Skip to main content

Chapter_13

Aaron L9/17/25About 4 minCausal inferenceDeepSeek-R1-0528response_md
### Chapter 13: STANDARDIZATION AND THE PARAMETRIC G-FORMULA **Introduction** This chapter describes standardization as an alternative to IP weighting for estimating the average causal effect of smoking cessation (A) on weight gain (Y) using NHEFS observational data (n=1,629 smokers, 1,566 uncensored). Both methods rely on the same identifiability conditions but differ in modeling assumptions.

13.1 Standardization as an alternative to IP weighting

Technical Points:

  • Core Theory: Standardization estimates the average causal effect (E[Y^{a=1,c=0}] - E[Y^{a=0,c=0}]) by adjusting for confounders (L) (sex, age, race, education, smoking intensity/duration, physical activity, weight).
  • Key Formula: The standardized mean is (\sum_l E[Y|A=a,C=0,L=l] \times \Pr[L=l]). For continuous (L), this becomes (\int E[Y|A=a,C=0,L=l] , dF_L(l)).
  • Identifiability Conditions: Requires exchangeability, positivity, and consistency conditional on (L).
  • Application: Alternative to IP weighting for high-dimensional data with nondichotomous treatments.

Fine Points:

  • Positivity Impact: Deviations from positivity affect IP weighting and standardization differently. Standardization allows parametric extrapolation over structural zeroes but may introduce bias (Fine Point 13.1).
  • Data Context: Quitters ((A=1), n=403) and non-quitters ((A=0), n=1,163) differ in covariate distribution; observed associational difference (2.5 kg) ≠ causal effect.

13.2 Estimating the mean outcome via modeling

Technical Points:

  • Modeling Approach: For high-dimensional (L), parametric models (e.g., linear regression) estimate (E[Y|A=a,C=0,L=l]).
  • Model Specification: Includes linear/quadratic terms for continuous covariates (age, weight) and product terms (e.g., (A \times) smoking intensity).
  • Key Formula: Standardized mean is (E[E[Y|A=a,C=0,L]]), estimated via (\frac{1}{n} \sum_{i=1}^n \hat{E}[Y|A=a,C=0,L_i]).

Fine Points:

  • Nonparametric Limitation: Impractical for millions of strata; parametric smoothing is necessary.
  • Example Estimate: Predicted weight gain for a specific covariate combination (e.g., non-quitter, male, age 26) was 0.34 kg.

13.3 Standardizing the mean outcome to the confounder distribution

Technical Points:

  • Estimation Method:
    1. Expand dataset into three blocks: original data, all untreated ((A=0)), all treated ((A=1)).
    2. Fit outcome model (E[Y|A,L]) using original data.
    3. Predict outcomes for (A=0) and (A=1) blocks.
    4. Average predictions: (\text{mean}(Y|A=0)) and (\text{mean}(Y|A=1)).
  • Result: Standardized mean difference = 5.18 kg (treated) - 1.66 kg (untreated) = 3.5 kg.

Fine Points:

  • Bootstrap Confidence Interval: 95% CI [2.6, 4.5] kg via nonparametric bootstrap (Technical Point 13.1).
  • Efficiency: Uses empirical distribution of (L), avoiding explicit modeling of (\Pr[L=l]).

13.4 IP weighting or standardization?

Technical Points:

  • Equivalence: IP weighted and standardized means are equal only under nonparametric estimation.
  • Model Misspecification:
    • IP weighting models (\Pr[A=a,C=0|L]) (treatment model).
    • Standardization models (E[Y|A=a,C=0,L]) (outcome model).
  • Doubly Robust Estimators: Combine both models (e.g., augmented IP weighting, plug-in estimators) for consistency if either model is correct (Fine Point 13.2, Technical Points 13.2–13.3).

Fine Points:

  • Practical Advice: Use both methods; agreement suggests robustness to misspecification.
  • g-Formula: Standardization is a plug-in estimator of the parametric g-formula, generalizing to time-varying treatments.

13.5 How seriously do we take our estimates?

Technical Points:

  • Validity Requirements:
    1. Identifiability (exchangeability, positivity, consistency).
    2. No measurement error.
    3. Correct model specification.
  • Sensitivity Analysis: Essential for unverifiable assumptions (e.g., unmeasured confounding, selection bias).

Fine Points:

  • Caveats:
    • Effect estimate (3.5 kg) assumes no unmeasured confounding, perfect measurement, and correct models.
    • Real-world deviations (e.g., vague treatment versions, mismeasurement) may bias estimates.
  • Expert Judgment: Causal interpretations rely on untestable assumptions; skepticism and sensitivity analyses are critical.

Technical and Fine Points Summaries

Fine Point 13.1 (Structural Positivity)

  • Key Insight: Standardization tolerates structural non-positivity via parametric extrapolation, but biases confidence intervals. Standard errors are smaller than IP weighting under positivity violations.

Technical Point 13.1 (Bootstrapping)

  • Method: Resample with replacement (1,000 samples) to estimate standard errors for confidence intervals.
  • Application: Used for the standardized mean difference (95% CI: [2.6, 4.5] kg).

Fine Point 13.2 (Doubly Robust Plug-in Estimator)

  • Method: Combines IP weights (W^A = 1/f(A|L)) and outcome model with covariate (R = W^A \cdot I(A=1) - W^A \cdot I(A=0)).
  • Advantage: Consistent if either treatment or outcome model is correct.

Technical Point 13.2 (Augmented IP Weighted Estimator)

  • Formula:

    E^[Ya=1]DR=1n∑i=1n[b^(Li)+Aiπ^(Li)(Yi−b^(Li))]\hat{E}[Y^{a=1}]_{DR} = \frac{1}{n} \sum_{i=1}^n \left[ \hat{b}(L_i) + \frac{A_i}{\hat{\pi}(L_i)} (Y_i - \hat{b}(L_i)) \right]

  • Property: Second-order bias allows machine learning for high-dimensional (L).

Technical Point 13.3 (AIPW and TMLE)

  • Relationship: Augmented IPW and targeted minimum loss-based estimation (TMLE) are equivalent under specific model constraints. TMLE uses "clever covariates" for doubly robust plug-in estimation.

Note: All conclusions are derived directly from the source text. Assumptions (e.g., exchangeability) are untestable; sensitivity analyses are recommended.