A Socio-Demographic Latent Space Approach to Spatial Data for Small Area Estimation When Geography is Important but not All-Important
Scott Holan, University of Missouri Abstract:
Many models for spatial and spatio-temporal data assume that “near things are more related than distant things,”
which is known as the first law of geography. While geography may be important, it may not be all-important, for
at least two reasons. First, technology helps bridge distance, so that regions separated by large distances may
be more similar than would be expected based on geographical distance. Second, geographical, political, and
social divisions can make neighboring regions dissimilar. We develop a flexible Bayesian approach for learning
from spatial data which units are close in an unobserved socio-demographic space and hence which units are similar.
As a by-product, the Bayesian approach helps quantify the relative importance of socio-demographic space relative
to geographical space. To demonstrate the proposed approach, we present simulations along with an application to
county-level data on median household income in the U.S. state of Florida.
A Joint Confidence Set for a Ranking Based on Ordered Means
Thomas Mathew, University of Maryland Baltimore County Abstract:
This talk presents the development of a measure of uncertainty associated with an estimated ranking of several
populations, following up on the novel ideas proposed in Klein, Wright, and Wieczorek (2020, JRSS-C), to be
referred to as KWW, Wright (2024, Census Bureau Research Report) and in Wright (2025, Statistics and Probability
Letters). In their work, KWW have developed a measure of uncertainty based on confidence intervals for the
population means, and by considering the overlap/nonoverlap of the confidence intervals. Later, Wright (2024)
derived a new joint confidence region for a ranking based on confidence intervals for the population mean
differences, focusing on the overlap/non-overlap of the intervals with the number zero. The present work explores
possible improvements as follows: (i) consider differences among the ordered population means, rather than the
population means themselves, (ii) compute lower confidence limits for the ordered population mean differences
(i.e., an orthant confidence set), and (iii) tighten the joint confidence set by distributing the confidence
level unequally among the differences of the ordered population means, with a smaller difference receiving a
smaller confidence level. Because the tightness of the confidence set is determined by the number of lower
confidence limits that are equal to zero, the strategy proposed in (iii) results in a tighter confidence set
because it produces a larger lower confidence limit for a smaller difference, thereby reducing the chance of
a lower limit that is zero. The methodology will be applied for ranking the different states in the country in terms
of the mean travel time to work, based on an American Community Survey dataset.
This work is joint with Michael Lucagbo (University of the Philippines Diliman).
Generalized Fiducial Inference for Hybrid Discrete-Continuous Count Data
Kendall Thomas, University of North Carolina at Chapel Hill Abstract:
High-frequency athlete-tracking data can be summarized as quantile cubes (Thomas
and Hannig, 2026): counts of time spent in bins of velocity, acceleration, and movement angle.
In our application, each athlete–match session has 100 such bins, with counts ranging
from under a hundred to several thousand. Whether a bin’s negative binomial distribution
is well approximated by a normal distribution depends on its shape parameter, the product
of its mean and dispersion, rather than on the count alone. We develop a hybrid generalized
fiducial framework that uses exact negative binomial structural equations where the estimated
shape is small and moment-matched normal approximations where it is large. This
approach provides uncertainty quantification for hundreds of regression and dispersion parameters
without specifying a prior distribution. Only the normal components contribute
to the fiducial Jacobian, which has a closed-form block-diagonal structure, and we prove a
Bernstein–von Mises theorem for the correctly specified hybrid model. Simulations show
near-nominal coverage, agreement with a flat-prior Bayesian analysis, and the importance
of partitioning by estimated shape rather than by the mean or observed responses. In 216
sessions from 17 professional women’s soccer athletes, the estimated position effects concentrate
at the velocity extremes, while the estimated age effects follow an acceleration
gradient. However, an athlete-level permutation check indicates that these patterns are not
distinguishable from those obtained under athlete-label permutations.
This work is joint with Jan Hannig ( University of North Carolina at Chapel Hill ).
Inference on Normal Coefficients of Variation and Survival Probabilities: One- and Two-Sample Cases
Samuel Simon, University of Louisiana at Lafayette Abstract:
Simple and accurate methods of constructing confidence intervals (CIs) for normal coefficient of variation
(CV) are proposed. A CI based on a normal approximation is developed, which is not only straightforward to
compute, but also as accurate as the available exact but complicated CI on the basis of noncentral t distribution.
We also propose methods of computing CIs for the ratio and the difference of two coefficients of variation.
These confidence intervals are also simple and accurate in terms of coverage probability and precision. The
results are extended to estimate survival probabilities in one- and two-sample problems. The methods are
illustrated using two examples.
Extrinsic Data Analysis
Vic Patrangenaru, Florida State University Abstract:
We consider all sorts of shape data, processed images and RNA/DNA sequences to be points on an object space,
which is embeddable in a Hilbert space. Thus simple linear operations and projections, combined with large
sample and nonparametric bootstrap principles, lead to new statistical methods
for random objects, and reveal their nontrivial topological structure. Examples of object spaces include real
and complex (shape spaces of 3D views of finite configurations), Hilbert manifolds, stratified spaces (such as
phylogenetic tree spaces), spaces of covariance operators, and space of perceived colors, to mention just a
few examples. The metric structure of
an object space makes it possible to develop notions of a
extrinsic mean and extrinsic covariance, via projections and differentiability principles, respectively, and
ultimately a nonparametric approach to statistical inference for object data. We propose to develop new methods
or expand upon existing approaches to Object Data Analysis for shape, 2D and 3D
digital image and RNA or DNA data, with an eye towards the computational speed of algorithms required for our
methodologies, and real-world applications.
This work is joint with Robert L. Paige.
Advances in BHV4 Analyses and AI Possibilities for Phylogenetic Tree Building
Garett Ordway, Pensacola Christian College Abstract:
We discuss the development of the novel representation called the Spiky Projective Excavated Dodecahedron (SPED).
Then we show the development of the extrinsic statistical analysis on the space of Billera- Holmes-Vogtmann tree
space using the Veronese- Whitney (VW) embedding on the SPED. A couple applications include the original yeast
genome dataset to study the phylogenetic trees of four distinct yeast clades and Covid-19 variants. Finally, we
investigate the frontier of AI in genomic trees.
Extrinsic MANOVA for Oriented Projective Shape Analysis of 3D Faces
Musab Alamoudi, Florida State University Abstract:
Nonparametric MANOVA provides a framework for comparing mean vectors from multiple populations. This framework
can be extended to testing the equality of extrinsic means for random objects on smooth manifolds. In
three-dimensional oriented projective shape analysis, a labeled configuration of k landmarks containing an
oriented projective frame is represented on a product of spheres (S3)q, where q = k - 5. In this work, we
apply manifold MANOVA to the case of 3D oriented projective shape (OPS) of 3D landmark configurations.
AI techniques are used here for retrieving 3D facial surface data from regular digital camera images. For
g groups, under equality of the extrinsic means, the large-sample null distribution of the statistic is
chi-square with 3q(g-1) degrees of freedom. Because facial data sets typically contain only a small number
of independent reconstructions per subject, the statistic is calibrated in practice by nonparametric bootstrap.
The methodology is applied to 38 independent 3D reconstructions from seven subjects, using nine anatomical
facial landmarks. Well-conditioned, non-coplanar projective frames produced strong and stable differentiation
among subjects, whereas nearly coplanar frames led to unstable inference. The OPS framework also provides a
fast methodology for facial differentiation through the oriented projective frame approach.
