Repository Public
1"""Reference values for Statistics/Regression. References: statsmodels 0.14 (OLS, WLS, Logit,2robust covariances, statsmodels.stats.diagnostic, OLSInfluence) and linearmodels PanelOLS."""3import warnings45import numpy as np6import pandas as pd7import statsmodels.api as sm8from statsmodels.stats import diagnostic as dg9from statsmodels.stats.outliers_influence import OLSInfluence, variance_inflation_factor10from statsmodels.stats.stattools import durbin_watson, jarque_bera1112from gen_common import dump1314warnings.simplefilter("ignore")15rng = np.random.default_rng(20261121)16out = {}171819def summary(res, ci=0.95):20 c = res.conf_int(alpha=1 - ci)21 c = np.asarray(c)22 return {23 "params": np.asarray(res.params), "bse": np.asarray(res.bse), "t": np.asarray(res.tvalues), "p": np.asarray(res.pvalues),24 "ci_lo": c[:, 0], "ci_hi": c[:, 1], "cov": np.asarray(res.cov_params()),25 "fvalue": float(np.squeeze(res.fvalue)), "f_pvalue": float(np.squeeze(res.f_pvalue)),26 }272829# ---------------------------------------------------------------- OLS ---------------------------30n = 6031x1 = rng.normal(10, 3, n)32x2 = 0.5 * x1 + rng.normal(0, 2, n)33x3 = rng.uniform(0, 5, n)34e = np.zeros(n)35u = rng.normal(0, 1, n) * (0.5 + 0.15 * x1)36for t in range(n):37 e[t] = (0.5 * e[t - 1] if t else 0) + u[t]38y = 2.0 + 1.5 * x1 - 0.8 * x2 + 0.3 * x3 ** 2 + e39X = np.column_stack([x1, x2, x3])40Xc = sm.add_constant(X)41groups = np.repeat(np.arange(10), 6)42res = sm.OLS(y, Xc).fit()43infl = OLSInfluence(res)44ols = {45 "y": y, "X": X, "groups": groups, **summary(res),46 "ci90_lo": np.asarray(res.conf_int(alpha=0.10))[:, 0], "ci90_hi": np.asarray(res.conf_int(alpha=0.10))[:, 1],47 "rsquared": res.rsquared, "rsquared_adj": res.rsquared_adj, "llf": res.llf, "aic": res.aic, "bic": res.bic,48 "mse_resid": res.mse_resid, "ssr": res.ssr, "tss": res.centered_tss, "df_model": res.df_model, "df_resid": res.df_resid,49 "resid": res.resid, "fitted": res.fittedvalues, "hat": infl.hat_matrix_diag, "cooks": infl.cooks_distance[0],50 "student": infl.resid_studentized_internal,51 "vif": [variance_inflation_factor(Xc, j) for j in range(1, Xc.shape[1])], "cond": res.condition_number,52 "predict_X": [[9.0, 4.0, 2.5], [14.0, 8.0, 0.5]], "predict": res.predict(np.array([[1, 9.0, 4.0, 2.5], [1, 14.0, 8.0, 0.5]])),53}54for ct in ["HC0", "HC1", "HC2", "HC3"]:55 ols[ct] = summary(sm.OLS(y, Xc).fit(cov_type=ct))56 ols[ct + "_t"] = summary(sm.OLS(y, Xc).fit(cov_type=ct, use_t=True))57ols["HAC4"] = summary(sm.OLS(y, Xc).fit(cov_type="HAC", cov_kwds={"maxlags": 4}))58ols["HAC4_small"] = summary(sm.OLS(y, Xc).fit(cov_type="HAC", cov_kwds={"maxlags": 4, "use_correction": True}))59ols["HAC0"] = summary(sm.OLS(y, Xc).fit(cov_type="HAC", cov_kwds={"maxlags": 0}))60ols["HAC4_t"] = summary(sm.OLS(y, Xc).fit(cov_type="HAC", cov_kwds={"maxlags": 4}, use_t=True))61ols["cluster"] = summary(sm.OLS(y, Xc).fit(cov_type="cluster", cov_kwds={"groups": groups}))62ols["cluster_t"] = summary(sm.OLS(y, Xc).fit(cov_type="cluster", cov_kwds={"groups": groups}, use_t=True))63ols["cluster_nocorr"] = summary(sm.OLS(y, Xc).fit(cov_type="cluster", cov_kwds={"groups": groups, "use_correction": False}))6465# diagnostics66bp = dg.het_breuschpagan(res.resid, Xc)67bp0 = dg.het_breuschpagan(res.resid, Xc, robust=False)68wh = dg.het_white(res.resid, Xc)69bg1 = dg.acorr_breusch_godfrey(res, nlags=1)70bg3 = dg.acorr_breusch_godfrey(res, nlags=3)71lb = dg.acorr_ljungbox(res.resid, lags=[1, 