arXiv subjects

Jeffrey D. Blume

Publications and source records attributed to Jeffrey D. Blume.

8 recordsLinked to original sources

arXiv2026

Diagnostics for Individual-Level Prediction Instability in Machine Learning for Healthcare

In healthcare, predictive models increasingly inform patient-level decisions, yet little attention is paid to the variability in individual risk estimates and its impact on treatment decisions. For overparameterized models, now standard in machine learning, a substantial source of variability often goes undetected. Even when the data and model architecture are held fixed, randomness introduced by optimization and initialization can lead to materially different risk estimates for the same patient. This problem is largely obscured by standard evaluation practices, which rely on aggregate performance metrics (e.g., log-loss, accuracy) that are agnostic to individual-level stability. As a result, models with indistinguishable aggregate performance can nonetheless exhibit substantial procedural arbitrariness, which can undermine clinical trust. We propose an evaluation framework that quantifies individual-level prediction instability by using two complementary diagnostics: empirical prediction interval width (ePIW), which captures variability in continuous risk estimates, and empirical decision flip rate (eDFR), which measures instability in threshold-based clinical decisions. We apply these diagnostics to simulated data and GUSTO-I clinical dataset. Across observed settings, we find that for flexible machine-learning models, randomness arising solely from optimization and initialization can induce individual-level variability comparable to that produced by resampling the entire training dataset. Neural networks exhibit substantially greater instability in individual risk predictions compared to logistic regression models. Risk estimate instability near clinically relevant decision thresholds can alter treatment recommendations. These findings that stability diagnostics should be incorporated into routine model validation for assessing clinical reliability.

cs.LG↗

arXiv2022

Sequential monitoring using the Second Generation P-Value with Type I error controlled by monitoring frequency

Many adaptive monitoring schemes adjust the required evidence toward a hypothesis to control Type I error. This shifts focus away from determining scientific relevance with an uncompromised degree of evidence. We propose sequentially monitoring the Second Generation P-Value (SGPV) on repeated intervals until establishing evidence for scientific relevance (SeqSGPV). SeqSGPV encompasses existing strategies to monitor Region of Practical Equivalence (ROPE) or Region of Equivalence (ROE) hypotheses. Hence, our focus is to formalize sequential SGPV monitoring; establish a novel set of scientific hypotheses, called PRISM, which is a ROE with a ROPE surrounding the null hypothesis; and use monitoring frequency and a novel affirmation step to control Type I error. Under immediate and delayed outcomes, we assess finite and limiting SeqSGPV operating characteristics when monitoring PRISM, ROPE, and null-bound ROE hypotheses. In extensive simulations, SeqSGPV PRISM monitoring reduced wait time for fully sequential monitoring, average sample size, and reversals of null hypothesis conclusions under the null. With real-world data, we design a SeqSGPV-monitored randomized trial. SeqSGPV is method-agnostic and easy to implement. Adjusting monitoring frequency/affirmation and monitoring a one-sided PRISM synergistically control Type I error. PRISM monitoring and adjusting monitoring frequency to control Type I error may have application beyond SeqSGPV.

stat.ME↗

arXiv2021

Variable Selection in GLM and Cox Models with Second-Generation P-Values

Variable selection has become a pivotal choice in data analyses that impacts subsequent inference and prediction. In linear models, variable selection using Second-Generation P-Values (SGPV) has been shown to be as good as any other algorithm available to researchers. Here we extend the idea of Penalized Regression with Second-Generation P-Values (ProSGPV) to the generalized linear model (GLM) and Cox regression settings. The proposed ProSGPV extension is largely free of tuning parameters, adaptable to various regularization schemes and null bound specifications, and is computationally fast. Like in the linear case, it excels in support recovery and parameter estimation while maintaining strong prediction performance. The algorithm also preforms as well as its competitors in the high dimensional setting (n>p). Slight modifications of the algorithm improve its performance when data are highly correlated or when signals are dense. This work significantly strengthens the case for the ProSGPV approach to variable selection.

stat.ME↗

arXiv2021

Variable Selection with Second-Generation P-Values

Many statistical methods have been proposed for variable selection in the past century, but few balance inference and prediction tasks well. Here we report on a novel variable selection approach called Penalized regression with Second-Generation P-Values (ProSGPV). It captures the true model at the best rate achieved by current standards, is easy to implement in practice, and often yields the smallest parameter estimation error. The idea is to use an l0 penalization scheme with second-generation p-values (SGPV), instead of traditional ones, to determine which variables remain in a model. The approach yields tangible advantages for balancing support recovery, parameter estimation, and prediction tasks. The ProSGPV algorithm can maintain its good performance even when there is strong collinearity among features or when a high dimensional feature space with p > n is considered. We present extensive simulations and a real-world application comparing the ProSGPV approach with smoothly clipped absolute deviation (SCAD), adaptive lasso (AL), and mini-max concave penalty with penalized linear unbiased selection (MC+). While the last three algorithms are among the current standards for variable selection, ProSGPV has superior inference performance and comparable prediction performance in certain scenarios. Supplementary materials are available online.

