Contributed Poster Presentations: ENAR

Logistic regression models binary outcomes in biomedical studies to obtain interpretable odds ratios. Separation occurs when predictors perfectly or nearly perfectly classify cases and controls, causing maximum likelihood estimates to diverge. We develop DISCO (DIagnosis of Separation and Correction of Odds-ratio inflation), a detection-and-estimation framework. DISCO provides pre-hoc diagnosis to detect separation, identify problematic predictors, and quantify severity. A key theoretical insight is that under separation, the coefficient vector's direction remains identifiable, so pairwise ratios of nonzero coefficients are well-defined despite individual magnitudes diverging. We propose a Bayesian estimator with multivariate exponential power prior that penalizes coefficients. We construct a simulation framework generating separation cases with ground truth for fair comparison. Simulations show DISCO detects more problematic data than existing methods, and our estimator controls bias better than alternatives. In HIV-risk data, DISCO outperforms competitors by identifying problematic variables and producing finite, sign-preserving odds ratios. 

Many ordinal discrete diagnostic tests, such as breast cancer tumor grade, are prognostic factors but are often subjective and lack reproducibility. With multiple independent ratings, it is important to evaluate and compare performance among raters.. Here first we consider testing whether one of the raters can act as the gold standard, with the alternative that the gold standard is latent and the observed ratings are modeled by a latent class model (Kruskal, 1977; Dawid & Skene, 1979). When indeed none of the raters can act as the gold standard, we consider both the maximum likelihood method and the Bayesian method for estimating and comparing the concordance between each rater and the latent gold standard. Both simulation studies and application to tumor grade and molecular subtypes of breast cancer are used to illustrate these methods. 

Machine learning (ML) models are used in neuroimaging studies to predict biological and psychometric phenotypes, such as age and psychopathology factor scores. Neuroscientists use Pearson's correlation between the predicted and actual feature to quantify model accuracy for these ML models; however, Pearson's correlation is not accurate when using ML models due to their slow. To address this, we use a model-agnostic semiparametric "one-step" (OS) modification of Pearson's correlation to model associations between functional and structural neuroimaging data with age and psychopathology in the Reproducible Brain Charts (RBC) dataset. We use random forest and ridge regression machine learning models and are able to provide model accuracy scores with confidence intervals that are less biased and more replicable. Our method allows valid model comparisons not possible with Pearson's correlation and shows that functional and structural imaging are highly predictive of age, but they do not improve prediction accuracy for psychopathology. 

Many NIH-sponsored laboratories carry out nontargeted metabolite studies, focusing only on a subset of the entire human metabolome and reporting summary data for the subset of metabolites. To form a complete picture of the entire metabolome or a substantial part of it, one must put the summary analysis on different subsets, obtained from the different laboratories, together in a consistent and efficient manner. We propose a methodology for integrating findings at various laboratories in a cooperative research partnership, in a scenario where only estimated Spearman rank correlation matrices, not raw data, can be shared between partner organizations. We assume the laboratories will study different, but intersecting sets of metabolites and consider how to impute missing values for the metabolite data not collected at a given lab. 

Bayesian kernel machine regression (BKMR) has become a widely used tool for studying the health effects of complex exposure mixtures, with extensions to over dispersed count outcomes now available. However, existing BKMR models for count data handle spatial and temporal structure separately rather than as a unified random field, leaving the joint space-time dependency in the data unmodeled. No existing method therefore combines nonlinear mixture modeling, a neighborhood-based spatio-temporal random effect, and an over dispersed count likelihood in a single computationally workable framework.

We address this gap by proposing SVD-PICAR NB-BKMR, which extends the PICAR framework from the spatial to the spatio-temporal setting through a Kronecker product of spatial and temporal bases. However, this spatiotemporal basis contains redundant components that expand the dimension of the latent random field, produce highly correlated posterior draws, and make the full-rank ICAR field ill-conditioned, weakening exposure-effect estimates and slowing MCMC sampling. We resolve this by incorporating a truncated singular value decomposition (SVD) into the spatio-temporal PICAR basis, retaining only the dominant orthogonal directions, reducing the latent field to a manageable dimension, and separating background spatio-temporal variation from the nonlinear exposure-response pattern.
We compare three competing models: a fixed effect model (M1), a full-rank mixed effect model (M2), and the proposed SVD-PICAR mixed effect model (M3). Models are evaluated through simulation and an application to county level thyroid cancer and pesticide exposure data from Nebraska (1992–2014). In simulation, M3 achieves the best predictive fit by the Watanabe-Akaike Information Criterion (WAIC), the highest MCMC efficiency by effective sample size per second (ESS/sec), and the most stable exposure identification. In the Nebraska application, Atrazine and Alachlor emerge as the leading contributors to thyroid cancer risk, with nonlinear exposure-response relationships that remain robust after spatio-temporal correction. The SVD-PICAR framework offers a principled and practically workable approach for mixture analysis in spatio-temporal count data settings.