arXiv subjects

Stephen Salerno

Publications and source records attributed to Stephen Salerno.

11 recordsLinked to original sources

arXiv2026

Spatially Robust Inference with Predicted and Missing at Random Labels

When outcome data are expensive or onerous to collect, scientists increasingly substitute predictions from machine learning and AI models for unlabeled cases, a process which has consequences for downstream statistical inference. While recent methods provide valid uncertainty quantification under independent sampling, real-world applications involve missing at random (MAR) labeling and spatial dependence. For inference in this setting, we propose a doubly robust estimator with cross-fit nuisances. We show that cross-fitting induces fold-level correlation that distorts spatial variance estimators, producing unstable or overly conservative confidence intervals. To address this, we propose a jackknife spatial heteroscedasticity and autocorrelation consistent (HAC) variance correction that separates spatial dependence from fold-induced noise. Under standard identification and dependence conditions, the resulting intervals are asymptotically valid. Simulations and benchmark datasets show substantial improvement in finite-sample calibration, particularly under MAR labeling and clustered sampling.

stat.ML↗

arXiv2026

PanCanBench: A Comprehensive Benchmark for Evaluating Large Language Models in Pancreatic Oncology

Large language models (LLMs) have achieved expert-level performance on standardized examinations, yet multiple-choice accuracy poorly reflects real-world clinical utility and safety. As patients and clinicians increasingly use LLMs for guidance on complex conditions such as pancreatic cancer, evaluation must extend beyond general medical knowledge. Existing frameworks, such as HealthBench, rely on simulated queries and lack disease-specific depth. Moreover, high rubric-based scores do not ensure factual correctness, underscoring the need to assess hallucinations. We developed a human-in-the-loop pipeline to create expert rubrics for de-identified patient questions from the Pancreatic Cancer Action Network (PanCAN). The resulting benchmark, PanCanBench, includes 3,130 question-specific criteria across 282 authentic patient questions. We evaluated 22 proprietary and open-source LLMs using an LLM-as-a-judge framework, measuring clinical completeness, factual accuracy, and web-search integration. Models showed substantial variation in rubric-based completeness, with scores ranging from 46.5% to 82.3%. Factual errors were common, with hallucination rates (the percentages of responses containing at least one factual error) ranging from 6.0% for Gemini-2.5 Pro and GPT-4o to 53.8% for Llama-3.1-8B. Importantly, newer reasoning-optimized models did not consistently improve factuality: although o3 achieved the highest rubric score, it produced inaccuracies more frequently than other GPT-family models. Web-search integration did not inherently guarantee better responses. The average score changed from 66.8% to 63.9% for Gemini-2.5 Pro and from 73.8% to 72.8% for GPT-5 when web search was enabled. Synthetic AI-generated rubrics inflated absolute scores by 17.9 points on average while generally maintaining similar relative ranking.

cs.CL↗

arXiv2025

Do We Really Even Need Data? A Modern Look at Drawing Inference with Predicted Data

As artificial intelligence and machine learning tools become more accessible, and scientists face new obstacles to data collection (e.g., rising costs, declining survey response rates), researchers increasingly use predictions from pre-trained algorithms as substitutes for missing or unobserved data. Though appealing for financial and logistical reasons, using standard tools for inference can misrepresent the association between independent variables and the outcome of interest when the true, unobserved outcome is replaced by a predicted value. In this paper, we characterize the statistical challenges inherent to drawing inference with predicted data (IPD) and show that high predictive accuracy does not guarantee valid downstream inference. We show that all such failures reduce to statistical notions of (i) bias, when predictions systematically shift the estimand or distort relationships among variables, and (ii) variance, when uncertainty from the prediction model and the intrinsic variability of the true data are ignored. We then review recent methods for conducting IPD and discuss how this framework is deeply rooted in classical statistical theory. We then comment on some open questions and interesting avenues for future work in this area, and end with some comments on how to use predicted data in scientific studies that is both transparent and statistically principled.

