src/causal/robustness.py

from sklearn.linear_model import LogisticRegression
from sklearn.utils import resample

def make_dr_balanced():
    return DRLearner(
        model_regression=RandomForestRegressor(
            n_estimators=300, min_samples_leaf=5, random_state=42, n_jobs=-1
        ),
        model_propensity=LogisticRegression(max_iter=1000, solver="lbfgs", class_weight="balanced"),
        cv=1,
        random_state=42
    )

robust_results = {}
placebo = {}
rng = np.random.default_rng(123)

for t in T_cols:
    T_all = C[t].astype(float).to_numpy()
    thresh = np.median(T_all)
    T_bin = (T_all > thresh).astype(int)
    T_tr, T_test = T_bin[idx_tr], T_bin[idx_te]

    dr = make_dr_balanced()
    dr.fit(y_tr, T_tr, X=X_tr)

    prop = LogisticRegression(max_iter=1000, solver="lbfgs", class_weight="balanced")
    prop.fit(X_tr, T_tr)
    e_hat = prop.predict_proba(X_te)[:, 1]

    mask = (e_hat > 0.05) & (e_hat < 0.95)
    if mask.sum() < max(50, int(0.2 * len(X_te))):
        mask = (e_hat > 0.02) & (e_hat < 0.98)

    ate_trim = float(dr.ate(X_te[mask]))

    boots = []
    n_trim = int(mask.sum())
    for _ in range(300):
        b = rng.integers(0, n_trim, size=n_trim)
        boots.append(float(dr.ate(X_te[mask][b])))
    ci_trim = (float(np.percentile(boots, 2.5)), float(np.percentile(boots, 97.5)))

    robust_results[t] = {"ATE_trim": ate_trim, "CI_trim": ci_trim, "trim_keep": int(mask.sum())}

    T_tr_shuff = rng.permutation(T_tr)
    dr_p = make_dr_balanced()
    dr_p.fit(y_tr, T_tr_shuff, X=X_tr)
    ate_placebo = float(dr_p.ate(X_te))
    placebo[t] = ate_placebo

out = {"robust": robust_results, "placebo": placebo}
with open("results/causal_section3_robustness.json", "w") as f:
    json.dump(out, f, indent=2)

print(" Robustness checks saved → results/causal_section3_robustness.json")
for k, v in robust_results.items():
    print(f"{k}: ATE_trim={v['ATE_trim']:.3f}, CI_trim={tuple(np.round(v['CI_trim'],3))}, kept={v['trim_keep']}")
print("Placebo (ATE; expected ≈ 0):", {k: round(v, 3) for k, v in placebo.items()})

import json, numpy as np, pandas as pd, matplotlib.pyplot as plt, seaborn as sns, pathlib as p
from sklearn.model_selection import train_test_split
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.linear_model import LinearRegression

plt.rcParams["figure.dpi"] = 150
sns.set_style("whitegrid")

RES = p.Path("results"); RES.mkdir(exist_ok=True, parents=True)
rob_path = RES/"causal_section3_robustness.json"
rob = json.load(open(rob_path))

rows=[]
for k,v in rob.get("robust", {}).items():
    rows.append({
        "transporter": k,
        "ATE_trim": float(v["ATE_trim"]),
        "CI_low": float(v["CI_trim"][0]),
        "CI_high": float(v["CI_trim"][1]),
        "kept_n": int(v.get("trim_keep", v.get("kept", np.nan)))
    })
df = pd.DataFrame(rows).sort_values("ATE_trim", ascending=False)
df.to_csv(RES/"ED_Table_S3_causal_robustness.csv", index=False)