5, 10], boxpierce=True)72lb_df = dg.acorr_ljungbox(res.resid, lags=[5], model_df=2)73jb = jarque_bera(res.resid)74r2 = dg.linear_reset(res, power=2, use_f=True)75r3 = dg.linear_reset(res, power=3, use_f=True)76r3c = dg.linear_reset(res, power=3, use_f=False)77r4 = dg.linear_reset(res, power=4, use_f=True)78ols["diag"] = {79 "bp": list(bp), "bp_original": list(bp0), "white": list(wh), "dw": durbin_watson(res.resid),80 "bg1": list(bg1), "bg3": list(bg3),81 "lb_lags": [1, 5, 10], "lb_stat": lb["lb_stat"].values, "lb_p": lb["lb_pvalue"].values,82 "bp_stat": lb["bp_stat"].values, "bp_p": lb["bp_pvalue"].values,83 "lb5_df2": [lb_df["lb_stat"].values[0], lb_df["lb_pvalue"].values[0]],84 "jb": list(jb),85 "reset2": [float(r2.fvalue), float(r2.pvalue)], "reset3": [float(r3.fvalue), float(r3.pvalue)],86 "reset3_chi2": [float(r3c.statistic), float(r3c.pvalue)], "reset4": [float(r4.fvalue), float(r4.pvalue)],87}88out["ols"] = ols8990# no intercept91r0 = sm.OLS(y, X).fit()92out["ols_noconst"] = {**summary(r0), "rsquared": r0.rsquared, "rsquared_adj": r0.rsquared_adj, "llf": r0.llf, "aic": r0.aic, "bic": r0.bic,93 "df_model": r0.df_model, "df_resid": r0.df_resid, "tss": r0.uncentered_tss, "HC1": summary(sm.OLS(y, X).fit(cov_type="HC1"))}94# simple regression (one regressor): closed forms95rs = sm.OLS(y, sm.add_constant(x1)).fit()96out["ols_simple"] = {**summary(rs), "rsquared": rs.rsquared, "white": list(dg.het_white(rs.resid, sm.add_constant(x1)))}9798# ---------------------------------------------------------------- WLS ---------------------------99w = 1.0 / (0.5 + 0.15 * x1) ** 2100rw = sm.WLS(y, Xc, weights=w).fit()101iw = OLSInfluence(rw)102out["wls"] = {103 "w": w, **summary(rw), "rsquared": rw.rsquared, "rsquared_adj": rw.rsquared_adj, "llf": rw.llf, "aic": rw.aic, "bic": rw.bic,104 "mse_resid": rw.mse_resid, "ssr": rw.ssr, "tss": rw.centered_tss, "resid": rw.resid, "fitted": rw.fittedvalues,105 # statsmodels' OLSInfluence is not reliable for WLS (its hat diagonal matches neither the weighted nor the106 # unweighted projection), so the reference leverage is computed directly: diag of the projector on √w·X.107 "hat": np.diag((Xc * np.sqrt(w)[:, None]) @ np.linalg.inv((Xc * w[:, None]).T @ Xc) @ (Xc * np.sqrt(w)[:, None]).T),108 "HC1": summary(sm.WLS(y, Xc, weights=w).fit(cov_type="HC1")),109 "HC3": summary(sm.WLS(y, Xc, weights=w).fit(cov_type="HC3")),110}111112# ---------------------------------------------------------------- categorical -------------------113region = rng.choice(["north", "south", "east", "west"], size=n)114size = rng.choice(["S", "M", "L"], size=n)115effect = pd.Series(region).map({"north": 0.0, "south": 2.0, "east": -1.0, "west": 4.0}).values116yc = 1 + 0.7 * x1 + effect + rng.normal(0, 1, n)117D = pd.get_dummies(pd.DataFrame({"region": region, "size": size}), drop_first=True).astype(float)118Xd = sm.add_constant(np.column_stack([x1, D.values]))119rc = sm.OLS(yc, Xd).fit()120out["categorical"] = {"y": yc, "x1": x1, "region": list(region), "size": list(size), "columns": list(D.columns), **summary(rc), "rsquared": rc.rsquared}121122# ---------------------------------------------------------------- panel -------------------------123from linearmodels.panel import PanelOLS124125126def panel_case(N, T, drop):127 ent = np.repeat(np.arange(N), T)128 tim = np.tile(np.arange(2000, 2000 + T), N)129 a = rng.normal(0, 2, N)[ent]130 g = np.linspace(0, 1.5, T)[tim - 2000]131 p1 = rng.normal(0, 1, N * T) + 