stat.ML↗

arXiv2025

A Moment-Based Generalization to Post-Prediction Inference

Artificial intelligence (AI) and machine learning (ML) are increasingly used to generate data for downstream analyses, yet naively treating these predictions as true observations can lead to biased results and incorrect inference. Wang et al. (2020) proposed a method, post-prediction inference, which calibrates inference by modeling the relationship between AI/ML-predicted and observed outcomes in a small, gold-standard sample. Since then, several methods have been developed for inference with predicted data. We revisit Wang et al. in light of these recent developments. We reflect on their assumptions and offer a simple extension of their method which relaxes these assumptions. Our extension (1) yields unbiased point estimates under standard conditions and (2) incorporates a simple scaling factor to preserve calibration variability. In extensive simulations, we show that our method maintains nominal Type I error rates, reduces bias, and achieves proper coverage.

stat.ME↗

arXiv2024

ipd: An R Package for Conducting Inference on Predicted Data

Summary: ipd is an open-source R software package for the downstream modeling of an outcome and its associated features where a potentially sizable portion of the outcome data has been imputed by an artificial intelligence or machine learning (AI/ML) prediction algorithm. The package implements several recent proposed methods for inference on predicted data (IPD) with a single, user-friendly wrapper function, ipd. The package also provides custom print, summary, tidy, glance, and augment methods to facilitate easy model inspection. This document introduces the ipd software package and provides a demonstration of its basic usage. Availability: ipd is freely available on CRAN or as a developer version at our GitHub page: github.com/ipd-tools/ipd. Full documentation, including detailed instructions and a usage `vignette' are available at github.com/ipd-tools/ipd. Contact: jtleek@fredhutch.org and tylermc@uw.edu

stat.ME↗

arXiv2024

What's the Weight? Estimating Controlled Outcome Differences in Complex Surveys for Health Disparities Research

In this work, we are motivated by the problem of estimating racial disparities in health outcomes, specifically the average controlled difference (ACD) in telomere length between Black and White individuals, using data from the National Health and Nutrition Examination Survey (NHANES). To do so, we build a propensity for race to properly adjust for other social determinants while characterizing the controlled effect of race on telomere length. Propensity score methods are broadly employed with observational data as a tool to achieve covariate balance, but how to implement them in complex surveys is less studied - in particular, when the survey weights depend on the group variable under comparison (as the NHANES sampling scheme depends on self-reported race). We propose identification formulas to properly estimate the ACD in outcomes between Black and White individuals, with appropriate weighting for both covariate imbalance across the two racial groups and generalizability. Via extensive simulation, we show that our proposed methods outperform traditional analytic approaches in terms of bias, mean squared error, and coverage when estimating the ACD for our setting of interest. In our data, we find that evidence of racial differences in telomere length between Black and White individuals attenuates after accounting for confounding by socioeconomic factors and utilizing appropriate propensity score and survey weighting techniques. Software to implement these methods and code to reproduce our results can be found in the R package svycdiff, available through the Comprehensive R Archive Network (CRAN) at cran.r-project.org/web/packages/svycdiff/, or in a development version on GitHub at github.com/salernos/svycdiff.

stat.ME↗

arXiv2024

Some models are useful, but for how long?: A decision theoretic approach to choosing when to refit large-scale prediction models

Large-scale prediction models using tools from artificial intelligence (AI) or machine learning (ML) are increasingly common across a variety of industries and scientific domains. Despite their effectiveness, training AI and ML tools at scale can cost tens or hundreds of thousands of dollars (or more); and even after a model is trained, substantial resources must be invested to keep models up-to-date. This paper presents a decision-theoretic framework for deciding when to refit an AI/ML model when the goal is to perform unbiased statistical inference using partially AI/ML-generated data. Drawing on portfolio optimization theory, we treat the decision of {\it recalibrating} a model or statistical inference versus {\it refitting} the model as a choice between ``investing'' in one of two ``assets.'' One asset, recalibrating the model based on another model, is quick and relatively inexpensive but bears uncertainty from sampling and may not be robust to model drift. The other asset, {\it refitting} the model, is costly but removes the drift concern (though not statistical uncertainty from sampling). We present a framework for balancing these two potential investments while preserving statistical validity. We evaluate the framework using simulation and data on electricity usage and predicting flu trends.

stat.ME↗

arXiv2024

From Narratives to Numbers: Valid Inference Using Language Model Predictions from Verbal Autopsy Narratives

