First Course Ding: Chapter 11

The central role of the propensity score

This translation keeps the chapter’s basic workflow:

Show code 
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

import crabbymetrics as cm

np.set_printoptions(precision=4, suppress=True)


def repo_root():
    for candidate in [Path.cwd().resolve(), *Path.cwd().resolve().parents]:
        if (candidate / "ding_w_source").exists():
            return candidate
    raise FileNotFoundError("could not locate ding_w_source from the current working directory")


def expit(x):
    return 1.0 / (1.0 + np.exp(-x))

1 NHANES School-Meal Example

Show code 
data = pd.read_csv(repo_root() / "ding_w_source" / "repl" / "nhanes_bmi.csv")

y = data["BMI"].to_numpy(dtype=float)
d = data["School_meal"].to_numpy(dtype=np.int32)
feature_names = [
    "age",
    "ChildSex",
    "black",
    "mexam",
    "pir200_plus",
    "WIC",
    "Food_Stamp",
    "fsdchbi",
    "AnyIns",
    "RefSex",
    "RefAge",
]
x = data[feature_names].to_numpy(dtype=float)
x = (x - x.mean(axis=0)) / np.where(x.std(axis=0) > 1e-12, x.std(axis=0), 1.0)

naive = cm.OLS()
naive.fit(d.astype(float)[:, None], y)

adjusted = cm.OLS()
adjusted.fit(np.column_stack([d.astype(float), x]), y)

lin_design = np.column_stack([d.astype(float), x, d[:, None].astype(float) * x])
lin = cm.OLS()
lin.fit(lin_design, y)

ps_model = cm.Logit(alpha=1.0, max_iterations=300)
ps_model.fit(x, d)
ps_summary = ps_model.summary()
pscore = expit(ps_summary["intercept"] + x @ np.asarray(ps_summary["coef"]))

quantiles = np.quantile(pscore, [0.2, 0.4, 0.6, 0.8])
quintile = np.digitize(pscore, quantiles, right=True)


def stratified_ate(y, d, strata):
    total = 0.0
    for level in np.unique(strata):
        idx = strata == level
        if np.any(d[idx] == 1) and np.any(d[idx] == 0):
            total += idx.mean() * (y[idx & (d == 1)].mean() - y[idx & (d == 0)].mean())
    return float(total)


stratified = stratified_ate(y, d, quintile)


def stratified_neyman(y, d, strata):
    estimate = 0.0
    variance = 0.0
    for level in np.unique(strata):
        idx = strata == level
        treated_k = idx & (d == 1)
        control_k = idx & (d == 0)
        if not (treated_k.any() and control_k.any()):
            continue
        weight = idx.mean()
        tau_k = y[treated_k].mean() - y[control_k].mean()
        var_k = y[treated_k].var(ddof=1) / treated_k.sum() + y[control_k].var(ddof=1) / control_k.sum()
        estimate += weight * tau_k
        variance += weight**2 * var_k
    return estimate, np.sqrt(variance)


def strata_from_score(score, n_strata):
    cutpoints = np.quantile(score, np.arange(1, n_strata) / n_strata)
    return np.digitize(score, cutpoints, right=True)


stratification_rows = []
for n_strata in [5, 10, 20, 50, 80]:
    est, se = stratified_neyman(y, d, strata_from_score(pscore, n_strata))
    stratification_rows.append({"strata": n_strata, "estimate": est, "se": se})


def ipw_estimates(y, d, pscore, lower=0.0, upper=1.0):
    p = np.clip(pscore, lower, upper)
    ht = float(np.mean(d * y / p - (1 - d) * y / (1 - p)))
    hajek = float(
        np.mean(d * y / p) / np.mean(d / p)
        - np.mean((1 - d) * y / (1 - p)) / np.mean((1 - d) / (1 - p))
    )
    return ht, hajek


ipw_rows = []
for label, bounds in {
    "none": (1e-6, 1.0 - 1e-6),
    "0.01-0.99": (0.01, 0.99),
    "0.05-0.95": (0.05, 0.95),
    "0.10-0.90": (0.10, 0.90),
}.items():
    ht, hajek = ipw_estimates(y, d, pscore, *bounds)
    ipw_rows.append({"truncation": label, "HT": ht, "Hajek": hajek})

treated = d == 1
control = ~treated
bal = cm.BalancingWeights(objective="entropy", autoscale=True, max_iterations=300, tolerance=1e-8, l2_norm=0.02)
bal.fit(x[control], x[treated])
weights = np.asarray(bal.summary()["weights"])
att_bal = float(y[treated].mean() - np.dot(weights, y[control]))

pd.DataFrame(
    {
        "estimate": [
            float(naive.summary()["coef"][0]),
            float(adjusted.summary()["coef"][0]),
            float(lin.summary()["coef"][0]),
            stratified,
            att_bal,
        ]
    },
    index=[
        "Naive OLS",
        "Fisher OLS",
        "Lin OLS",
        "Propensity-stratified ATE",
        "Entropy-balancing ATT",
    ],
)
estimate
Naive OLS 0.533904
Fisher OLS 0.061248
Lin OLS -0.016954
Propensity-stratified ATE -0.115245
Entropy-balancing ATT -0.184490

