analysis(matlab): add log(day) model + Cohen's f + power to every variation
Add variation_logpower.m (self-contained) + make_variation_logpower.m, dropping logpowersim.m + logpower_result.txt into all 28 variation folders. Matches the paper's power code: models behavior ~ stim + log(day) + stim:log(day) + (1|rat) (log(day+1), since our day 0 = paper Day 1), reports Cohen's f (partial eta^2 of the interaction) and the interaction under residual/Satterthwaite/honest random-slope DF, then runs the Monte-Carlo power sim on the log-day ground truth (per-animal cluster-honest + LME power). Notable: under log(day) the accumulating divergence is captured more sharply, so several full-window scenarios reach honest significance that were n.s. under raw day (e.g. right_only_d0_13 honest p=0.003, unmerge_d0_13 0.021, unmerge_d0_10 0.036); Cohen's f is small-medium (~0.10-0.34). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Experiments DB Dev
@@ -0,0 +1,36 @@
|
||||
==============================================================================
|
||||
LOG-DAY MODEL + COHEN'S f + POWER -- prev_f_full
|
||||
==============================================================================
|
||||
model: behavior ~ stim + log(day) + stim:log(day) + (1|rat) (success COUNT)
|
||||
log(day) uses 1-indexed training day (our day 0 = paper "Day 1")
|
||||
observed groups: stim n=12, control n=12 nrep=120, alpha=0.05
|
||||
|
||||
--- fitted on real data ---
|
||||
stim x log(day) interaction: F(1,227)=8.356 p(resid)=0.004218 p(Satt)=0.004253 (df=208)
|
||||
honest per-animal random slope (log-day): F(1,24.3)=3.32 p=0.08079
|
||||
Cohen's f (interaction, partial eta^2=0.012) = 0.112 (small-medium; f: .10 small, .25 medium, .40 large)
|
||||
|
||||
--- power simulation (log-day ground truth: stim:logday=+2.39, ratSD=4.33, resSD=4.34) ---
|
||||
|
||||
true stim:log(day) = +2.39 (100% of observed)
|
||||
N/group | per-animal power | LME power
|
||||
------------------------------------------
|
||||
3 | 0.15 | 0.28
|
||||
5 | 0.33 | 0.46
|
||||
8 | 0.61 | 0.65
|
||||
12 | 0.82 | 0.83 <- observed
|
||||
16 | 0.92 | 0.94
|
||||
24 | 0.99 | 0.98
|
||||
|
||||
true stim:log(day) = +1.19 (50% of observed)
|
||||
N/group | per-animal power | LME power
|
||||
------------------------------------------
|
||||
3 | 0.05 | 0.09
|
||||
5 | 0.12 | 0.17
|
||||
8 | 0.20 | 0.26
|
||||
12 | 0.33 | 0.32 <- observed
|
||||
16 | 0.46 | 0.49
|
||||
24 | 0.46 | 0.49
|
||||
|
||||
Read the per-animal column as the honest power; the LME column matches the
|
||||
paper's power code (anova interaction p, observation-level DF) and is optimistic.
|
||||
@@ -0,0 +1,148 @@
|
||||
% Variation log-day analysis + Cohen's f + power simulation.
|
||||
%
|
||||
% The paper's power code models behavior against LOG training day, not raw day:
|
||||
% behavior ~ stim + log(day) + stim:log(day) + (1|rat).
|
||||
% Their day is 1-indexed (1..10); our data.csv day is 0-indexed (day 0 = paper
|
||||
% "Day 1"), so log(day + 1) reproduces their transform exactly.
|
||||
%
|
||||
% This script (a) refits that log-day model on data.csv, (b) reports the
|
||||
% interaction (residual DF, Satterthwaite DF, and the honest per-animal
|
||||
% random-slope test) and Cohen's f -- the partial-eta^2 effect size of the
|
||||
% interaction, var(fitted_full) - var(fitted_no_interaction) over var(behavior)
|
||||
% -- and (c) runs the Monte-Carlo power simulation on the log-day model,
|
||||
% scoring per-animal (cluster-honest) and LME power across N.
|
||||
% Writes logpower_result.txt. Run: matlab -batch "logpowersim"
|
||||
% (Copy of analysis/matlab/variation_logpower.m; see make_variation_logpower.m.)
