Prediction-based Inference in Electronic Health Record (EHR)-linked Biobanks

Xingran Chen, Cheng-Han Yang, Zhenke Wu, Bhramar MukherjeearXiv 2026 · DOI · arXiv

TL;DR

This paper evaluates how "prediction-based" (PB) inference—using machine learning to fill in missing lab results—performs in genetic studies using Electronic Health Record (EHR) data. While these methods significantly boost statistical power, the authors reveal they can introduce bias in realistic clinical scenarios where data is "missing not at random," ultimately providing a decision tree to help researchers choose the safest analytical path.

What Problem Does This Paper Address?

Electronic Health Records (EHRs) linked to biobanks are a gold mine for Genome-Wide Association Studies (GWAS). However, EHR data is notoriously "messy." Lab tests (like HbA1c or Vitamin D levels) aren't recorded for everyone; they are ordered by doctors based on clinical suspicion. This creates a "clinically informative" missingness pattern: a patient suspected of diabetes is more likely to have an HbA1c test than a healthy one.

In statistics, this is often Missing Not At Random (MNAR). Traditional methods either throw away the missing data (Complete-Case Analysis) or use Multiple Imputation. Recently, a new class of Prediction-Based (PB) methods has emerged, using external ML models to "rectify" or augment estimates. The problem? We didn't fully understand if these PB methods remain valid when the missingness is driven by the very outcome we are trying to study.

Methods

The authors conducted a structured review of eight recent PB methods and categorized them into three frameworks:

  1. Estimating-Equation Augmentation: Methods like PPI and PPI++ that add a "bias-correction" term directly into the math used to find the association.
  2. Estimator Augmentation: Methods like PS-PPI that calculate a standard estimate and then adjust it using a "rectifier" term based on predicted values.
  3. Joint Modeling: Specifically the SynSurr method, which models the true outcome and the predicted surrogate simultaneously.

They compared these against five baselines (including Complete-Case Analysis and Multiple Imputation) across 10 different simulated observation models and a massive real-world dataset. Crucially, they derived three new theorems to define the exact mathematical conditions under which these methods succeed or fail.

Datasets

  • Simulated Data: 10 different "observation process" models ranging from Missing Completely at Random (MCAR) to complex MNAR scenarios involving interaction effects.
  • All of Us (AoU) Research Program: A large-scale US biobank. The authors performed GWAS on six laboratory biomarkers:
  • White Blood Cell (WBC) count
    • Hemoglobin A1c (HbA1c)
    • Alanine aminotransferase (ALT)
    • Albumin
    • C-reactive protein (CRP)
    • Vitamin D

Key Insights and Ideas

  • The PB "Power Boost": When data is missing at random, PB methods are vastly superior. They achieve much higher statistical power (e.g., ~0.97 in simulations) compared to traditional methods by effectively "reclaiming" the information in the unlabeled (missing) portion of the biobank.
  • The MNAR Trap: PB methods actually require stronger assumptions than simple Complete-Case Analysis (CCA) to stay valid under MNAR. Specifically, the ML model's prediction error must be independent of the reason the data is missing. If the ML model uses features that are also driving the doctor's decision to order the test, the final genetic estimate will be biased.
  • The "Sick Patient" Bias: The paper highlights that "complete-case" subsets in EHRs usually represent sicker individuals. PB methods allow researchers to extend their findings to a more representative, broader population by including those who didn't need the test.
  • A Practical Decision Tree: The authors synthesized their theoretical findings into a decision tree (Figure 2), providing a "cheat sheet" for biobank researchers to determine when to use PB methods vs. when to stick to conservative CCA.

Key Results

  • Efficiency Gains: In the All of Us data, PB methods like SynSurr and PS-PPI consistently reduced standard errors compared to CCA, making the findings more precise.
  • Replication: PB methods successfully replicated known genetic associations for all six biomarkers while identifying additional significant variants that CCA missed (e.g., for ALT and Vitamin D).
  • Naive Imputation Warning: The "Naive" approach—simply plugging in ML predictions and running a standard analysis—was a disaster, showing a 100% Type I error rate (false positives) in almost all scenarios.
  • Robustness of Logistic Regression: Interestingly, for binary outcomes (yes/no), associations were more robust to missingness bias because the bias often gets "absorbed" by other variables in the model.

Limitations and Future Work

  • Cross-Sectional Only: The study focused on a single "snapshot" in time. EHR data is longitudinal (recorded over years), and PB theory for time-series data is still in its infancy.
  • Outcome-Only Missingness: The authors assumed only the biomarker (the $Y$ variable) was missing. In reality, covariates (the $X$ variables) are often missing too.
  • The "Black Box" Dependency: The success of PB methods is heavily tied to the accuracy of the underlying ML model. Future work is needed to develop strategies that improve prediction without compromising the ability to generalize results to the whole population.

Source: Prediction-based Inference in Electronic Health Record (EHR)-linked Biobanks (Chen et al., 2026).

Summary updated

Abstract

Electronic health record (EHR)-linked biobank data facilitate large-scale scientific discoveries such as genome-wide association study (GWAS) on a multitude of phenotypic traits and biomarkers routinely captured in EHR. However, heterogeneous missingness in biomarkers may bias analyses when used as phenotypes. Recently proposed prediction-based (PB) inference methods incorporate external machine ML models to impute missing biomarkers and thereby improve the statistical power and estimation precision in association analyses. Yet, it remains unclear that if these methods are still preferable when the outcome $Y$ is generated under a clinically informative observation process. A comprehensive comparative evaluations and theoretical understanding of the existing PB methods under such realistic EHR missingness mechanisms are lacking. In this paper, we conduct a structured review of 8 recently developed PB methods and categorize them based on their frameworks. We systematically evaluate 9 methods, including 4 PB methods and 5 baseline methods from the traditional missing-data approaches, under 10 different outcome observation process models across different missing assumptions for continuous and binary outcomes. Our results show that PB methods can substantially improve statistical power and estimation efficiency when their missingness assumptions hold, but they may require stronger assumptions than CCA to control type I error. We provide theoretical results to characterize the scenarios under which CCA or PB methods may remain valid. Finally, we apply CCA, weighted CCA, and two PB methods to perform GWAS of six laboratory biomarkers in the All of Us data. The results demonstrate that PB methods can replicate known genetic associations while improving efficiency relative to (weighted) CCA. Furthermore, they extend inferential results to a more representative study population.
BibTeX
@misc{chen2026prediction,
  author = {Xingran Chen and Cheng-Han Yang and Zhenke Wu and Bhramar Mukherjee},
  title = {{Prediction-based Inference in Electronic Health Record (EHR)-linked Biobanks}},
  year = {2026},
  eprint = {2603.14356},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.14356},
}
Figure