Technical Summary: Inverse Probability Weighting in a Post-Bayesian World

Problem Statement

The paper addresses the challenge of selection bias in Bayesian inference. In many practical scenarios, researchers possess a large-scale dataset ($X$, size $n$) drawn from a biased distribution $G$ (e.g., registry data), while a smaller, high-quality dataset ($X'$, size $m$) is available from an unbiased target distribution $Q$ (e.g., a clinical cohort).

Standard Bayesian updating on the large dataset $X$ minimizes the Kullback-Leibler (KL) divergence between the model and the biased distribution $G$, leading to posterior estimates that are systematically biased relative to the target population $Q$. While Inverse Probability Weighting (IPW) is a well-established frequentist technique for correcting such biases, its integration into Bayesian frameworks has been historically difficult. Traditional Bayesian methods rely on pure likelihood and priors, and the introduction of external weights appears to contradict Bayes' rule and its theoretical guarantees (e.g., consistency, asymptotic normality).

Methodology

The authors propose a Generalised Bayesian framework that incorporates IPW to redefine the inferential target. The core methodology involves three steps:

1. Ratio Estimation via Classification:
   Instead of assuming known densities, the authors train a classifier (e.g., logistic regression) to distinguish between samples from the biased distribution $G$ and the unbiased distribution $Q$. This classifier estimates the density ratio $r(x) = q(x)/g(x)$ for each data point $x_i$ in the large dataset. This ratio represents the odds of a data point belonging to the ideal population relative to the biased population.

2. Reweighted Likelihood Update:
   The standard Bayesian posterior is modified by reweighting the likelihood contribution of each data point $x_i$ by the estimated ratio $w_i = r(x_i)$. The generalised posterior $\pi_n(\theta)$ is defined as:
   $$\pi_n(\theta) \propto \pi(\theta) \prod_{i=1}^n p(x_i|\theta)^{w_i}$$
   On the log scale, this corresponds to minimizing a weighted empirical risk:
   $$\log \pi_n(\theta) = \log \pi(\theta) - n L_n(\theta; w)$$
   where $L_n(\theta; w)$ is the weighted negative log-likelihood.

3. Theoretical Justification (KL Projection):
   The authors demonstrate that this reweighting is mathematically equivalent to shifting the inference target. Conventional Bayesian inference minimizes $KL(g(\cdot) || p(\cdot|\theta))$. By applying weights $w_i \approx q(x_i)/g(x_i)$, the procedure minimizes:
   $$L(\theta; w) = E_G [r(X)(-\ell_\theta(X))] = -E_Q[\ell_\theta(X)] = KL(q(\cdot) || p(\cdot|\theta)) + c$$
   Thus, the posterior converges to the parameter $\theta^*$ that minimizes the KL divergence between the model and the unbiased target distribution $Q$, rather than the biased source $G$.

4. Practical Stabilization:
   To address instability caused by extreme weights (which can reduce the Effective Data Size, EDS), the authors recommend:

   • Clipping: Truncating extreme log-odds (e.g., $\pm 10$).
   • Tempering: Using a tempered ratio $r(x)^\alpha$ (with $0 < \alpha < 1$) to stabilize inference, albeit at the cost of targeting a slightly different distribution.
   • Cross-fitting: Using K-fold cross-validation to estimate weights to prevent overfitting and ensure data independence between weight estimation and posterior updating.

Key Theoretical Contributions

The paper provides rigorous theoretical results under standard M-estimation assumptions (regularity, identifiability, Glivenko-Cantelli conditions):

• Existence and Uniqueness: The reweighted population risk function attains a unique global minimum under convexity and coercivity conditions.
• Posterior Concentration: The generalised posterior concentrates asymptotically around the true minimiser $\theta^*$ of the reweighted risk.
• Asymptotic Normality (Bernstein-von Mises): The posterior distribution is asymptotically normal. Crucially, the authors distinguish between the posterior variance and the frequentist sampling variance:
  • The posterior covariance is $I^{-1}/n$ (where $I$ is the expected Hessian).
  • The frequentist "sandwich" variance of the estimator is $I^{-1}JI^{-1}/n$ (where $J$ involves the variance of the weighted scores).
  • Under reweighting, $J \neq I$ generally. Therefore, to achieve valid frequentist credible intervals, the authors advocate using the sandwich-corrected covariance rather than the raw posterior variance.
• Plug-in Weight Effects: The paper analyzes the impact of estimating weights from a finite sample of size $m$. If $m/n \to \infty$, the weight estimation noise is negligible. If $m/n \to c \in (0, \infty)$, an additional variance inflation term $K_c$ appears in the asymptotic distribution, which must be accounted for.

Empirical Results

The authors validate the approach through three examples:

1. Gaussian Simulation: Inference on the mean and variance of a Gaussian model where data is drawn from a biased distribution.

   • Result: Standard Bayes exhibited significant bias. The exact odds reweighting successfully debiased the estimates but showed higher variance and lower stability (lower EDS). Tempered odds ($\alpha=0.75$) offered a trade-off with slightly higher bias but improved stability.

2. Multivariable Regression (Simulated Registry): A linear regression model with selection bias where patients with higher outcomes were less likely to be included in the registry.

   • Result: The odds reweighting consistently produced the smallest bias for regression coefficients, closely matching the ground truth derived from the unbiased sample. It achieved near-perfect coverage for specific parameters, though with higher variance than tempered methods.

3. Real-World Application (Prostate Cancer): Predicting 10-year prostate cancer mortality using PSA levels from Norwegian registry data.

   • Context: The registry data was biased towards older men, whereas the target was the general population of men aged 39+.
   • Result: The method successfully projected the inference from the biased registry to the general population distribution. The authors note that while the true underlying distribution of cancer onset is unobserved, the method provided a principled upper bound on the predictive influence of PSA, complementing the lower bound provided by the raw registry data.

Significance and Claims

The paper claims to provide a coherent justification for using IPW within a post-Bayesian framework. Its primary contributions are:

• Reframing Bias Correction: It reinterprets frequentist IPW not as an ad-hoc modification of Bayes' rule, but as a mechanism to shift the KL-divergence minimization target from the observed data distribution to a user-specified target distribution.
• Bridging Paradigms: It unifies generalised Bayesian inference (which relaxes strict model assumptions) with frequentist bias-correction techniques, allowing Bayesian methods to handle selection bias in large-scale "Big N" datasets.
• Practical Utility: The authors demonstrate that the method is immediately applicable using standard software (e.g., brms in R) and off-the-shelf classifiers, making it accessible for epidemiological and causal inference problems previously considered intractable for Bayesian approaches due to selection bias.
• Theoretical Rigor: By establishing asymptotic normality and providing the correct sandwich variance for credible intervals, the paper ensures that the method maintains frequentist validity while retaining the interpretability of Bayesian posteriors.

The authors remain modest regarding the scope, acknowledging that optimal classifier design, weight stability in small samples, and extensions to non-i.i.d. models (e.g., time series) require further investigation. They position this work as a foundational step toward extending Bayesian reasoning beyond traditional model misspecification to include data collection biases.