"""Re-run the key tables of Reinstein & Snyder (JIE 2005) in Python. Tables I-V. Specifications are copied from the authors' Stata files original_code/revision4.do (Tables IV, V) and original_code/SELECT2.DO (Table III). Ports `reg ..., robust cluster(weekend)` and `probit ..., robust cluster(weekend)` to statsmodels. SCREENS is rescaled to thousands to match the paper's units. Run from the package root: python3 replication/replicate.py data/movie.dta Writes replication/output/keycoefs_tables4_5.csv """ import sys from pathlib import Path import numpy as np import pandas as pd import statsmodels.api as sm pd.set_option("display.width", 250) pd.set_option("display.max_columns", 20) d = pd.read_stata(sys.argv[1], convert_categoricals=False) d["lwk1cpi"] = np.log(d.wk1cpi) d["lwk1per"] = np.log(d.wk1cpi / d.openscr) d["lgross"] = np.log(d.grosscpi) d["screens"] = d.openscr / 1000 d["weekend"] = d.opendate + 6 - d.openwday d = d[d.pre != 1].copy() print("N after dropping pre-opening reviews:", len(d)) GENRE = "advent animated children comedy crime document drama fantasy noir horror musical mystery romance scifi thrill war western".split() PROD = "disney fox sony mgm_ua gram orion univers warner paramnt".split() YEARS = [f"year{i}" for i in range(2, 10)] M11 = [f"month{i}" for i in range(1, 12)] M10 = [f"month{i}" for i in range(1, 11)] KEY = ["dur", "oudur", "uudur", "ou", "uu"] def design(df, xvars): X = df[xvars].astype(float) X = X.loc[:, X.std() > 0] # Stata drops all-constant columns X = sm.add_constant(X) # drop exactly collinear columns, as Stata does, keeping earlier ones keep = [] for c in X.columns: trial = X[keep + [c]].values if np.linalg.matrix_rank(trial) == len(keep) + 1: keep.append(c) return X[keep] def ols(df, y, xvars): X = design(df, xvars) res = sm.OLS(df[y].astype(float), X).fit( cov_type="cluster", cov_kwds={"groups": df.weekend.values}) return res def probit(df, y, xvars): df = df.copy() xs = list(xvars) # Stata: a dummy that perfectly predicts the outcome is dropped with its observations changed = True while changed: changed = False for c in xs: v = df[c] if set(v.unique()) <= {0, 1} and v.sum() > 0: outs = df.loc[v == 1, y].unique() if len(outs) == 1: df = df[v != 1] xs.remove(c) changed = True break X = design(df, xs) res = sm.Probit(df[y].astype(float), X).fit( disp=0, cov_type="cluster", cov_kwds={"groups": df.weekend.values}) return res def col(res, names): out = {} for n in names: if n in res.params: out[n] = f"{res.params[n]:6.2f} ({res.bse[n]:.2f}){stars(res.pvalues[n])}" else: out[n] = "" out["N"] = int(res.nobs) return out def stars(p): return "***" if p < .01 else "**" if p < .05 else "*" if p < .10 else "" # ---------------- Table I ---------------- t1 = pd.DataFrame({ "TOTREV": d.grosscpi / 1e6, "OPENREV": d.wk1cpi / 1e6, "SISKEL UP": d.sdummy, "EBERT UP": d.edummy, "ONE UP": d.ou, "TWO UP": d.uu, "DURING": d.dur, "SCREENS": d.screens, "FOURDAY": d.fourday, "MALTIN": d.maltin}) print("\nTABLE I (paper means: 30.6 6.6 .32 .41 .31 .21 .81 1.49 .07 2.29)") print(t1.describe().T[["mean", "std", "max", "min"]].round(2)) # ---------------- Table II ---------------- t2 = pd.DataFrame({"lnTOTREV": np.log(d.grosscpi), "lnOPENREV": d.lwk1cpi, "MALTIN": d.maltin, "SISKEL": d.sdummy, "EBERT": d.edummy, "DURING": d.dur}) print("\nTABLE II correlations") print(t2.corr().round(2)) # ---------------- Table III ---------------- base = GENRE + PROD sub = d[d.month != 12] p = { "(1)": probit(sub, "dur", ["lwk1cpi", "sdummy", "edummy"] + base + ["maltin"] + YEARS + ["screens", "fourday"] + M10), "(2)": probit(sub, "dur", ["lwk1cpi", "ou", "uu"] + base + ["maltin"] + YEARS + ["screens", "fourday"] + M10), "(3)": probit(sub, "dur", ["lwk1cpi", "sdummy", "edummy"] + base + YEARS + ["screens", "fourday"] + M10), "(4)": probit(sub, "dur", ["lwk1cpi", "ou", "uu"] + base + YEARS + ["screens", "fourday"] + M10), "(5)": probit(sub, "dur", YEARS + ["fourday"] + M10), } names3 = ["lwk1cpi", "sdummy", "edummy", "ou", "uu", "maltin", "screens", "fourday"] t3 = pd.DataFrame({k: col(v, names3) for k, v in p.items()}) t3.loc["loglik"] = [round(v.llf) for v in p.values()] t3.loc["pseudoR2"] = [round(v.prsquared, 2) for v in p.values()] print("\nTABLE III probit for DURING") print(t3) # ---------------- Table IV ---------------- ctrl_all = ["maltin"] + GENRE + PROD + YEARS + ["screens", "fourday"] + M11 ctrl_gen = ["maltin"] + PROD + YEARS + ["screens", "fourday"] + M11 spec = KEY m4 = { "All": ols(d, "lwk1cpi", spec + ctrl_all), "Wide": ols(d[d.openscr > 1533], "lwk1cpi", spec + ctrl_all), "Narrow": ols(d[d.openscr <= 1533], "lwk1cpi", spec + ctrl_all), "Drama": ols(d[d.drama == 1], "lwk1cpi", spec + ctrl_gen), "Action": ols(d[d.action == 1], "lwk1cpi", spec + ctrl_gen), "Comedy": ols(d[d.comedy == 1], "lwk1cpi", spec + ctrl_gen), } names4 = KEY + ["maltin", "screens", "fourday", "const"] t4 = pd.DataFrame({k: col(v, names4) for k, v in m4.items()}) t4.loc["R2"] = [round(v.rsquared, 2) for v in m4.values()] print("\nTABLE IV ln(OPENREV)") print(t4) # ---------------- Table V ---------------- dr = d[d.drama == 1] m5 = { "(1) lnOPEN": ols(dr, "lwk1cpi", spec + ctrl_gen), "(2) no MALTIN": ols(dr, "lwk1cpi", spec + [c for c in ctrl_gen if c != "maltin"]), "(3) per screen": ols(dr, "lwk1per", spec + [c for c in ctrl_gen if c != "screens"]), "(4) lnTOTREV": ols(dr, "lgross", spec + ctrl_gen), } t5 = pd.DataFrame({k: col(v, names4) for k, v in m5.items()}) t5.loc["R2"] = [round(v.rsquared, 2) for v in m5.values()] print("\nTABLE V dramas") print(t5) # ---------------- machine-readable key coefficients ---------------- rows = [] for tab, models in (("IV", m4), ("V", m5)): for k, r in models.items(): for v in KEY: rows.append(dict(table=tab, model=k, var=v, b=r.params[v], se=r.bse[v], p=r.pvalues[v], N=int(r.nobs), r2=r.rsquared)) OUT = Path(__file__).resolve().parent / "output" OUT.mkdir(exist_ok=True) pd.DataFrame(rows).to_csv(OUT / "keycoefs_tables4_5.csv", index=False)