This work is joint with Victor Patrangenaru (Florida State University) and Robert Paige (Missouri University of Science and Technology).
Object Data Analysis for Protein Structures
Ruiyang Sun, Florida State University Abstract:
Protein conformations are naturally represented as high-dimensional geometric objects rather than collections of
independent scalar measurements. Nevertheless, structural differences between proteins are commonly quantified
using pairwise summaries such as root-mean-square deviation (RMSD). We develop an object data analysis framework
for studying protein conformational variability by representing three-dimensional protein structures as shape-valued
observations. After removing translational and rotational effects through geometric alignment, we examine structural
variation using shape-based distances, mean configurations, and low-dimensional representations of conformational
variability. The methodology is applied to a curated collection of apo and holo protein structures, along with
corresponding AlphaFold predictions. We investigate whether object- and shape-based representations reveal aspects
of conformational diversity that are not captured by conventional RMSD-based comparisons, and whether predicted
structures reproduce geometric features observed in experimental conformations. The proposed analysis provides
a population-level statistical perspective on protein structural variability and demonstrates the potential role
of object data analysis in the evaluation of generated molecular structures.
Data-Efficient Generative Modeling for LPBF Process Design
Xiyuan Liu, Louisiana Tech Abstract:
Machine learning (ML) and generative modeling are increasingly used in manufacturing design. However, during
early-stage process development, experimental datasets are often limited, making it challenging to estimate
complex process-structure relationships. In this study, we propose a lightweight bidirectional generative
framework that models the joint probability distribution between laser powder bed fusion (LPBF) process parameters
(X) and melt-pool geometry (Y) using limited datasets. The framework uses a modified Conditional Variational
Autoencoder (CVAE) with sparse interactions among process parameters and physics-informed relationships in the
latent space. This approach provides a lower-dimensional representation of the joint distribution and reduces
the risk of overfitting with limited data. The proposed framework supports both forward and inverse inference.
The forward model estimates the distribution of melt-pool width and depth and reconstructs melt-pool geometry
for digital-twin development, while the inverse model estimates process-parameter ranges conditional on a
prescribed melt-pool geometry. The numerical CVAE achieves R2=0.955 using the full dataset and R2=0.87
with an 80/20 train-test split. For image-based reconstruction, the two-stage model achieves R2=0.7790.
These results demonstrate that the proposed framework provides a data-efficient approach for estimating joint
process-geometry distributions and developing forward and inverse digital twins for LPBF process design, with
the potential to guide future experimental design.
A composite likelihood-based approach for multiple changepoint detection in snow cover count time series
QiQi Lu, Virginia Commonwealth University Abstract:
Accurate assessment of the Earth’s snow cover is critical for understanding climate
change, yet historical snow-cover records are complicated by overdispersion, serial dependence,
and inhomogeneities arising from changes in observation technology, features
that previous analyses have failed to address simultaneously. In this talk, we
model the annual number of snow-covered weeks at a fixed location as a beta-binomial
count time series, coupling a latent Gaussian autoregressive process with a time-varying
marginal distribution whose mean accommodates both trends and multiple unknown
changepoints. Because the full likelihood is computationally intractable, we develop an
inference procedure based on a first-order pairwise composite likelihood and identify
changepoint configurations by optimizing a composite likelihood information criterion
using a genetic algorithm. Simulation studies demonstrate the performance of the
proposed method, and we apply it to the Rutgers Northern Hemisphere 24-km Weekly
Snow Cover Extent dataset to investigate changes in the annual number of snow-covered
weeks over time.
This work is joint with Edward Akuffo Addo Dankwa (Mississippi State University), Evan Ling (George Mason University),
Penelope Prochnow (Mississippi State University), and Jonathan Woody (Mississippi State University)
Addressing Imbalance Data: Statistics Meets Computer Vision
Hossein Moradi Rekabdarkolaee, Bowling Green State University Abstract:
Infectious bovine keratoconjunctivitis (IBK), commonly known as pinkeye, is a highly contagious ocular disease
in cattle that causes significant welfare and economic losses. Early automated detection from farm images can
enable timely veterinary intervention. However, these data could be imbalance. In this talk, we will compare
statistical and Deep Learning algorithms that address the imbalance problem in classification.
Impact of Measurement Error on Machine Learning Performance and Interpretability
Tamer Oraby, University of Texas Rio Grande Valley Abstract:
Measurement error is widespread in modern data pipelines, yet its effects on machine-learning
prediction and interpretation are often overlooked. We examine how four common error types, namely classical
measurement error, feature misclassification, Berkson error caused by model- or algorithm-driven assignments,
and group-level summary assignment error, affect predictive accuracy and permutation-based feature importance.
Our study shows that measurement error generally reduces predictive performance and significantly distorts
the estimated influence and ranking of key features. The severity and pattern of this impact depend on the
error type, level, the machine learning method used, and the prediction task. These findings demonstrate that
measurement error can undermine both the predictive power and interpretability of models, even when they are
well-tuned. Using the computational resources of Lonestar6, we systematically examined a wide array of error
mechanisms, learning algorithms, and simulation scenarios. Our results emphasize the importance of detecting
and mitigating measurement error before drawing conclusions about model robustness, generalizability, or the
importance of specific variables.
This work is joint with Zeinab Mohamed (Oberlin College).
CAFT: A Compositional Log-Linear Model for Microbiome Data with Zero Cells
Mo Li, University of Louisiana at Lafayette Abstract:
Differential abundance analysis is fundamental in microbiome research. However, microbiome data are compositional,
highly sparse (with many zeros), and often affected by differential experimental bias across taxa, making valid
inference difficult. Standard methods frequently rely on relative abundances or pseudocounts, which can introduce
spurious associations and undermine false discovery rate (FDR) control. We propose the Compositional Accelerated
Failure Time (CAFT) model, a new framework that treats zero counts as censored observations below a detection limit
and therefore avoids pseudocounts. CAFT is inherently robust to multiplicative technical bias and addresses
compositional effects through properly constructed score tests. Simulations show that CAFT improves type I error
and FDR control relative to competing compositional methods while maintaining strong power, with or without
technical bias. Applications to real microbiome datasets further illustrate its utility. CAFT provides a robust,
efficient, and powerful approach for compositional differential abundance analysis.
Frontiers of Statistics in Science and Engineering: 2035 and Beyond National Academies of Science, Engineering and Medicine Consensus Study
Katherine Bennett Ensor, Rice University Abstract:
The National Academies of Sciences, Engineering, and Medicine committee convened to examine the discipline of statistics, its accomplishments,
and its future directions, leading to the release of Frontiers of Statistics in Science and Engineering: 2035 and Beyond in February 2026.
The project was sponsored by the National Institutes of Health (NIH), the National Science Foundation (NSF), and the National Security
Agency (NSA), which gave the committee a three-part charge: assess where the statistical sciences stand today, describe where they
need to go, and recommend specifically how the nation should invest to get there. The report's first sentence sums up the committee's
case for the field: "Statistics is the science through which data become wisdom." I was honored to serve as chair of this rock star committee,
working alongside vice-chair Lance Waller at Emory University. I will share findings and recommendations as well as a bit about the
process and the role NASEM plays in the technology infrastructure of the country.
