function L = tdcs_paper_lme(mergeKey, cfg, windowKey)
%TDCS_PAPER_LME Replicate the paper's linear mixed model, verbatim formula:
%       behavior ~ stim + day + stim:day + (1|rat)
%   fit with fitlme on the Box-B2 (stim = 1, tDCS) vs Box-A2 (stim = 0, control)
%   subset of the MERGEKEY grouping ('unmerged'|'mergeA2'|'mergeB2'|'mergeNaive').
%   `behavior` = successful reaches (COUNT), `day` = training day (raw; day 0 =
%   the paper's "Day 1"), `rat` = subject. Reports the three effects (stim x day
%   interaction, day, stim) as t(df) / F(1) / p with the interaction 95% CI,
%   alongside the paper's reference values, and writes a results/ txt file.
%
%   L = TDCS_PAPER_LME(MERGEKEY, CFG) uses the full training range (default) and
%   writes results/paper_<MERGEKEY>.txt.
%
%   L = TDCS_PAPER_LME(MERGEKEY, CFG, WINDOWKEY) restricts to a day window --
%   WINDOWKEY is 'full' (day <= 26), 'd0_10' (day <= 10), or 'd0_13' (day <= 13,
%   the last day Box-A2 has data). The windowed fits give the two anchor groups
%   EQUAL day coverage, so the stim:day interaction is not confounded by the
%   full-range coverage imbalance; windowed runs write
%   results/paper_<MERGEKEY>_<WINDOWKEY>.txt.
%
%   (This is the same model as tdcs_lme -- success ~ day*tDCS + (1|subject) --
%   written with the paper's exact term order and variable names.)
%
%   L fields: .model .nRats .nObs .windowKey and .stim/.day/.interaction effect
%   structs (.estimate .se .t .df .p from Coefficients; .F .df1 .df2 .Fp from
%   ANOVA; .ci = coefficient 95% CI).

if nargin < 3 || isempty(windowKey)
    windowKey = 'full';
end
if ~ismember(windowKey, {'full', 'd0_10', 'd0_13'})
    error('tdcs_paper_lme:badWindow', ...
        'windowKey must be ''full'', ''d0_10'', or ''d0_13'' (got "%s").', windowKey);
end

S = tdcs_scenario_data([mergeKey '_' windowKey]);
A = S(ismember(S.group, {cfg.anchorLow, cfg.anchorHigh}), :);

tbl = table();
tbl.behavior = A.success;
tbl.day = A.day;
tbl.stim = double(A.group == cfg.anchorHigh);   % Box-B2 = 1, Box-A2 = 0
tbl.rat = A.subject;

model = fitlme(tbl, 'behavior ~ stim + day + stim:day + (1|rat)');
C = model.Coefficients; An = anova(model); ci = coefCI(model);
Ans = anova(model, 'DFMethod', 'satterthwaite');   % Satterthwaite denominator DF

L.model = model;
L.windowKey = windowKey;
L.nRats = numel(unique(tbl.rat));
L.nObs = height(tbl);
L.maxDayCtrl = max(tbl.day(tbl.stim == 0));
L.maxDayStim = max(tbl.day(tbl.stim == 1));
L.stim = localTerm(C, An, Ans, ci, 'stim');
L.day = localTerm(C, An, Ans, ci, 'day');
L.interaction = localTerm(C, An, Ans, ci, 'day:stim');   % MATLAB canonicalizes stim:day -> day:stim
% Honest test: refit with a per-rat random SLOPE so the interaction DF
% collapses toward the animal count (see tdcs_random_slope_interaction).
L.interRS = tdcs_random_slope_interaction(tbl, ...
    'behavior ~ stim + day + stim:day + (day|rat)', 'day:stim');

localReport(L, mergeKey, cfg);

end

function e = localTerm(C, An, Ans, ci, name)
i = strcmp(C.Name, name);
if ~any(i)
    error('tdcs_paper_lme:missingTerm', 'No "%s" coefficient (have: %s).', ...
        name, strjoin(C.Name, ', '));
end
ai = strcmp(An.Term, name);
si = strcmp(Ans.Term, name);
e = struct('estimate', C.Estimate(i), 'se', C.SE(i), 't', C.tStat(i), ...
    'df', C.DF(i), 'p', C.pValue(i), 'F', An.FStat(ai), 'df1', An.DF1(ai), ...
    'df2', An.DF2(ai), 'Fp', An.pValue(ai), 'ci', ci(i, :), ...
    'dfSatt', Ans.DF2(si), 'pSatt', Ans.pValue(si));
end

function localReport(L, mergeKey, cfg)
window = L.windowKey;
if strcmp(window, 'full')
    scenarioName = ['paper_' mergeKey];
    windowLabel = 'full range';
else
    scenarioName = ['paper_' mergeKey '_' window];
    windowLabel = strrep(window, 'd0_10', 'days 0-10');
    windowLabel = strrep(windowLabel, 'd0_13', 'days 0-13');
end