2 Overlap

Show code 
fig, ax = plt.subplots(figsize=(6, 4))
ax.hist(pscore[d == 1], bins=25, density=True, alpha=0.6, label="Treated")
ax.hist(pscore[d == 0], bins=25, density=True, alpha=0.6, label="Control")
ax.set_xlabel("Estimated propensity score")
ax.set_ylabel("Density")
ax.set_title("Estimated overlap in the NHANES example")
ax.legend()
fig.tight_layout()
plt.show()

Show code 
for level in np.unique(quintile):
    idx = quintile == level
    print(
        f"Quintile {int(level) + 1}: n = {idx.sum():4d}, treated share = {float(d[idx].mean()): .4f}"
    )
Quintile 1: n =  467, treated share =  0.1349
Quintile 2: n =  465, treated share =  0.4172
Quintile 3: n =  466, treated share =  0.5730
Quintile 4: n =  466, treated share =  0.7725
Quintile 5: n =  466, treated share =  0.8584

3 Stratification And IPW

The R script reports propensity-score subclassification for several numbers of strata, then compares Horvitz-Thompson and Hajek IPW under increasing propensity-score truncation. The estimates below use the same formulas.

Show code 
pd.DataFrame(stratification_rows).set_index("strata")
estimate se
strata
5 -0.115245 0.283263
10 -0.177911 0.282360
20 -0.192190 0.279109
50 -0.287762 0.277245
80 -0.196331 NaN
Show code 
pd.DataFrame(ipw_rows).set_index("truncation")
HT Hajek
truncation
none -1.497137 -0.151211
0.01-0.99 -1.497137 -0.151211
0.05-0.95 -1.481276 -0.147388
0.10-0.90 -0.712506 -0.051208

4 Balance Check

The balance diagnostic asks the same question as the R plots: after reducing the covariates to a scalar propensity score, do the residual treated-control differences in the original covariates shrink?

Show code 
def hajek_balance(column, d, pscore):
    p = np.clip(pscore, 1e-6, 1.0 - 1e-6)
    treated_mean = np.mean(d * column / p) / np.mean(d / p)
    control_mean = np.mean((1 - d) * column / (1 - p)) / np.mean((1 - d) / (1 - p))
    return float(treated_mean - control_mean)


def stratified_balance(column, d, strata):
    total = 0.0
    for level in np.unique(strata):
        idx = strata == level
        treated_k = idx & (d == 1)
        control_k = idx & (d == 0)
        if treated_k.any() and control_k.any():
            total += idx.mean() * (column[treated_k].mean() - column[control_k].mean())
    return float(total)


balance_rows = []
for j, name in enumerate(feature_names):
    balance_rows.append(
        {
            "covariate": name,
            "unadjusted": float(x[treated, j].mean() - x[control, j].mean()),
            "stratified": stratified_balance(x[:, j], d, quintile),
            "hajek_ipw": hajek_balance(x[:, j], d, pscore),
        }
    )

balance = pd.DataFrame(balance_rows)
balance
covariate unadjusted stratified hajek_ipw
0 age 0.055745 0.000400 -0.028061
1 ChildSex -0.026365 -0.003588 -0.002667
2 black 0.250013 -0.004957 -0.003932
3 mexam 0.341891 0.004654 -0.003997
4 pir200_plus -0.836903 -0.017041 0.011753
5 WIC 0.370816 0.011726 -0.030595
6 Food_Stamp 0.710649 0.066151 -0.044231
7 fsdchbi 0.411863 0.003839 -0.041801
8 AnyIns -0.138846 0.004704 -0.016292
9 RefSex -0.219006 -0.040892 -0.005085
10 RefAge -0.171938 -0.023001 -0.000938
Show code 
fig, ax = plt.subplots(figsize=(9, 4))
xpos = np.arange(len(feature_names))
width = 0.25
for offset, column, label in [
    (-width, "unadjusted", "Unadjusted"),
    (0.0, "stratified", "5-strata"),
    (width, "hajek_ipw", "Hajek IPW"),
]:
    ax.bar(xpos + offset, np.abs(balance[column]), width=width, label=label)
ax.axhline(0.1, color="black", linestyle="--", linewidth=1.0)
ax.set_xticks(xpos)
ax.set_xticklabels([f"x{j + 1}" for j in xpos])
ax.set_ylabel("Absolute mean difference on standardized covariates")
ax.set_title("Propensity-score balance diagnostics")
ax.legend()
fig.tight_layout()

The chapter’s message is still the right one:

  • the outcome regression can move after covariate adjustment
  • the propensity score is a one-dimensional overlap diagnostic
  • once the score is estimated, blocking and weighting become natural follow-up estimators