Revisiting Equity–Bond Diversification under Tail Risk and Structural Market Change: Evidence from Global ETFs and Option-Implied Dynamics
Farzana Afroz, Texas Tech University Abstract:
Balanced equity–bond portfolios face challenges during periods of inflation, monetary tightening,
and changing market dependence. This study examines thirty balanced and multi-asset
exchange-traded funds listed in the United States, Canada, and Australia using daily U.S.-dollar
returns over 2020–2024. A rolling-window framework compares minimum-variance and Conditional
Value-at-Risk (CVaR) allocations under long-only and constrained long–short specifications
with an equally weighted benchmark. The analysis evaluates cumulative performance, riskadjusted
returns, maximum drawdown, Value-at-Risk (VaR), and CVaR, complemented by filtered
extreme-tail estimation and conditional volatility modeling. Over the reported evaluation
period, optimized portfolios exhibit lower empirical VaR and CVaR than the equally weighted
benchmark, while long–short portfolios exhibit lower tail-loss measures than their corresponding
long-only strategies. The long–short 95% CVaR portfolio records the smallest maximum
drawdown and the highest reported Sharpe, Sortino, and Rachev ratios, although all reported
Sharpe and Sortino ratios are negative. GARCH and FIGARCH comparisons indicate substantial
volatility persistence, with information criteria favoring GARCH and squared-residual
diagnostics showing no significant remaining autocorrelation at the examined lags. A historically
estimated Normal Inverse Gaussian model provides an illustrative framework for European
option valuation and model-generated implied-volatility surfaces. These surfaces represent implications
of the specified pricing model rather than observed option-market expectations. The
findings highlight the value of evaluating multiple dimensions of portfolio risk, while investment
interpretations remain subject to trading costs, short-selling costs, and leverage considerations.
This work is joint with Abootaleb Shirvani (Kean University), Nicholas Appiah (Texas Tech University),
Svetlozar T. Rachev (Texas Tech University), and Frank J. Fabozzi (Johns Hopkins University).
A Study on the Influence of International Trade, Monetary Policy and Regime Shift, towards the Economic Growth in Sri Lanka
Ruwani Denagama Vitharanage, Texas Tech University Abstract:
Economic growth refers to the increase in goods and services produced by a nation over time. Sri Lanka, as a
developing country, has faced economic challenges due to the 30-year civil war that ended in 2009, the Easter
Attacks in 2019, and the COVID-19 pandemic. Therefore, restoring economic activities and achieving sustainable
growth have become important topics.
This study examines, the influence of international trade, monetary policy, and regime shifts on Sri Lanka’s
economic growth using econometric and machine learning models. Previous studies have mainly examined international
trade and monetary policy separately. This study extends literature by investigating their joint effects on
economic growth while considering macroeconomic factors such as inflation, interest rates, and trade openness
during 2007–2019. GDP is used as the response variable, while explanatory variables include continuous time
series variables selected through a literature review and categorical variables representing monetary policy
and regime shifts.
Among the time series models, SARIMA and VARX(7) were used, while Random Forest, XGBoost, and LSTM models were
applied as machine learning algorithms. Feature selection was incorporated into the machine learning models to
improve performance. The results reveal that consumer goods, intermediate goods, the Trade Openness Index, and
inflation, are the most significant factors influencing Sri Lanka’s economic growth. In contrast, SDFR and SLFR
rates and regime shifts have no significant effect. The LSTM model using Random Forest-selected features was
identified as the best model for forecasting GDP. Although the VARX model achieved a 99.94% coefficient of
variation for GDP predictions within the study period, correlated error terms make it unsuitable for forecasting.
These findings can assist relevant authorities in developing strategies to support the recovery and sustainable
growth of Sri Lanka’s economy.
Stirling Numbers of the Second Kind: Properties and Applications in Mathematics and Statistics
Subhash Bagui, University of West Florida Abstract:
The Stirling numbers of the second kind have numerous applications in mathematics and statistics,
yet these applications are not widely known among college students in these fields. This work aims to
promote the use of Stirling numbers of the second kind among college students in mathematics and statistics.
We present the Stirling numbers of the second kind in a clear, accessible way so that college students
can grasp them easily. We derive the formula for the Stirling numbers of the second kind using the
concept of one-to-one and onto functions. Using this formula, we derive the standard recurrence relation.
By the multinomial theorem, we express a monomial rⁿ in terms of the Stirling numbers of the second kind
and the falling factorial of r. Using this generating function, we give a closed form for the sums of
integral powers of integers. Finally, for applications in statistics, we provide closed-form expressions
for the nth (raw) moments of several discrete distributions, including Binomial, Poisson, Geometric, and
Negative Binomial.
This work is joint with KL Mehra (University of Alberta).
Contact Weighted Training and Multi Context Encoding for All Atom Inverse Protein Folding
Mithun Ranjan Kar, University of Louisiana at Lafayette Abstract:
Inverse protein folding seeks to design amino acid sequences that fold into a target structure while
preserving functional interactions. For all-atom protein complexes, this requires modeling interactions
with chemically diverse partners, including metal ions, nucleotides, and small molecules. ADFLIP addresses
this problem using discrete flow matching over all-atom structures and achieves strong interaction
residue recovery. However, its training objective assigns equal importance to all residues, even though
interaction residues account for only a small fraction of the sequence and largely determine binding
specificity. Moreover, it represents all non-protein partners using a shared context encoder despite their
distinct structural and chemical characteristics.
We improve ADFLIP with two complementary modifications. First, we introduce contact weighted
training, which increases the contribution of residues within 5°A of a non-protein partner, encouraging
the model to focus on interaction-critical regions during optimization. Second, we propose multi-context
encoding, which learns separate contextual representations for metal ions, nucleotides, and small
molecules while preserving the original discrete flow formulation and inference procedure.
Experiments on the ADFLIP benchmark show that our approach consistently improves interaction
residue recovery across all partner types. Compared with our reproduced ADFLIP baseline, interaction
residue recovery increases from 74.24% to 75.76% for metal ions, from 49.43% to 51.08% for nucleotides,
and from 60.74% to 63.16% for small molecules. These improvements are achieved while maintaining
competitive sequence modeling quality, with perplexities of 2.64, 3.10, and 2.74, compared with 2.61,
4.86, and 3.57 for the ADFLIP baseline. Our results demonstrate that emphasizing interaction residues
and learning partner-specific contextual representations effectively enhances all-atom inverse protein
folding.
This work is joint with Kumer Das (University of Louisiana at Lafayette).
Quantifying the Pressure Experienced by Batters and Bowlers in T20 Cricket
Ananda Manage, Sam Houston State University Abstract:
As in many sports, cricket players experience varying levels of pressure throughout a match, with pressure often
intensifying during critical stages of the game. This study proposes a mathematical framework for quantifying the
match pressure experienced by batters and bowlers in Twenty20 (T20) cricket. A Batting Pressure Index and a Bowling
Pressure Index are derived from the solutions to two separate systems of differential equations developed under
assumptions that reflect the practical dynamics of the game. Both indices are expressed on standardized scales,
allowing pressure to be compared across different stages of an innings and across matches. Although the proposed
framework is developed specifically for T20 cricket, it can be adapted to other formats of cricket through appropriate
modifications.
