"""F2-conf-01: school-level programme effect with cluster-robust uncertainty.
Reads data/schools.csv only; writes out/results.json (deterministic, no randomness)."""
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_raw = len(d)
issues = []
# Audit checks (no exclusions needed if all pass)
assert d[["school_id", "student_id", "treated", "score"]].notna().all().all(), "missing values"
assert not d.student_id.duplicated().any(), "duplicate student ids"
assert (d.groupby("school_id").treated.nunique() == 1).all(), "treatment varies within school"
assert set(d.treated.unique()) <= {0, 1}
sizes = d.groupby("school_id").size()
if sizes.nunique() != 1:
    issues.append({"issue": "unbalanced school sizes", "handling": "kept; CR1 handles it"})

# OLS of score on treated with intercept
m_iid = smf.ols("score ~ treated", data=d).fit()
G = d.school_id.nunique()
m_cl = smf.ols("score ~ treated", data=d).fit(cov_type="cluster",
                                             cov_kwds={"groups": d.school_id}, use_t=True)
tau = float(m_iid.params["treated"])
se_iid = float(m_iid.bse["treated"])
se_cl = float(m_cl.bse["treated"])  # CR1: G/(G-1)*(N-1)/(N-K) small-sample correction
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

# Sensitivity: two-sample t-test on school means (balanced design)
sm = d.groupby("school_id").agg(score=("score", "mean"), treated=("treated", "first"))
tt = stats.ttest_ind(sm.score[sm.treated == 1], sm.score[sm.treated == 0], equal_var=True)

# ICC of residual: one-way ANOVA estimator on OLS residuals by school
d["resid"] = m_iid.resid
m = int(sizes.iloc[0]) if sizes.nunique() == 1 else float(sizes.mean())
gm = d.groupby("school_id").resid.mean()
msb = m * ((gm - d.resid.mean()) ** 2).sum() / (G - 1)
msw = ((d.resid - d.school_id.map(gm)) ** 2).sum() / (len(d) - G)
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": int(sm.treated.sum()),
    "n_analysed": int(len(d)),
    "feasibility": "complete",
    "feasibility_reason": None,
    "data_issues": issues,
    "n_input_rows": {"schools.csv": int(n_raw)},
    "random_seed": None,
    "t_df": int(df),
    "t_stat_cluster": float(tstat),
    "p_value_iid": float(m_iid.pvalues["treated"]),
    "se_ratio_cluster_to_iid": se_cl / se_iid,
    "design_effect_from_icc": float(deff),
    "school_mean_ttest_t": float(tt.statistic),
    "school_mean_ttest_p": float(tt.pvalue),
    "students_per_school": m,
}
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))