"""F2-conf-01: school-level programme effect with cluster-robust uncertainty.
Run from workspace root: ~/wp29/venv/bin/python out/analysis.py"""
import json
import numpy as np
import pandas as pd
import statsmodels.formula.api as smf
from scipy import stats

d = pd.read_csv("data/schools.csv")
n_input = 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")
tconst = d.groupby("school_id")["treated"].nunique()
if (tconst > 1).any():
    issues.append({"issue": "treated not constant within school", "handling": "reported; not altered"})
sizes = d.groupby("school_id").size()

G = d["school_id"].nunique()
G1 = int(d.groupby("school_id")["treated"].first().sum())

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) by school,
# inference with t distribution on G-1 degrees of freedom.
cl = smf.ols("score ~ treated", data=d).fit(cov_type="cluster",
        cov_kwds={"groups": d["school_id"], "use_correction": True, "df_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 (equivalent analysis at school level)
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((sm.y[sm.t == 1].mean() - sm.y[sm.t == 0].mean()) / tt.statistic)

# ICC of residual: one-way ANOVA estimator on OLS residuals by school
d = d.assign(r=ols.resid)
m = sizes.iloc[0] if sizes.nunique() == 1 else (len(d) - (sizes**2).sum() / len(d)) / (G - 1)
gm = d.groupby("school_id")["r"].mean()
msb = (sizes * (gm - d["r"].mean())**2).sum() / (G - 1)
msw = ((d["r"] - d["school_id"].map(gm))**2).sum() / (len(d) - G)
s2b = max((msb - msw) / m, 0.0)
icc = float(s2b / (s2b + msw))

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(len(d)),
    "feasibility": "complete",
    "feasibility_reason": None,
    "data_issues": issues,
    "n_input_rows": {"schools.csv": int(n_input)},
    "random_seed": None,
    "_diagnostics": {
        "t_stat_cluster": float(tstat), "df_cluster": int(df),
        "p_value_iid": float(ols.pvalues["treated"]),
        "se_school_means_ttest": se_means, "p_school_means_ttest": float(tt.pvalue),
        "design_effect_implied": float((se_cl / se_iid)**2),
        "design_effect_icc": float(1 + (m - 1) * icc),
        "students_per_school": float(m),
        "msb": float(msb), "msw": float(msw),
    },
}
with open("out/results.json", "w") as f:
    json.dump(res, f, indent=2)
print(json.dumps(res, indent=2))