Application of Market Basket Analysis to Identify Co-Occurring Substance Use Among Patients Admitted to Publicly Funded Substance Use Treatment Programs
Chesmi Kumbalatara, Southern Illinois University Abstract:
Co-occurring substance use (polydrug use) is increasingly recognized as a feature of substance use disorders,
highlighting the need to identify these common patterns in treatment populations. This study aimed to identify
these patterns among patients admitted to publicly funded substance use treatment programs in the U.S. The study
also assessed the applicability of Market Basket Analysis (MBA), a data mining technique used to find patterns in
what customers buy, for identifying drug-use patterns in large datasets.
The study analyzed 2024 admissions data of patients aged 12 and older from the Treatment Episode Data Set (TEDS), a
national data system that tracks annual admissions and discharges for substance use treatment in state-licensed or
certified facilities. To identify co-occurring substance use, two statistics of MBA were used:
Support% (how frequently a drug combination appears across all admissions) and Confidence% (conditional probability
of using one drug also using another).
Of 1,583,823 patients, 30.4%, 9.1%, and 7.3% used alcohol, heroin, and marijuana, respectively. Six drugs were
used >1% as either primary (i.e., initiated first), secondary, or tertiary. Confidence% for alcohol users to also
use marijuana was 26%, and for marijuana users to also use alcohol was 46%. Differences in support% and confidence%
for common use patterns across sex, race/ethnicity, and health insurance type will be discussed. For ages <20,
support% for co-occurring alcohol and marijuana use was 27% with confidence% of 71% for alcohol users to also use
marijuana. Pre-pandemic data have similar patterns and will be presented.
MBA is useful for detecting common substance use patterns and as an interactive, data-driven visualization
technique to demonstrate them to health professionals at local clinics. Identifying co-occurring substance use
can help develop timely interventions that prevent these common patterns at the local level.
Bayesian Spatial Causal Modeling of Asthma Prevalence in the United States
Rajitha Edirisingha Mudiyanselage, Texas Tech University Abstract:
Asthma remains a major public health concern in the United States, with substantial geographic variation across
counties. Long-term exposure to fine particulate matter (PM2.5) is widely associated with respiratory health risks,
motivating its investigation in relation to county-level asthma prevalence, however, estimating its causal effect
on asthma prevalence is challenging due to spatial confounding and correlated environmental exposures.
In this study, we propose a Bayesian spatial causal inference framework that integrates generalized propensity score
(GPS) methods with spatial modeling to estimate the exposure-response relationship between PM2.5 and county-level
asthma prevalence. Spatial dependence is modeled using conditional auto-regressive (CAR) priors, while confounding
is addressed through spline-based adjustment of the generalized propensity score.
Our findings indicate that incorporating both spatial random effects and propensity score adjustment substantially
improves model fit and reduces bias in estimating treatment effects. The results suggest a statistically significant
positive association between long-term PM2.5 exposure and asthma prevalence. These findings highlight the importance
of jointly addressing spatial dependence and confounding when conducting causal inference in environmental health
studies.
It’s Not Your Parents’ Income, It’s Your Neighbors’ Diplomas - A Cluster Analysis of Neighborhood Groups in Forsyth County, NC
Quoc Bui, Winston-Salem State University Abstract:
Two children are born on the same day in Forsyth County. One arrives in Census Tract 8.01 in East
Winston-Salem, where only 18% of adults have a college degree, where most families are headed
by single parents, and where her chance of reaching the top income bracket is just 4%. Ten miles
west in Clemmons, another child enters a neighborhood where 58% of adults have degrees and her
odds of moving up are four times higher. These two children, separated by a short drive, face very
different futures, not because of their talents or their parents’ choices, but because of where they
were born.
Winston-Salem ranks among the worst cities in America for upward mobility. A child born into
a low-income family here has only a 4.5% chance of reaching the top income bracket as an adult.
Yet this county-wide statistic hides enormous variation from one neighborhood to the next.
Some census tracts offer children real pathways to opportunity; others seem to trap families across
generations. This research asks: what makes these neighborhoods different?
Preliminary findings suggest that adult educational attainment is the variable category that most
strongly differentiates neighborhood types in Forsyth County, and it is also the factor most correlated
with economic mobility within the county. This makes intuitive sense: education opens doors
to higher-paying jobs and greater economic security. However, there is a problem. School quality
varies dramatically across the county, with disadvantaged neighborhoods served by schools that
have less experienced teachers and higher turnover. Closing these school quality gaps would require
sustained investment over many years, along with policy changes that face significant political obstacles.
By contrast, recent research suggests a more feasible lever: economic connectedness. If
cross-class friendships can shape aspirations and provide access to information and job opportunities,
then community programs that foster interactions between residents of different neighborhoods
may offer a more immediate pathway to improving mobility, even while longer-term efforts to improve
schools continue.
This work is joint with Bao Maddux (Winston-Salem State University).
Real-world Impacts from Statistical Research in Industry, Government, and Academe
Karen Kafadar, University of Virginia Abstract:
Advances in statistics arise often, though not exclusively, through academic research. The problems that motivate these
advances can come from researchers in other university departments as well as in industry, government, and public health.
In this talk, I will briefly describe research that arose from projects at Hewlett-Packard, National Cancer Institute, and UVA's
Dept of Neurophysiology. These examples demonstrate that research in statistics can arise unexpectedly in many scientific
disciplines and require theoretical results across the spectrum of statistics to solve real-world problems.
Divide, Shrink, and Conquer: Estimation Strategies for Large-Scale Data
S. Ejaz Ahmed, Brock University, Ontario, Canada Abstract:
Modern datasets are often too large, high-dimensional, and distributed to be conveniently stored and analyzed in a single location.
This creates both computational and storage challenges and motivates statistical methods that can extract and combine
information without requiring the full dataset to be centralized. In this talk, I present a divide-and-conquer approach to
shrinkage estimation for sparse linear regression. The basic idea is simple: divide a large dataset into smaller subsets,
perform estimation locally, and then combine the resulting information through a single round of communication. We
develop pretest- and Stein-type shrinkage estimators that exploit sparsity while reducing estimation variability and
substantially limiting the amount of information that needs to be stored or transferred. The theoretical properties of
the proposed one-shot estimators are established, including consistency and asymptotic normality. Their finite-sample
performance is investigated through extensive simulation studies and illustrated using the Million Song Year Prediction
Dataset, a large-scale dataset containing audio features from one million contemporary songs. The application considers
the challenging task of predicting the release year of a song from its acoustic characteristics, providing a natural setting
for evaluating distributed estimation in a high-dimensional data environment. The talk demonstrates how divide-and-conquer
and shrinkage can work together to address the statistical challenges created by modern large-scale data: divide the data,
shrink intelligently, and conquer the estimation problem.
Newer Innovations in Randomized Response Technique (RRT) Research
Sat Gupta, University of North Carolina Greensboro Abstract:
In this talk, we will discuss various new RRT innovations that have taken place in the last 20 years
such as optionality, mixture models, trust enhancement, unified measure of model quality, measurement
errors in the binary domain, and adverse effect of the use of auxiliary information in the RRT domain
Impact of Partial Randomized Response Technique on Stratified Calibration Estimators with Different Types of Distance Measures in the Presence of Measurement Error
Trisandhya Pidugu, Bharati Vidyapeeth's College of Engineering, New Delhi Abstract:
Calibration is a methodological approach that enhances the accuracy of population parameter estimates
through the integration of auxiliary information. In this study, a new ratio calibration method has
been developed with distance measures under the partial randomized response technique in presence of
measurement error to address the problem of estimating the population mean under stratified approach.
