Robust inference for risk heterogeneity under group imbalance

Technical Summary: Robust Inference for Risk Heterogeneity under Group Imbalance

Problem Statement
In biomedical data, population-level heterogeneity is ubiquitous, where risk patterns often differ significantly across demographic or clinical subgroups. A critical challenge arises when analyzing these disparities in the presence of severe group imbalance, where one subgroup (the minority) is underrepresented compared to another (the majority). Existing approaches for detecting risk heterogeneity, such as standard likelihood-based methods or linear adjustment frameworks (e.g., Maity et al., 2024), face two primary limitations in this context:

1. Sensitivity to Model Misspecification: These methods often rely on a baseline model for the majority group. If this baseline model is misspecified, the resulting estimates of heterogeneity are biased.
2. Regularization Bias: In high-dimensional settings, regularized estimation (e.g., Lasso) is necessary but introduces bias by shrinking coefficients toward zero. This is particularly problematic for the minority group, where sample sizes are small, and the bias from the majority group's nuisance parameter estimation propagates to the heterogeneity estimates.

The authors aim to develop an inference framework capable of handling covariate-adjusted heterogeneity across subgroups with large numbers of covariates, potential underrepresentation, and non-linear associations, while providing reliable statistical inference (confidence intervals and p-values).

Methodology
The paper proposes a robust inference framework based on Neyman orthogonality (Chernozhukov et al., 2018). The core objective is to estimate the parameter vector $\beta$ representing the difference in risk between a minority group ($T=0$) and a majority group ($T=1$), conditional on covariates $X$.

• Model Formulation: The heterogeneity is modeled as a linear function of covariates on the link-transformed scale:
  $$\gamma(E[Y|X, T=0]) - \gamma(E[Y|X, T=1]) = \alpha^T X$$
  Here, $\gamma$ is a prespecified link function (e.g., logit for binary outcomes). The parameter of interest is a subset $\beta$ of $\alpha$, while the baseline risk function $\xi(X)$ for the majority group and the remaining coefficients $\delta$ are treated as nuisance parameters.

• Neyman Orthogonal Score: To achieve robustness against estimation errors in nuisance parameters ($\xi, \delta, \pi$), the authors construct a score function $S(Z, \beta, \delta, \xi, \pi)$ that satisfies the Neyman orthogonality condition. This condition ensures that the score function is locally insensitive to small perturbations in the nuisance parameters. The score is derived by modifying the linear adjustment approach to include a correction term involving the derivative of the loss function with respect to the nuisance parameters.

• Estimation Procedure:

  1. Nuisance Estimation: Estimate the baseline model $\hat{\xi}$ using majority group data, and estimate the nuisance coefficients $\hat{\delta}$ and propensity score $\hat{\pi}$ using appropriate regularized or machine learning methods.
  2. Debiased Estimation: Solve the score equation $\sum S(Z_i, \beta, \hat{\delta}, \hat{\xi}, \hat{\pi}) = 0$ for $\beta$. Since this equation is implicit, an iterative algorithm (Algorithm 1) is provided to approximate the solution.
  3. Variance Estimation: The asymptotic variance is estimated using the sandwich estimator formula, allowing for the construction of valid confidence intervals.

• Theoretical Guarantees: Under assumptions of bounded covariates and specific convergence rates for the nuisance parameter estimators (which can be slower than $\sqrt{N}$), the proposed debiased estimator $\hat{\beta}_{debias}$ is shown to be consistent and asymptotically normal. The effective sample size for inference is determined by the size of the minority group.

Key Contributions

1. Robust Inference Tool: The framework provides a method to perform statistical inference on risk heterogeneity that is robust to both baseline model misspecification and regularization bias.
2. Handling Imbalance: It explicitly accommodates scenarios where subgroups are severely underrepresented, a common issue in clinical datasets.
3. Flexibility: The method is agnostic to the specific algorithm used for the baseline model $\xi(\cdot)$, provided it meets convergence rate conditions. It supports a wide range of outcome types (binary, continuous, count) via generalized linear models.
4. Debiasing Mechanism: Unlike standard two-step procedures, this approach unifies the correction for regularization bias and nuisance parameter bias through the Neyman orthogonal score.

Results

• Simulation Studies:

  • The authors simulated data with $N=30,000$ and varying minority group proportions ($10\%, 20\%, 40\%$).
  • Bias Reduction: The proposed debiased estimator achieved near-zero bias across all levels of group imbalance and baseline model choices (Penalized GLM, XGBoost, Neural Networks). In contrast, the standard linear adjustment estimator exhibited substantial bias, particularly when the minority group was small or the baseline model was misspecified.
  • Inference Validity: The 95% confidence intervals constructed from the debiased estimator achieved empirical coverage close to the nominal 95% level. QQ-plots confirmed the asymptotic normality of the estimator.
  • Variance: As expected, variance increased as the minority group proportion decreased, but the estimator remained stable.

• Data Application (eICU Database):

  • The method was applied to the eICU Collaborative Research Database to analyze in-hospital mortality differences between Caucasian patients and underrepresented groups (Asian, Hispanic, Native American).
  • Findings: Standard linear adjustment methods (using Lasso) shrank all heterogeneity coefficients to zero, failing to detect significant differences.
  • Discovery: The proposed method identified clinically meaningful heterogeneities that standard methods missed:
    • Native American patients admitted with Overdose had significantly higher adjusted log-odds of mortality compared to Caucasians.
    • Asian patients admitted with Cardiac Arrest showed significantly higher mortality risk.
    • Hispanic patients admitted with Diabetic Ketoacidosis (DKA) exhibited significantly higher mortality risk.
  • These findings align with existing literature regarding specific ethnic disparities in these conditions.

Significance and Claims
The paper claims that the proposed framework offers a critical advancement for precision medicine and health equity research. By enabling reliable inference in the presence of group imbalance and high-dimensional covariates, the method allows researchers to uncover "clinically meaningful" heterogeneities that are otherwise obscured by the bias inherent in standard regularized approaches. The authors emphasize that their approach does not require perfect specification of the majority group's risk model, making it a practical and robust tool for real-world biomedical data analysis where such models are often complex and imperfect. The methodology extends beyond binary outcomes, suggesting broad applicability to continuous and count data in healthcare settings.