0.5 * a132 p2 = rng.normal(0, 1, N * T) + 0.3 * g133 u = rng.normal(0, 1, N * T) * (1 + 0.3 * np.abs(p1))134 yy = 1 + 1.2 * p1 - 0.7 * p2 + a + g + u135 keep = np.ones(N * T, bool)136 if drop:137 keep[rng.choice(N * T, size=drop, replace=False)] = False138 ent, tim, p1, p2, yy = ent[keep], tim[keep], p1[keep], p2[keep], yy[keep]139 df = pd.DataFrame({"y": yy, "p1": p1, "p2": p2, "ent": ent, "tim": tim}).set_index(["ent", "tim"])140 case = {"y": yy, "p1": p1, "p2": p2, "entity": ent, "time": tim}141 for label, te in [("oneway", False), ("twoway", True)]:142 mod = PanelOLS(df["y"], df[["p1", "p2"]], entity_effects=True, time_effects=te)143 r = mod.fit()144 rr = mod.fit(cov_type="robust")145 rcl = mod.fit(cov_type="clustered", cluster_entity=True)146 # LSDV with statsmodels147 dummies = [pd.get_dummies(ent, drop_first=False).astype(float).values]148 if te:149 dummies.append(pd.get_dummies(tim, drop_first=True).astype(float).values)150 Xl = np.column_stack([p1, p2] + dummies)151 ls = sm.OLS(yy, Xl).fit()152 lsc = sm.OLS(yy, Xl).fit(cov_type="cluster", cov_kwds={"groups": ent})153 lsct = sm.OLS(yy, Xl).fit(cov_type="cluster", cov_kwds={"groups": ent}, use_t=True)154 case[label] = {155 "params": r.params.values, "se": r.std_errors.values, "t": r.tstats.values, "cov": r.cov.values,156 "df_resid": r.df_resid, "s2": r.s2, "rsquared": r.rsquared, "ssr": r.resid_ss,157 "f": r.f_statistic.stat, "f_p": r.f_statistic.pval, "f_pooled": r.f_pooled.stat, "f_pooled_p": r.f_pooled.pval,158 "f_pooled_df": r.f_pooled.df,159 "effects": r.estimated_effects.values.ravel(), "resid": r.resids.values,160 "se_robust": rr.std_errors.values, "f_robust": rr.f_statistic_robust.stat, "f_robust_p": rr.f_statistic_robust.pval,161 "se_cluster": rcl.std_errors.values, "f_cluster": rcl.f_statistic_robust.stat,162 "lsdv_params": ls.params[:2], "lsdv_se": ls.bse[:2], "lsdv_p": ls.pvalues[:2], "lsdv_ci": ls.conf_int()[:2],163 "lsdv_cluster_se": lsc.bse[:2], "lsdv_cluster_p": lsc.pvalues[:2],164 "lsdv_cluster_t_p": lsct.pvalues[:2], "lsdv_cluster_t_ci": lsct.conf_int()[:2],165 "lsdv_entity_effects": ls.params[2:2 + len(np.unique(ent))],166 }167 return case168169170out["panel_balanced"] = panel_case(12, 6, 0)171out["panel_unbalanced"] = panel_case(15, 7, 17)172173# ---------------------------------------------------------------- logit -------------------------174m = 250175z1 = rng.normal(0, 1, m)176z2 = rng.normal(1, 2, m)177z3 = (rng.uniform(size=m) < 0.4).astype(float)178eta = -0.5 + 1.1 * z1 - 0.45 * z2 + 0.9 * z3179yb = (rng.uniform(size=m) < 1 / (1 + np.exp(-eta))).astype(float)180Z = np.column_stack([z1, z2, z3])181lg = sm.Logit(yb, sm.add_constant(Z)).fit(disp=0, method="newton", tol=1e-13, maxiter=200)182ci = lg.conf_int()183lg0 = sm.Logit(yb, Z).fit(disp=0, method="newton", tol=1e-13, maxiter=200)184out["logit"] = {185 "y": yb, "X": Z, "params": lg.params, "bse": lg.bse, "z": lg.tvalues, "p": lg.pvalues, "ci_lo": ci[:, 0], "ci_hi": ci[:, 1],186 "cov": lg.cov_params(), "llf": lg.llf, "llnull": lg.llnull, "prsquared": lg.prsquared, "llr": lg.llr, "llr_pvalue": lg.llr_pvalue,187 "aic": lg.aic, "bic": lg.bic, "pred_table_05": lg.pred_table(0.5), "pred_table_03": lg.pred_table(0.3), "fitted": lg.predict(),188 "predict_X": [[0.2, 1.0, 1.0], [-1.5, 3.0, 0.0]], "predict": lg.predict(np.array([[1, 0.2, 1.0, 1.0], [1, -1.5, 3.0, 0.0]])),189 "noconst_params": lg0.params, "noconst_bse": lg0.bse, "noconst_llf": lg0.llf,190}191192dump("regression", out)193