Additionally, to show the impact of measurement error under the Impact of Partial Randomized Response
Technique (PRRT) model, the proposed calibration estimators are compared with similar direct
estimators (i.e. in absence of measurement error and randomization). The effectiveness of the proposed
calibration estimators is evaluated through a simulation study using a real data.
Keywords: Auxiliary information, Calibration estimators, Mean estimation, Measurement error,
Randomized Response Technique, Stratified random sampling.
This work is joint with Sat Gupta (University of North Carolina at Greensboro)
Trust Enhancement in Binary Randomized Response Technique (RRT) Models Abstract:
Mala Gupta, University of Texas Dallas Abstract:
Building on recent studies on trust enhancement in quantitative Randomized Response Technique (RRT)
models, this study introduces trust enhancement in a binary mixture RRT model. Our approach
emphasizes building respondent trust alongside preserving anonymity. To support our theoretical
findings, we conduct a simulation study. The overarching goal of our work is to shift the focus
from merely protecting anonymity and accounting for lack of trust to actively foster trust.
Simulation results show that the proposed model has better quality, and succeeds in building
respondent trust that results in more efficient estimates. We will also discuss opportunities
to carry this work forward.
This work is joint with Nikita Jayaraj (Texas A&M University).
Optional Mixture Binary RRT Models Accounting for Trust Deficit and Measurement Errors
Pujita Sapra, High Point State University Abstract:
Randomized Response Technique (RRT) models have been designed to protect
respondents’ privacy in surveys that include sensitive questions. Recent studies
have shown that despite the promised privacy protection, some respondents may
not fully trust the model and may choose not to participate in the survey or provide
inaccurate responses. There have been studies which account for potential lack of
trust in RRT, but the underlying model becomes very complex and increases the
possibility of measurement errors. Recent studies have examined models that not
only account for trust deficit but also simultaneously account for measurement
errors. In the present study, we explore how much additional gains can be accomplished
by using Optional RRT models compared to Non-Optional RRT models
for tackling trust deficit and measurement errors simultaneously.
This work is joint with Caroline Wyrick (UNCG), Sadia Khalil (Lahore College For Women University), Qi Zhang (Independent Researcher), and Sat Gupta (UNCG)
Estimating Age Misrepresentation Among Collegiate Student Athletes in India: A Field Work Study Using a Mixture Binary RRT Model
Geeta Kalucha, University of Delhi Abstract:
Age misrepresentation fundamentally undermines fairness and resource allocation within Indian sports. This study
applies a Mixture Binary Randomized Response Technique (RRT) among university athletes at the campus of a Delhi
University college to estimate the true prevalence of age fraud while assessing trust in the method’s privacy
guarantee. Using the direct-question approach anonymously, prevalence was found to be 0.300 (variance 0.000105),
trust in RRT's privacy protection was 0.750 (variance 0.002526), and the RRT- based estimated prevalence was
0.260606 (variance 0.005615). These findings reveal substantial sensitivity surrounding age fraud and strong
acceptability of RRT’s privacy guarantees. The methodological advantages, practical implementation guidance, and
policy relevance for integrating RRT into national anti-age fraud frameworks are discussed in depth.
Bayesian Tele-connected Spatial Clustering of Multivariate Spatial Data with Applications to Disease-Mapping
Srijato Bhattacharyya, Texas A&M University Abstract:
Spatial clustering is crucial in disease mapping by identifying subregions with different patterns of disease
incidence or mortality. This study proposes a novel Bayesian spatial clustering method for multivariate spatial
disease data, which allows for understanding geographic variations of multivariate disease patterns while
accounting for both spatial information and dependence among multiple disease measurements. We develop a new
random tele-connected graph partition model with an unknown number of clusters, which is capable of encouraging
locally contiguous clusters and allowing for remote subregions to be clustered together. We use this prior in a
Bayesian hierarchical model to detect spatial clusters and estimate cluster-specific disease patterns and
dependence across the multivariate disease variables. We develop a tailored Markov chain Monte Carlo (MCMC)
algorithm for posterior inference, utilizing efficient doubly split-merge samplers taking advantage of graph
algorithms. We illustrate our method with simulation studies and apply it to investigate the clustering patterns
of county-level prostate cancer mortality rate decline across six southern U.S. states from 1985 to 2014.
Non-Homogeneous Continuous-time Markov Chain: Estimations and Challenges
Joonha Chang, Louisiana State University Health Science Center, New Orleans Abstract:
Non-homogeneous continuous-time Markov chain (NH-CTMCs) model offers a highly flexible framework for modeling
longitudinal processes, yet they remain understudied due to severe computational bottlenecks. The primary
challenge lies in transition probability estimation, which directly complicates parameter estimation via maximum
likelihood estimation (MLE). While various numerical and analytical approaches have been proposed to evaluate
the transition matrix such as Uniform Accelerated (UA) Expansion, product integrals, the Peano-Baker series,
and the Magnus expansion, robust and computationally efficient solutions remain inaccessible for general state
spaces. In this talk, we address these challenges by presenting newly derived, well-approximated closed-form
transition probabilities for 2-state and 3-state NH-CTMC models that utilize log-logistic transition rates.
We will demonstrate how the closed-form solutions effectively overcome traditional computational hurdles to
enhance computations on MLE. Furthermore, we will view this contribution within the broader landscape of
NH-CTMC research, discussing the inherent mathematical difficulties in scaling these models. Finally, we will
discuss potential future directions for the field, highlighting the need for developing better approximation
techniques for product integrals and the Peano-Baker series to advance the broader application of NH-CTMCs
with general forms of rate functions.
SCG: Identifying Spatially Co-Expressed Genes Using Spatial Covariance Regression
Ihsan Buker, University of Alabama at Birmingham Abstract:
Spatial transcriptomics (ST) technologies enable the in situ measurement of gene expression. These platforms
have proven invaluable for understanding molecular mechanisms of drug action, identifying spatial
niches within tumors, and characterizing cell differentiation. Despite considerable methodological progress
in cell-type deconvolution, spatial domain identification, and spatially variable gene identification, gene
co-expression in spatial settings has received comparatively less attention.
From a biological perspective, gene co-expression often reflects shared regulatory or functional processes,
such as membership in the same pathway. Gene co-expression networks present a powerful method
for analyzing such patterns, representing the structured dependence among genes induced by an underlying
biological process. Existing methods largely estimate a single global co-expression network, ignoring spatial
heterogeneity. The small group of methods that do consider spatial heterogeneity often scale poorly or
analyze gene co-expression on an a priori specified list of gene pairs, as opposed to complete networks as
we do here.
We propose the Spatial Covariance Regression (SCR) framework, which estimates spot-specific gene
co-expression networks and identifies Spatially Co-expressed Genes (SCGs) across the tissue domain.
Starting from a normalized gene expression matrix, SCR employs a spatially varying Bayesian factor model
to infer spot-specific correlation matrices. These matrices yield spot-specific gene co-expression networks,
with genes as nodes and marginal correlations as edges.
A key computational challenge is the high dimensionality of modern ST platforms, namely, platforms
that profile a rapidly expanding gene panel across an increasing number of spatial locations. SCR
addresses this through a global–local decomposition of the spatially varying factor loadings, which considerably
reduces the number of parameters to be estimated. Scalable posterior inference is achieved via
a GPU-accelerated coordinate ascent variational inference (CAVI) algorithm, enabling SCR to scale to
platforms such as 10X Genomics Xenium while allowing for robust uncertainty quantification.