|
||||
|
||||
here = fileparts(mfilename('fullpath'));
|
||||
if isempty(here); here = pwd; end
|
||||
vname = regexprep(here, '.*[/\\]', '');
|
||||
|
||||
D = readtable(fullfile(here, 'data.csv'), 'TextType', 'string');
|
||||
FORMULA = 'behavior ~ stim + day + stim:day + (1|rat)'; % 'day' column = log(day+1)
|
||||
NS = [3 5 8 12 16 24];
|
||||
EFFMULS = [1 0.5];
|
||||
NREP = 120;
|
||||
|
||||
warnState = warning('off', 'all');
|
||||
rng(1);
|
||||
|
||||
logday = log(D.day + 1); % 0-indexed day -> their log(1-indexed day)
|
||||
tbl0 = table(D.success, logday, double(D.stim), categorical(D.subject), ...
|
||||
'VariableNames', {'behavior', 'day', 'stim', 'rat'});
|
||||
nStim = numel(unique(D.subject(D.stim == 1)));
|
||||
nCtrl = numel(unique(D.subject(D.stim == 0)));
|
||||
|
||||
bar = repmat('=', 1, 78);
|
||||
s = sprintf('%s\nLOG-DAY MODEL + COHEN''S f + POWER -- %s\n%s\n', bar, vname, bar);
|
||||
s = [s sprintf('model: behavior ~ stim + log(day) + stim:log(day) + (1|rat) (success COUNT)\n')];
|
||||
s = [s sprintf('log(day) uses 1-indexed training day (our day 0 = paper "Day 1")\n')];
|
||||
s = [s sprintf('observed groups: stim n=%d, control n=%d nrep=%d, alpha=0.05\n', nStim, nCtrl, NREP)];
|
||||
|
||||
if nStim < 2 || nCtrl < 2 || numel(unique(tbl0.day)) < 2
|
||||
s = [s sprintf('\nInsufficient data for this analysis (need >=2 animals/group and >=2 days).\n')];
|
||||
localFinish(s, here); warning(warnState); return
|
||||
end
|
||||
|
||||
% ---- fitted model on the real data ----
|
||||
full = fitlme(tbl0, FORMULA);
|
||||
An = anova(full); Asatt = anova(full, 'DFMethod', 'satterthwaite');
|
||||
ii = strcmp(An.Term, 'day:stim'); is = strcmp(Asatt.Term, 'day:stim');
|
||||
reduced = fitlme(tbl0, 'behavior ~ stim + day + (1|rat)');
|
||||
eta2part = max((var(fitted(full)) - var(fitted(reduced))) / var(tbl0.behavior), 0);
|
||||
cohenf = sqrt(eta2part / (1 - eta2part));
|
||||
|
||||
% honest per-animal random-slope interaction
|
||||
rsP = NaN; rsDf = NaN; rsF = NaN; rsOk = false;
|
||||
try
|
||||
mr = fitlme(tbl0, 'behavior ~ stim + day + stim:day + (day|rat)');
|
||||
Ar = anova(mr, 'DFMethod', 'satterthwaite'); ri = strcmp(Ar.Term, 'day:stim');
|
||||
rsF = Ar.FStat(ri); rsDf = Ar.DF2(ri); rsP = Ar.pValue(ri); rsOk = true;
|
||||
catch
|
||||
end
|
||||
if cohenf < 0.10; mag = 'small';
|
||||
elseif cohenf < 0.25; mag = 'small-medium';
|
||||
elseif cohenf < 0.40; mag = 'medium';
|
||||
else; mag = 'large'; end
|
||||
|
||||
s = [s sprintf('\n--- fitted on real data ---\n')];
|
||||
s = [s sprintf('stim x log(day) interaction: F(1,%d)=%.3f p(resid)=%.4g p(Satt)=%.4g (df=%.0f)\n', ...
|
||||
An.DF2(ii), An.FStat(ii), An.pValue(ii), Asatt.pValue(is), Asatt.DF2(is))];
|
||||
if rsOk
|
||||
s = [s sprintf(' honest per-animal random slope (log-day): F(1,%.1f)=%.2f p=%.4g\n', rsDf, rsF, rsP)];
|
||||
else
|
||||
s = [s sprintf(' honest per-animal random slope (log-day): did not converge\n')];
|
||||
end
|
||||
s = [s sprintf('Cohen''s f (interaction, partial eta^2=%.3f) = %.3f (%s; f: .10 small, .25 medium, .40 large)\n', ...
