feat(ch13): Stroop within-subject analysis (paired t-test + LMM) and plots

scripts/ch13_stroop_within.py

@@ -0,0 +1,95 @@
+# SPDX-License-Identifier: MIT
+"""
+Chapter 13 — Within-Subjects (Stroop): paired t-test vs. mixed model
+Usage:
+  python scripts/ch13_stroop_within.py --save-plots
+Options:
+  --data data/synthetic/psych_stroop_trials.csv
+  --min-rt 200 --max-rt 2000    # trial-level trims
+"""
+import argparse, os, math, numpy as np, pandas as pd
+import statsmodels.api as sm
+import statsmodels.formula.api as smf
+from scipy import stats
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+def load_trials(path):
+    df = pd.read_csv(path)
+    df["condition"] = df["condition"].astype("category")
+    return df
+
+def cohen_dz(a, b):
+    # within-subject effect size
+    diff = a - b
+    return diff.mean()/diff.std(ddof=1)
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--data", default="data/synthetic/psych_stroop_trials.csv")
+    ap.add_argument("--min-rt", type=float, default=200.0)
+    ap.add_argument("--max-rt", type=float, default=2000.0)
+    ap.add_argument("--save-plots", action="store_true")
+    args = ap.parse_args()
+
+    os.makedirs("outputs", exist_ok=True)
+    df = load_trials(args.data)
+
+    # Keep correct trials and trim RTs
+    before = len(df)
+    df = df[(df["correct"] == 1) & (df["rt_ms"].between(args.min_rt, args.max_rt))]
+    print(f"Filtered: kept {len(df)}/{before} trials (correct & {args.min_rt}-{args.max_rt} ms)")
+
+    # Log-transform RT (typical for skew)
+    df["log_rt"] = np.log(df["rt_ms"])
+
+    # Subject means per condition (paired t-test)
+    subj_means = df.groupby(["subject", "condition"], as_index=False)["log_rt"].mean()
+    wide = subj_means.pivot(index="subject", columns="condition", values="log_rt")
+    # paired t-test (incongruent > congruent expected)
+    tstat, pval = stats.ttest_rel(wide["incongruent"], wide["congruent"], nan_policy="omit")
+    dz = cohen_dz(wide["incongruent"], wide["congruent"])
+
+    print("\n=== Paired t-test on subject means (log RT) ===")
+    print(f"t = {tstat:.3f}, p = {pval:.3g}, Cohen's dz = {dz:.3f}")
+    delta_ms = (np.exp(wide["incongruent"].mean()) - np.exp(wide["congruent"].mean()))
+    print(f"Back-translated mean RT difference ≈ {delta_ms:.1f} ms (incongruent - congruent)")
+
+    # Mixed model on trial-level log RT
+    # random intercepts for subjects; (you can extend to random slopes)
+    md = smf.mixedlm("log_rt ~ C(condition)", df, groups=df["subject"])
+    m = md.fit(reml=False)
+    print("\n=== Mixed Model (trial-level log RT) ===")
+    print(m.summary())
+
+    # Simple marginal means (back-transform)
+    emmeans = df.groupby("condition")["log_rt"].mean().apply(np.exp)
+    print("\nEstimated marginal means (back-transformed ms):")
+    for k, v in emmeans.items():
+        print(f"  {k:<12} = {v:.1f} ms")
+
+    # Plots
+    if args.save_plots:
+        # per-subject spaghetti of means
+        plt.figure(figsize=(6,4))
+        for _, row in wide.iterrows():
+            plt.plot([0,1], [row["congruent"], row["incongruent"]], alpha=0.3)
+        plt.xticks([0,1], ["congruent","incongruent"])
+        plt.ylabel("Mean log RT")
+        plt.title("Within-subject means (log RT)")
+        plt.tight_layout(); plt.savefig("outputs/ch13_stroop_subject_means.png", dpi=150)
+
+        # violin/box of trial-level log RT by condition
+        plt.figure(figsize=(6,4))
+        data_c = [df.loc[df["condition"]==c,"log_rt"].values for c in ["congruent","incongruent"]]
+        plt.violinplot(data_c, showmeans=True)
+        plt.xticks([1,2], ["congruent","incongruent"])
+        plt.ylabel("Trial log RT")
+        plt.title("Trial-level distribution by condition")
+        plt.tight_layout(); plt.savefig("outputs/ch13_stroop_violins.png", dpi=150)
+
+        print("Saved plots -> outputs/ch13_stroop_subject_means.png, ch13_stroop_violins.png")
+
+if __name__ == "__main__":
+    main()