To identify SCGs, we apply Otsu’s thresholding to the posterior spatial standard deviations of the
edges, yielding a data-driven threshold for spatial variability. Gene pairs whose edge strength exhibits
spatial variation exceeding this threshold are classified as SCGs, with the false discovery rate controlled at 10%.
We evaluate SCR in terms of graph structure recovery and edge estimation accuracy across varying spatial
resolutions (n = 10,000 to 50,000 locations). Across all settings, SCR achieves high Matthews correlation
coefficients (MCC : 0.94–0.95) and true positive rates (TPR : 0.94–0.97), while maintaining consistently low
false positive rates (FPR < 0.01), demonstrating accurate discrimination between true and spurious edges
despite class imbalance. Recovery of spatially varying edges, assessed via the RV coefficient, is consistently
close to 1 across spatial locations and replicates, indicating strong concordance between estimated and true
spatially varying edges.
Applied to breast cancer ST data, SCR identifies biologically meaningful spatial coexpression
patterns across pathologist-annotated tissue regions (normal, intermediate, and tumor). Regionspecific
connectivity scores across cancer hallmark pathways reveal spatially heterogeneous network structure
that aligns with known tumor biology, including pathways related to epithelial–mesenchymal transition
and angiogenesis. Applied to an Alzheimer’s disease mouse brain ST dataset, SCR identifies
spatially varying co-expression among key neuroinflammatory genes across mutant and wild-type samples.
Region-specific connectivity differences reveal altered gene interaction patterns in disease-relevant brain
regions, consistent with known neuroinflammatory mechanisms.
SCR provides a scalable and statistically principled framework for estimating spatially varying gene coexpression
networks and identifying SCGs. By combining a global–local Bayesian factor model with efficient
variational inference, SCR scales to modern ST platforms and produces biologically interpretable results in
complex tissue environments. Python and R modules implementing SCR will be made publicly available.
Joint Smooth Direct Search for Learning Dynamic Treatment Regimes from Longitudinal Health Data
Nilson Chapagain, Texas A&M University Abstract:
Longitudinal health studies create opportunities to learn adaptive intervention strategies that tailor decisions
to an individual’s evolving history, but estimating such strategies becomes challenging when decisions occur at
multiple stages and involve more than two treatment options. We propose a joint smooth direct-search approach
that formulates multistage treatment-rule estimation as a weighted multiclass decision problem and replaces
the nonsmooth optimal-regime objective with Fisher-consistent smooth surrogate criteria. The method optimizes
treatment rules across stages simultaneously, allowing intermediate outcomes and later-stage responses to
inform earlier decisions while avoiding reliance on a fully specified outcome model. Through simulation studies,
we show that joint direct search performs competitively across a range of multistage settings. It is especially
advantageous when treatment effects are nonlinear, decision boundaries are not well captured by simple regression
or blip models, or early-stage choices affect the quality of later-stage decisions. At the same time, our results
clarify when standard methods remain competitive, particularly when their regression, contrast, or
weighted-classification structure is well aligned with the underlying treatment rule. We illustrate the method
using electronic health record data from adult ICU patients with sepsis in MIMIC-III, estimating two-stage adaptive
IV fluid resuscitation rules based on admission and early post-admission clinical histories. This work highlights
how flexible statistical learning methods can support dynamic decision-making in longitudinal biomedical data.
Bayesian Semiparametric Joint Multivariable Frailty Model for Multitype Recurrent Events and a Terminal Event
Mithun Kumar Acharjee, The University of Southern Mississippi Abstract:
Biomedical and public health studies frequently follow individuals who may experience several types of recurrent
events before a terminal event such as death. Joint analysis of these outcomes must account for within-subject
recurrence, event-specific unobserved heterogeneity, and dependence among the recurrent- and terminal-event
processes. We propose a Bayesian semiparametric multivariable frailty model for multitype recurrent events
and a terminal event. Recurrent events are modeled on the gap-time scale, whereas the terminal event is modeled
on the calendar-time scale. A correlated multivariable gamma frailty vector assigns a distinct latent
susceptibility to each recurrent event type and to the terminal event, while its correlation structure
captures dependence among the processes. Outcome-specific regression coefficients allow covariate effects
to differ across outcomes. To accommodate variation across successive recurrences, the recurrent-event
baseline cumulative hazards may vary by both event type and gap number. Gamma-process priors provide
flexible semiparametric estimation of these baseline functions without imposing restrictive parametric forms,
and posterior inference is conducted through Markov chain Monte Carlo computation. Simulation studies evaluate
the operating characteristics of the proposed method. We apply the framework to the Reasons for Geographic and
Racial Differences in Stroke (REGARDS) study to jointly examine recurrent myocardial infarction, heart failure,
and coronary revascularization together with all-cause mortality. The proposed approach provides an interpretable
and flexible framework for analyzing multitype recurrent-event and terminal-event data across a broad range of
biomedical applications.
High-dimensional Proportional Hazards Model with the Horseshoe+ Prior
Zhuanzhuan Ma, University of Texas Rio Grande Valley Abstract:
We study the use of horseshoe+ prior, a continuous shrinkage prior, to the high-dimensional proportional hazards
(PH) model for analyzing right censored data. This work represents the first attempt to apply a continuous
shrinkage prior in the setting of high-dimensional survival models and to compute the maximum a posteriori
(MAP) estimator. Since the objective function is highly nonconcave, direct optimization of the posterior is
infeasible. To address this challenge, we propose a hybrid minorize–maximize (MM) algorithm for optimizing the
model parameters, which are partitioned into blocks containing nonoverlapping parameters. This decomposition,
together with the careful construction of the minorizing function and the componentwise log-concavity of the
objective function, greatly simplifies the maximization problem. Simulation studies demonstrate the effectiveness
of the proposed algorithm, with favorable comparisons to several popular methods for high-dimensional survival
models. Moreover, a near-optimal posterior contraction rate for the high-dimensional PH model is established,
and a convergence rate for the MAP estimator is provided. Finally, we illustrate the practical utility of the
proposed approach by identifying significant genetic biomarkers in an ovarian cancer study.
This work is joint with Mai Dao (Wichita State University), Bo Yu-Chien Ning (Texas A&M University), and Samiran Sinha (Texas A&M University).
High-Dimensional Survival Analysis: Which Methods Work for Feature Selection and Prediction?
Samiran Sinha, Texas A&M University Abstract:
High-dimensional survival data present challenges for both feature selection and prediction in the presence of
right censoring. This talk compares several statistical and machine-learning methods through extensive simulations
and an application to bladder cancer gene-expression data. The results highlight how performance depends on sparsity,
signal strength, and predictor correlation, and provide practical guidance for analyzing high-dimensional
survival data.