In settings where most deaths occur outside the healthcare system, verbal autopsies (VAs) are a common tool to monitor trends in causes of death (COD). VAs are interviews with a surviving caregiver or relative that are used to predict the decedent's COD. Turning VAs into actionable insights for researchers and policymakers requires two steps (i) predicting likely COD using the VA interview and (ii) performing inference with predicted CODs (e.g. modeling the breakdown of causes by demographic factors using a sample of deaths). In this paper, we develop a method for valid inference using outcomes (in our case COD) predicted from free-form text using state-of-the-art NLP techniques. This method, which we call multiPPI++, extends recent work in "prediction-powered inference" to multinomial classification. We leverage a suite of NLP techniques for COD prediction and, through empirical analysis of VA data, demonstrate the effectiveness of our approach in handling transportability issues. multiPPI++ recovers ground truth estimates, regardless of which NLP model produced predictions and regardless of whether they were produced by a more accurate predictor like GPT-4-32k or a less accurate predictor like KNN. Our findings demonstrate the practical importance of inference correction for public health decision-making and suggests that if inference tasks are the end goal, having a small amount of contextually relevant, high quality labeled data is essential regardless of the NLP algorithm.

cs.CL↗

arXiv2024

Do We Really Even Need Data?

As artificial intelligence and machine learning tools become more accessible, and scientists face new obstacles to data collection (e.g. rising costs, declining survey response rates), researchers increasingly use predictions from pre-trained algorithms as outcome variables. Though appealing for financial and logistical reasons, using standard tools for inference can misrepresent the association between independent variables and the outcome of interest when the true, unobserved outcome is replaced by a predicted value. In this paper, we characterize the statistical challenges inherent to this so-called ``inference with predicted data'' problem and elucidate three potential sources of error: (i) the relationship between predicted outcomes and their true, unobserved counterparts, (ii) robustness of the machine learning model to resampling or uncertainty about the training data, and (iii) appropriately propagating not just bias but also uncertainty from predictions into the ultimate inference procedure.

stat.ME↗

arXiv2022

Deep Learning of Semi-Competing Risk Data via a New Neural Expectation-Maximization Algorithm

Prognostication for lung cancer, a leading cause of mortality, remains a complex task, as it needs to quantify the associations of risk factors and health events spanning a patient's entire life. One challenge is that an individual's disease course involves non-terminal (e.g., disease progression) and terminal (e.g., death) events, which form semi-competing relationships. Our motivation comes from the Boston Lung Cancer Study, a large lung cancer survival cohort, which investigates how risk factors influence a patient's disease trajectory. Following developments in the prediction of time-to-event outcomes with neural networks, deep learning has become a focal area for the development of risk prediction methods in survival analysis. However, limited work has been done to predict multi-state or semi-competing risk outcomes, where a patient may experience adverse events such as disease progression prior to death. We propose a novel neural expectation-maximization algorithm to bridge the gap between classical statistical approaches and machine learning. Our algorithm enables estimation of the non-parametric baseline hazards of each state transition, risk functions of predictors, and the degree of dependence among different transitions, via a multi-task deep neural network with transition-specific sub-architectures. We apply our method to the Boston Lung Cancer Study and investigate the impact of clinical and genetic predictors on disease progression and mortality.

stat.ML↗

arXiv2022

High-Dimensional Survival Analysis: Methods and Applications

In the era of precision medicine, time-to-event outcomes such as time to death or progression are routinely collected, along with high-throughput covariates. These high-dimensional data defy classical survival regression models, which are either infeasible to fit or likely to incur low predictability due to over-fitting. To overcome this, recent emphasis has been placed on developing novel approaches for feature selection and survival prognostication. We will review various cutting-edge methods that handle survival outcome data with high-dimensional predictors, highlighting recent innovations in machine learning approaches for survival prediction. We will cover the statistical intuitions and principles behind these methods and conclude with extensions to more complex settings, where competing events are observed. We exemplify these methods with applications to the Boston Lung Cancer Survival Cohort study, one of the largest cancer epidemiology cohorts investigating the complex mechanisms of lung cancer.

stat.ME↗