"""F2-conf-01: school-level programme effect with cluster-robust inference."""
import json, os
import numpy as np, pandas as pd
import statsmodels.formula.api as smf
from scipy import stats

ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
d = pd.read_csv(os.path.join(ROOT, "data", "schools.csv"))
n_rows = len(d)
issues = []
# --- data checks ---
if d.isna().any().any():
    issues.append({"issue": "missing values", "handling": "rows with missing values dropped"})
    d = d.dropna()
if d.student_id.duplicated().any():
    issues.append({"issue": "duplicate student_id", "handling": "duplicates dropped (first kept)"})
    d = d.drop_duplicates("student_id")
bad = d.groupby("school_id").treated.nunique()
if (bad > 1).any():
    issues.append({"issue": "treatment varies within school", "handling": "schools excluded"})
    d = d[~d.school_id.isin(bad[bad > 1].index)]
sizes = d.groupby("school_id").size()
assert sizes.nunique() == 1, "unbalanced"
G = d.school_id.nunique(); m = int(sizes.iloc[0]); N = len(d)
G1 = int(d.groupby("school_id").treated.first().sum())

# --- OLS, iid SE ---
ols = smf.ols("score ~ treated", data=d).fit()
tau = float(ols.params["treated"]); se_iid = float(ols.bse["treated"])
# --- cluster-robust (CR1, Stata-style small-sample correction), t with G-1 df ---
cl = smf.ols("score ~ treated", data=d).fit(cov_type="cluster",
        cov_kwds={"groups": d.school_id.values, "use_correction": True})
se_cl = float(cl.bse["treated"])
df = G - 1
tstat = tau / se_cl
p = float(2 * stats.t.sf(abs(tstat), df))
q = stats.t.ppf(0.975, df)
lo, hi = tau - q * se_cl, tau + q * se_cl

# cross-check: two-sample t-test on school means (exact analogue under balance)
sm = d.groupby("school_id").agg(y=("score", "mean"), t=("treated", "first"))
tt = stats.ttest_ind(sm.y[sm.t == 1], sm.y[sm.t == 0], equal_var=True)
se_means = float(tau / tt.statistic)

# --- ICC of residual: one-way ANOVA estimator on OLS residuals ---
d = d.assign(r=ols.resid)
rm = d.groupby("school_id").r.transform("mean")
ssw = ((d.r - rm) ** 2).sum(); msw = ssw / (N - G)
ssb = m * (d.groupby("school_id").r.mean() ** 2).sum(); msb = ssb / (G - 2)  # 2 params removed
s2b = max((msb - msw) / m, 0.0)
icc = float(s2b / (s2b + msw))
deff = 1 + (m - 1) * icc

res = {
    "task_id": "F2-conf-01", "tau_hat": tau, "se_iid": se_iid, "se_cluster": se_cl,
    "p_value": p, "tau_ci_low": float(lo), "tau_ci_high": float(hi),
    "effect_significant_5pct": bool(p < 0.05), "icc_hat": icc,
    "n_schools": int(G), "n_treated_schools": G1, "n_analysed": int(N),
    "feasibility": "complete", "feasibility_reason": None, "data_issues": issues,
    "n_input_rows": {"schools.csv": int(n_rows)}, "random_seed": None,
    "_diagnostics": {"t_stat_cluster": float(tstat), "df": df, "students_per_school": m,
        "se_school_means_ttest": se_means, "p_school_means_ttest": float(tt.pvalue),
        "p_iid": float(ols.pvalues["treated"]), "design_effect_from_icc": deff,
        "se_ratio_cluster_over_iid": se_cl / se_iid},
}
os.makedirs(os.path.join(ROOT, "out"), exist_ok=True)
with open(os.path.join(ROOT, "out", "results.json"), "w") as f:
    json.dump(res, f, indent=2)
print(json.dumps(res, indent=2))