Semi-Supervised Learning for Measurement Error Correction in Survival Outcomes via Gaussian Process Survival Models
Yu-Chien Bo Ning, Texas A&M University Abstract:
Large-scale electronic health records (EHRs) provide an invaluable resource for clinical research, yet their data
quality is often compromised by measurement error that can bias statistical estimation. Motivated by this problem,
we develop a Bayesian framework to correct measurement error in survival outcomes under a semi-supervised learning
(SSL) setting. The data consist of a validation dataset, containing a relatively small number of observations with
both error-prone and error-free survival times, and a main study dataset, containing a large number of observations
with only error-prone survival times. For example, a typical dataset has a sample size of hundreds or thousands for
the validation study and more than ten thousand for the main study. At the core of our framework is a semiparametric
Gaussian process (GP) survival model, which corrects for the measurement error and predicts error-free survival times
for the main study. We demonstrate the effectiveness of the proposed framework through finite-sample simulations
under varying degrees of measurement error in survival time and misclassification in censoring status. We further
establish posterior contraction results for both the proposed SSL method and a Bayesian nonparametric deconvolution
approach to the survival density, and show that the relative efficiency of the two methods depends on the ratio of
validation-to-main-study sample sizes and the degree of ill-posedness induced by measurement error. Finally, we apply
the proposed method to the dataset from Yale New Haven Hospital.
This work is joint with Molin Wang (Harvard T. H Chan School of Public Health) and Ge Lin (Indiana University), and Raymond Carroll (Texas A&M University).
Quantifying Mediated Treatment Effects to Support Accelerated Approval Decision-Making in Oncology
Yiqing Chen, Texas A&M University Abstract:
Cancer is a leading cause of death worldwide and in the United States, responsible for nearly 10 million
and over 600,000 deaths annually, respectively. The relationships among treatment, treatment
response, intermediate outcomes, and survival endpoints are often complex in clinical settings. The FDA’s
Accelerated Approval (AA) pathway allows drugs to be approved based on surrogate endpoints that can be assessed
earlier than overall survival (OS), enabling faster access to promising therapies for patients in urgent need.
However, nearly 20% of indications have been withdrawn from AA after failing to confirm survival benefit in phase
3 trials. The FDA has strengthened guidance in recent years to encourage faster confirmatory trials and expedited
withdrawals when needed, as seen in the 2023 FDORA legislation. This work proposes novel quantitative framework
to support AA decision-making which explicitly evaluates the putative evidence for response “outperformance”
through mediation of progression and death. The method extends traditional mediation analysis of the response
to survival relationships using a multi-state framework that integrates progression and recurrence as intermediate
outcomes. In simulation studies across six mechanistic scenarios, the proposed framework achieved 87% decision
accuracy compared to 62% under traditional ORR-based criteria, with particular gains in scenarios where survival
benefit was absent or not captured by the surrogate endpoint.
From Insight to Impact: Agentic AI and the Decision Architecture of Trustworthy Action
Ari Zitin, SAS Institute Abstract:
As AI systems move from generating insights to executing decisions, the central challenge facing organizations is no
longer model accuracy alone, but decision quality at scale. When models act, inference becomes operational. This talk
examines how the rise of agentic AI changes the role of statistics, shifting emphasis from insight generation to the
design of decision systems that are trustworthy, explainable, and accountable.
Building on recent domain-specific advances in areas such as clinical research and public health, the talk introduces
decisioning as the missing architecture that connects statistical rigor to real-world impact. Decisioning defines how
model outputs are translated into actions through evidence, constraints, human oversight, and continuous validation.
Using illustrative examples across multiple industries, the keynote highlights where statisticians play a critical role as
decision architects, shaping uncertainty thresholds, escalation logic, auditability, and long-term monitoring.
The session reframes agentic AI as a decision-system problem rather than a modeling problem, and argues that the
future impact of AI depends on how effectively statistical principles are embedded into the systems that act on our behalf.
Statistical Analysis of Factors Associated with Early Initiation of Breastfeeding in Bangladesh: Evidence from BDHS 2022
Nihar Ranjan Sikder, University of North Florida Abstract:
Early initiation of breastfeeding (EIBF), defined as initiating breastfeeding within one hour of birth, is an
important maternal and child health practice. This study investigates the prevalence of EIBF and identifies
factors associated with its practice among mothers in Bangladesh using data from the 2022 Bangladesh Demographic
and Health Survey (BDHS). The study included 4,758 women who had given birth. Descriptive statistics were used to
summarize EIBF practices and selected socioeconomic, demographic, maternal, infant, and healthcare-related
characteristics. Pearson chi-square tests were conducted to examine bivariate associations between EIBF and
the explanatory variables. A binary logistic regression model was then fitted to identify factors significantly
associated with EIBF while accounting for multiple covariates. The regression analysis showed that skin-to-skin
contact immediately after birth was significantly associated with increased odds of EIBF (OR = 1.619, p < 0.001).
In contrast, cesarean delivery was associated with substantially lower odds of EIBF (OR = 0.404, p < 0.001).
Compared with home deliveries, deliveries in government facilities (OR = 0.577, p < 0.001) and private/NGO
facilities (OR = 0.528, p < 0.001) were also associated with lower odds of EIBF. Significant regional variation
in EIBF practices was observed across the administrative divisions of Bangladesh. Maternal education, age, and
wealth index were also statistically significant predictors after accounting for other factors in the model.
The findings demonstrate how statistical methods can be used to identify important factors associated with
breastfeeding practices in nationally representative survey data. The results particularly highlight skin-to-skin
contact, mode of delivery, healthcare facility type, and geographic variation as important areas for further
investigation and targeted interventions to improve EIBF practices in Bangladesh.
Fourier Ratio Bounds and Sensitivity to Sensor Placement
Christopher Housholder, Missouri State University Abstract:
Many problems in signal recovery, sampling, and learning depend on understanding when a function is concentrated
in frequency space. Classical sparsity assumptions are often too rigid for smooth continuous data and especially
so after the data is discretized onto a finite grid. Recent work develops the notion of the Fourier Ratio which
uses the \ell^1 and \ell^2 norms of the Fourier transform as a measure of spectral compressibility to give a
sparsity-free way to measure the effective size of Fourier support. We further study these estimates in several
directions. First, Holder regularity C^{0,\gamma} gives a fixed-grid bound of order N^{1-\gamma}\sqrt{\log N},
and fractional Sobolev regularity W^{s,p} with 0<s<1 and sp>2 gives the corresponding exponent
N^{1-s+2/p}\sqrt{\log N} through Sobolev--Morrey embedding. Then we introduce a small pre-discretization
perturbation in which the continuous function is shifted before being sampled on the grid. We use this shift
to choose a discretization with better L^2-mass control, which leads to a sharper Fourier Ratio bound in terms
of the L^2 energy of the gradient. We then extend the same energy method to compact Riemannian manifolds using
Laplace-Beltrami spectral decompositions, obtaining a sharp \sqrt{\log L} critical estimate on compact surfaces
and an explicit version on the sphere S^2 using spherical harmonics. Finally, we study sensitivity to sensor
placement by treating independent sensor displacements as measurement error and combine the resulting estimates
with random sampling to prove corresponding stable signal recovery results on the periodic square and for
bandlimited functions on compact manifolds. These estimates give a more flexible version of the original
discretization method and suggest further directions for smooth data sampling under less rigid geometric
assumptions.
Distribution and Quantile Estimation in High-Dimensional Sparse Linear Models
Shuxia Sun, Wright State University Abstract:
Assumptions for the distributions of random errors play an essential role in the statistical inference of regression
models, and residuals are often used to construct related statistics to assess or justify those assumptions.