stat.ME↗

arXiv2020

False Discovery Rate Computation: Illustrations and Modifications

False discovery rates (FDR) are an essential component of statistical inference, representing the propensity for an observed result to be mistaken. FDR estimates should accompany observed results to help the user contextualize the relevance and potential impact of findings. This paper introduces a new user-friendly R package for computing FDRs and adjusting p-values for FDR control. These tools respect the critical difference between the adjusted p-value and the estimated FDR for a particular finding, which are sometimes numerically identical but are often confused in practice. Newly augmented methods for estimating the null proportion of findings - an important part of the FDR estimation procedure - are proposed and evaluated. The package is broad, encompassing a variety of methods for FDR estimation and FDR control, and includes plotting functions for easy display of results. Through extensive illustrations, we strongly encourage wider reporting of false discovery rates for observed findings.

stat.ME↗

arXiv2017

Second-generation p-values: improved rigor, reproducibility, & transparency in statistical analyses

Verifying that a statistically significant result is scientifically meaningful is not only good scientific practice, it is a natural way to control the Type I error rate. Here we introduce a novel extension of the p-value - a second-generation p-value - that formally accounts for scientific relevance and leverages this natural Type I Error control. The approach relies on a pre-specified interval null hypothesis that represents the collection of effect sizes that are scientifically uninteresting or are practically null. The second-generation p-value is the proportion of data-supported hypotheses that are also null hypotheses. As such, second-generation p-values indicate when the data are compatible with null hypotheses, or with alternative hypotheses, or when the data are inconclusive. Moreover, second-generation p-values provide a proper scientific adjustment for multiple comparisons and reduce false discovery rates. This is an advance for environments rich in data, where traditional p-value adjustments are needlessly punitive. Second-generation p-values promote transparency, rigor and reproducibility of scientific results by a priori specifying which candidate hypotheses are practically meaningful and by providing a more reliable statistical summary of when the data are compatible with alternative or null hypotheses.

stat.ME↗

arXiv2017

Bagged Empirical Null p-values: A Method to Account for Model Uncertainty in Large Scale Inference

When conducting large scale inference, such as genome-wide association studies or image analysis, nominal $p$-values are often adjusted to improve control over the family-wise error rate (FWER). When the majority of tests are null, procedures controlling the False discovery rate (Fdr) can be improved by replacing the theoretical global null with its empirical estimate. However, these other adjustment procedures remain sensitive to the working model assumption. Here we propose two key ideas to improve inference in this space. First, we propose $p$-values that are standardized to the empirical null distribution (instead of the theoretical null). Second, we propose model averaging $p$-values by bootstrap aggregation (Bagging) to account for model uncertainty and selection procedures. The combination of these two key ideas yields bagged empirical null $p$-values (BEN $p$-values) that often dramatically alter the rank ordering of significant findings. Moreover, we find that a multidimensional selection criteria based on BEN $p$-values and bagged model fit statistics is more likely to yield reproducible findings. A re-analysis of the famous Golub Leukemia data is presented to illustrate these ideas. We uncovered new findings in these data, not detected previously, that are backed by published bench work pre-dating the Gloub experiment. A pseudo-simulation using the leukemia data is also presented to explore the stability of this approach under broader conditions, and illustrates the superiority of the BEN $p$-values compared to the other approaches.

stat.ME↗

arXiv2017

Missing Data and Prediction

Missing data are a common problem for both the construction and implementation of a prediction algorithm. Pattern mixture kernel submodels (PMKS) - a series of submodels for every missing data pattern that are fit using only data from that pattern - are a computationally efficient remedy for both stages. Here we show that PMKS yield the most predictive algorithm among all standard missing data strategies. Specifically, we show that the expected loss of a forecasting algorithm is minimized when each pattern-specific loss is minimized. Simulations and a re-analysis of the SUPPORT study confirms that PMKS generally outperforms zero-imputation, mean-imputation, complete-case analysis, complete-case submodels, and even multiple imputation (MI). The degree of improvement is highly dependent on the missingness mechanism and the effect size of missing predictors. When the data are Missing at Random (MAR) MI can yield comparable forecasting performance but generally requires a larger computational cost. We see that predictions from the PMKS are equivalent to the limiting predictions for a MI procedure that uses a mean model dependent on missingness indicators (the MIMI model). Consequently, the MIMI model can be used to assess the MAR assumption in practice. The focus of this paper is on out-of-sample prediction behavior, implications for model inference are only briefly explored.

stat.ME↗