21, matching the paper's n.s.), and a strong day effect. The overall tDCS advantage is significant even at the cluster-honest per-animal level (rate p=0.003). Caveat: pooling Naive (no-stimulation control) with Box-A2 (an electrode condition) is a modeling choice.