plt.figure(figsize=(8, max(3.5, 0.42*len(df))))
ax = sns.barplot(data=df, y="transporter", x="ATE_trim", color="steelblue", orient="h")
for i, r in df.reset_index(drop=True).iterrows():
    plt.plot([r.CI_low, r.CI_high], [i, i], color="k", lw=1)
plt.axvline(0, color="red", ls="--", lw=1)
plt.xlabel("ATE (trimmed)"); plt.ylabel("")
plt.title("Trimmed ATEs (95% CI)")
plt.tight_layout()
plt.savefig(RES/"ED_Fig_trimmed_ATEs.png", dpi=300)
plt.show()


placebo_samples = []

if "placebo_samples" in rob and rob["placebo_samples"]:
    for t, arr in rob["placebo_samples"].items():
        vals = np.array(arr, dtype=float).ravel().tolist()
        placebo_samples.append(pd.DataFrame({"transporter": t, "ATE_placebo": vals}))

elif "placebo" in rob and rob["placebo"]:
    for t, val in rob["placebo"].items():
        v = float(val)
        jitter = v + 0.002*np.random.default_rng(123).standard_normal(50)
        placebo_samples.append(pd.DataFrame({"transporter": t, "ATE_placebo": jitter.tolist()}))

else:
    import pandas as pd
    C = pd.read_csv("data/processed/causal_table.csv")
    cov_cont = ["ethanol_pct","ROS","NaCl_mM","H2O2_uM","PDR1_reg","YAP1_reg"]
    cov_cat  = ["batch"]
    X_df = C[cov_cont+cov_cat].copy()
    X_df[cov_cont] = X_df[cov_cont].astype(float)
    X_df[cov_cat]  = X_df[cov_cat].astype(str)

    ct = ColumnTransformer([
        ("num", StandardScaler(), cov_cont),
        ("cat", OneHotEncoder(sparse_output=False, handle_unknown="ignore"), cov_cat)
    ])
    X_all = ct.fit_transform(X_df)
    y = C["outcome"].astype(float).to_numpy()

    idx = np.arange(len(y))
    idx_tr, idx_te = train_test_split(idx, test_size=0.2, random_state=42)
    X_tr, X_te = X_all[idx_tr], X_all[idx_te]
    y_tr, y_te = y[idx_tr], y[idx_te]

    ols = LinearRegression().fit(X_tr, y_tr)
    r_te = y_te - ols.predict(X_te)

    rng = np.random.default_rng(123)
    T_cols = [c for c in C.columns if c.endswith("_expr")]

    for t in T_cols:
        T = C[t].astype(float).to_numpy()
        # binarize like Section 3
        T_bin = (T > np.median(T)).astype(int)
        T_tr = T_bin[idx_tr]

        vals=[]
        for _ in range(200):
            T_perm = rng.permutation(T_tr)
            T_te = rng.permutation(T_bin[idx_te])
            m1 = r_te[T_te==1].mean() if (T_te==1).any() else 0.0
            m0 = r_te[T_te==0].mean() if (T_te==0).any() else 0.0
            vals.append(m1 - m0)
        placebo_samples.append(pd.DataFrame({"transporter": t, "ATE_placebo": vals}))

P = pd.concat(placebo_samples, ignore_index=True)

plt.figure(figsize=(8, 4.5))
ax = sns.histplot(P, x="ATE_placebo", hue="transporter", element="step",
                  stat="density", common_norm=False, bins=30, alpha=0.35)
ax.set_xlabel("Placebo ATE"); ax.set_ylabel("Density")
ax.set_title("Placebo ATE distributions")
plt.tight_layout()
plt.savefig(RES/"ED_Fig_placebo_hist.png", dpi=300)
plt.show()

stats = P.groupby("transporter")["ATE_placebo"].agg(["mean","std"]).reset_index()
stats.to_csv(RES/"ED_Table_placebo_stats.csv", index=False)

print("Saved:",
      RES/"ED_Table_S3_causal_robustness.csv",
      RES/"ED_Fig_trimmed_ATEs.png",
      RES/"ED_Fig_placebo_hist.png",
      RES/"ED_Table_placebo_stats.csv")