|
||||
eta2part, cohenf, mag)];
|
||||
|
||||
% ---- power simulation on the log-day ground truth ----
|
||||
cn = full.CoefficientNames; be = full.fixedEffects;
|
||||
b0 = be(strcmp(cn, '(Intercept)')); bStim = be(strcmp(cn, 'stim'));
|
||||
bDay = be(strcmp(cn, 'day')); bInt = be(strcmp(cn, 'day:stim'));
|
||||
psi = covarianceParameters(full); sRat = sqrt(psi{1}); sRes = sqrt(full.MSE);
|
||||
days = unique(tbl0.day); % the log(day) grid
|
||||
|
||||
s = [s sprintf('\n--- power simulation (log-day ground truth: stim:logday=%+.2f, ratSD=%.2f, resSD=%.2f) ---\n', ...
|
||||
bInt, sRat, sRes)];
|
||||
for eMul = EFFMULS
|
||||
bI = bInt * eMul;
|
||||
s = [s sprintf('\n true stim:log(day) = %+.2f (%.0f%% of observed)\n', bI, eMul * 100)]; %#ok<AGROW>
|
||||
s = [s sprintf(' %-8s | per-animal power | LME power\n %s\n', 'N/group', repmat('-', 1, 42))]; %#ok<AGROW>
|
||||
for N = NS
|
||||
sigPA = 0; sigL = 0;
|
||||
for r = 1:NREP
|
||||
tb = localSim(N, days, b0, bStim, bDay, bI, sRat, sRes);
|
||||
sigPA = sigPA + (localPAp(tb) < 0.05);
|
||||
try
|
||||
m = fitlme(tb, FORMULA); Cm = m.Coefficients;
|
||||
sigL = sigL + (Cm.pValue(strcmp(Cm.Name, 'day:stim')) < 0.05);
|
||||
catch
|
||||
end
|
||||
end
|
||||
star = '';
|
||||
if N == nStim || N == nCtrl; star = ' <- observed'; end
|
||||
s = [s sprintf(' %-8d | %5.2f | %5.2f%s\n', N, sigPA / NREP, sigL / NREP, star)]; %#ok<AGROW>
|
||||
end
|
||||
end
|
||||
s = [s sprintf(['\nRead the per-animal column as the honest power; the LME column matches the\n' ...
|
||||
'paper''s power code (anova interaction p, observation-level DF) and is optimistic.\n'])];
|
||||
|
||||
localFinish(s, here);
|
||||
warning(warnState);
|
||||
|
||||
% ---------------------------------------------------------------- helpers
|
||||
function localFinish(s, here)
|
||||
fprintf('%s', s);
|
||||
fid = fopen(fullfile(here, 'logpower_result.txt'), 'w');
|
||||
fprintf(fid, '%s', s); fclose(fid);
|
||||
end
|
||||
|
||||
function tbl = localSim(N, days, b0, bStim, bDay, bI, sRat, sRes)
|
||||
nd = numel(days); rows = 2 * N * nd;
|
||||
rat = strings(rows, 1); day = zeros(rows, 1); stim = zeros(rows, 1); behavior = zeros(rows, 1);
|
||||
k = 0;
|
||||
for g = 0:1
|
||||
for sIdx = 1:N
|
||||
re = sRat * randn;
|
||||
rid = sprintf('g%d_r%d', g, sIdx);
|
||||
for d = 1:nd
|
||||
k = k + 1;
|
||||
rat(k) = rid; day(k) = days(d); stim(k) = g;
|
||||
behavior(k) = b0 + bStim * g + bDay * days(d) + bI * days(d) * g + re + sRes * randn;
|
||||
end
|
||||
end
|
||||
end
|
||||
tbl = table(categorical(rat), day, stim, behavior, ...
|
||||
'VariableNames', {'rat', 'day', 'stim', 'behavior'});
|
||||
end
|
||||
|
||||
function p = localPAp(tbl)
|
||||
rats = unique(tbl.rat); sl = zeros(numel(rats), 1); gr = zeros(numel(rats), 1);
|
||||
for i = 1:numel(rats)
|
||||
r = tbl.rat == rats(i);
|
||||
c = polyfit(tbl.day(r), tbl.behavior(r), 1); sl(i) = c(1);
|
||||
gr(i) = tbl.stim(find(r, 1));
|
||||
end
|
||||
[~, p] = ttest2(sl(gr == 1), sl(gr == 0), 'Vartype', 'unequal');
|
||||
end
|
||||
Reference in New Issue
Block a user