bar = repmat('=', 1, 78);
s = sprintf('%s\n', bar);
s = [s sprintf('PAPER LME REPLICATION -- %s (%s)\n', mergeKey, windowLabel)];
s = [s sprintf('%s\n', bar)];
s = [s sprintf('model: behavior ~ stim + day + stim:day + (1|rat)   [stim: %s=1 vs %s=0]\n', ...
    cfg.anchorHigh, cfg.anchorLow)];
s = [s sprintf('behavior = successful reaches (COUNT per session)\n')];
s = [s sprintf('N = %d rats, %d sessions   (day raw; day 0 = paper "Day 1")\n', L.nRats, L.nObs)];
s = [s sprintf('day coverage: stim(B2) 0..%d, control(A2) 0..%d\n', L.maxDayStim, L.maxDayCtrl)];
if abs(L.maxDayStim - L.maxDayCtrl) > 2
    s = [s sprintf(['** WARNING: unequal day coverage -- the full-range stim:day interaction\n' ...
        '   extrapolates the control group''s line and is CONFOUNDED here (the paper''s\n' ...
        '   groups had equal coverage). See the _d0_13 fair-window and phased analyses. **\n'])];
else
    s = [s sprintf(['(Equal day coverage -- the stim:day interaction over this window is NOT\n' ...
        ' confounded by the full-range coverage imbalance.)\n'])];
end
s = [s sprintf('\n')];
s = [s tdcs_model_summary(L.model, 'fitlme: behavior ~ stim + day + stim:day + (1|rat)') sprintf('\n')];
s = [s sprintf('\n%-26s %-18s %-12s %s\n', 'effect', 't(df) / F(df1)', 'p (resid)', 'Satterthwaite: p (df)')];
s = [s sprintf('%s\n', repmat('-', 1, 76))];
s = [s localRow('stim x day (interaction)', L.interaction)];
s = [s localRow('day (learning)', L.day)];
s = [s localRow('stim (main, Day 1)', L.stim)];
s = [s sprintf('interaction 95%% CI: [%+.2f, %+.2f]\n', L.interaction.ci(1), L.interaction.ci(2))];
s = [s localHonest(L.interRS)];
s = [s sprintf('\nINTERPRETATION\n')];
if L.interaction.p >= 0.05
    interTxt = 'slopes parallel -- no differential learning rate over this window';
elseif L.interaction.estimate > 0
    interTxt = 'tDCS (Box-B2) improves FASTER -- benefit accumulates over training';
else
    interTxt = 'tDCS (Box-B2) improves SLOWER -- groups converge';
end
s = [s sprintf('  - stim x day interaction: %s (p=%.4f, slope diff=%+.2f) -> %s.\n', ...
    localSigTxt(L.interaction.p), L.interaction.p, L.interaction.estimate, interTxt)];
s = [s sprintf('  - stim main effect on Day 1 (our day 0): %s (p=%.4f) -> groups %s on Day 1.\n', ...
    localSigTxt(L.stim.p), L.stim.p, localPick(L.stim.p < 0.05, 'already DIFFER', 'are comparable'))];
s = [s sprintf(['\nPaper (N=24): interaction t(227)=2.68, F(1)=7.12, p=0.008; ' ...
    'day t(227)=9.64,\n  F(1)=267.64, p=1.2e-18; stim t(227)=0.23, F(1)=0.053, p=0.81.\n'])];

fprintf('%s', s);
thisDir = fileparts(mfilename('fullpath'));
resDir = fullfile(thisDir, 'results');
if ~exist(resDir, 'dir'); mkdir(resDir); end
fid = fopen(fullfile(resDir, [scenarioName '.txt']), 'w');
if fid < 0; error('tdcs_paper_lme:fopen', 'Cannot open results file.'); end
cleanup = onCleanup(@() fclose(fid)); %#ok<NASGU>
fprintf(fid, '%s', s);
end

function r = localRow(name, e)
r = sprintf('%-26s t(%d)=%6.2f F(%d)=%7.3f  p=%.4g    p=%.4g (df=%.0f)\n', ...
    name, e.df, e.t, e.df1, e.F, e.p, e.pSatt, e.dfSatt);
end

function s = localHonest(rs)
if rs.ok
    s = sprintf(['\nHONEST LME -- per-rat random slope (day|rat): ' ...
        'interaction F(%d,%.1f)=%.2f, p=%.4g\n'], rs.df1, rs.df2, rs.F, rs.p);
else
    s = sprintf(['\nHONEST LME -- per-rat random slope (day|rat): ' ...
        'model did not converge for this window.\n']);
end
s = [s sprintf(['  Satterthwaite DF on the random-INTERCEPT model above stays ~= residual (the\n' ...
    '  slope''s error is at session level), so it does NOT fix pseudoreplication. A per-animal\n' ...
    '  random slope collapses the interaction DF toward the animal count -- the honest test.\n'])];
end

function t = localSigTxt(p)
if p < 0.05; t = 'SIGNIFICANT'; else; t = 'n.s.'; end
end

function t = localPick(b, yes, no)
if b; t = yes; else; t = no; end
end