In the setup of high-dimensional sparse models, where the number of covariates exceeds the sample size, residuals
can be highly affected by the selected model due to the presence of spurious variables. The recently introduced
refitted cross-validation procedure lessens the influence of spurious variables. In this paper, we propose error
distribution and quantile estimators in high-dimensional linear regression models and show that, under the
assumptions that non-zero coefficients are sparse and that the model selection has a sure screening property,
these estimators based on residuals from the refitted cross-validation method possess strong consistency
properties. Furthermore, our simulation studies indicate that the proposed two-stage refitted procedure improves
the performance of both estimators.
This work is joint with Fuxia Cheng (Illinois State University) and Yang Liu (Wright State University).
Semiparametric Validation of Earthquake Path Attenuation Using Perception-Based and Instrumental Data
Junhui Yang, Loyola Marymount University Abstract:
Ground-motion models describe how earthquake shaking varies with earthquake
characteristics, source-to-site distance, and local site conditions, and they are fundamental
to seismic hazard assessment and earthquake risk prediction. An important component of
these models is path attenuation, which characterizes how shaking decreases with distance
as seismic waves propagate away from an earthquake source. Because ground-motion
models are largely developed empirically from recordings at seismic stations, their
uncertainty can be substantial in regions where instrumental observations are sparse.
To supplement limited instrumental data, we investigate perception-based earthquake
shaking reports collected through the U.S. Geological Survey’s “Did You Feel It?” (DYFI)
system. DYFI data can provide much denser spatial coverage in populated areas because
observations are contributed by people who experienced the earthquake. However, these
reports are subjective measures of perceived shaking rather than direct physical
measurements of ground motion. Their suitability for quantitative ground-motion modeling
therefore requires careful statistical validation.
Using well-recorded California earthquakes, we first transform DYFI intensity observations
to equivalent peak ground acceleration and adjust for local site e_ects to place the
perception-based and instrumental data on a comparable scale. For each data source, we
then estimate the distance-dependent attenuation function using flexible semiparametric
regression, allowing the attenuation pattern to vary smoothly with distance without imposing
a pre-specified parametric decay form. We formulate the comparison as a statistical
hypothesis test of whether the two underlying attenuation functions are equal over the
distance range supported by both datasets. A resampling-based testing procedure, or
equivalently a comparison between common-curve and source-specific smooth models, is
used to quantify discrepancies between the estimated functions while accounting for
sampling variability. This framework moves beyond visual comparison of fitted attenuation
curves and provides formal statistical evidence for assessing whether perception-based
observations capture the same underlying path-attenuation behavior as instrumental
measurements. More broadly, the study illustrates how flexible statistical inference can be
used to validate heterogeneous data sources and evaluate whether dense but indirect
observations can reliably supplement sparse direct measurements in scientific modeling.
This work is joint with Pengfei Wang (Old Dominion University).
Credible Rank Distributions for Bayesian Unit-Level Small Area Estimation
Arghadeep Basu, University of Georgia Abstract:
Small area estimation often aims not only at predicting area level means but also at ranking areas to inform
resource allocation and help in making important policy decisions. Yet rankings derived from point estimates
alone ignore the substantial uncertainty inherent in small sample settings, potentially leading to misleading
priorities. In this paper, we develop a hierarchical Bayesian framework for unit-level data that extends the
classical nested-error regression model by incorporating flexible spatial and random-effect structures, including
penalized splines, area specific random regression coefficients, and spatially correlated intercepts. Then we
incorporate a joint posterior ranking methodology from the full joint posterior distribution of small area means
and construct a doubly stochastic ranking matrix whose entries give each area's posterior probability of
occupying every possible rank. Paired with simultaneous credible sets, both Cartesian and elliptical, this
allows practitioners to make probabilistic statements about relative orderings rather than relying on a single
point ranking. We test our proposed models on a unified data generating mechanism based on estimation and ranking
metrics. We then apply the frameworks to Acid Neutralizing Capacity measurements from 113 eight-digit Hydrologic
Unit Code watersheds in the northeastern United States, identifying the high and low ranking basins within the New
England, Mid-Atlantic, and Great Lakes regions.
A Hybrid MCMC-Optimization Framework for Efficient Sampling of Multimodal Bayesian Posterior Distributions
Lakshana Wickramasinghe, Lamar University Abstract:
Multimodal posterior distributions can arise from non-identifiability, parameter constraints, model misspecification,
and complex nonlinear dependence structures in the likelihood. In particular, high-dimensional copula models can
introduce complex nonlinear dependence structures into the joint likelihood, potentially producing highly
multimodal posterior distributions. Unlike optimization methods that seek a point estimate, MCMC aims to
characterize the full posterior distribution, including the relative probability mass and uncertainty associated
with multiple modes. These modes may represent distinct parameter regimes or competing explanations of the
observed system, making it important to adequately explore each mode and quantify its associated uncertainty.
However, conventional MCMC methods can mix poorly in multimodal posteriors, often becoming trapped within
individual modes and failing to adequately explore the posterior mass across modes. We propose an adaptive
MCMC framework that combines local proposals for efficient within-mode exploration with global proposals
designed to facilitate transitions between modes, using covariance-matrix adaptation to guide exploration of the
parameter space. We evaluate the proposed approach on synthetic multimodal posterior distributions and demonstrate
its ability to improve mode exploration and posterior characterization relative to conventional MCMC methods.
A Comparative Study of Vector Autoregressive, VAR, Bayesian Vector Autoregressive, BVAR, and Mixed-Frequency Bayesian Vector Autoregressive, MBVAR Models for Macroeconomic Forecasting Abstract:
Shree Rajkumar Saha, The Graduate Center City University of New York Abstract:
Vector autoregressions are the workhorse of macroeconomic forecasting. Still, the classical specification exhausts
degrees of freedom as the dimension grows, and aggregating monthly indicators to the quarterly frequency at which
GDP is published discards information. We compare three frameworks on four U.S. series from FRED and ALFRED: a
classical VAR with VECM and SVAR analysis, a Bayesian VAR under Minnesota and steady-state priors estimated by
Gibbs sampling, and the mixed-frequency Bayesian VAR of Schorfheide and Song (2015) and Ankargren et al. (2020).
Quarterly estimation covers 1959:Q1–2021:Q3 (T = 250); the mixed-frequency specification uses 488 monthly
observations over 1980–2020, a sample that includes the COVID-19 disruption. Beyond comparing impulse responses
and posterior diagnostics, we test forecast-accuracy differences formally using Diebold–Mariano statistics with
the Harvey–Leybourne–Newbold small-sample correction and the Clark and West (2007) adjustment for the nested
VAR/BVAR pair. Bayesian shrinkage improves short-horizon accuracy and uncertainty quantification relative to
the classical VAR, and the mixed-frequency specification yields tighter credible sets for GDP growth and inflation.
Estimating Two Missing Observations in Two-Way Fixed-Effects Linear Model Data
Mohammad Patwary, Butler University Abstract:
Accurately estimating missing observations in a two-way fixed-effects linear models is essential for preserving
the integrity of statistical analyses, ensuring valid interpretation of factor effects and interactions, and
maintaining the reliability of experimental conclusions. This study introduces a straightforward yet innovative
methodology for estimating multiple missing values within the framework of two-way fixed-effects linear models.
Special emphasis is placed on three distinct missingness scenarios involving two missing observations. The
proposed estimators are derived analytically by minimizing squared error loss, resulting in robust and
computationally efficient solutions. Numerical optimization is applied to verify the optimality of the estimators.
Simulation studies under varying levels of replication are conducted to assess the performance of the estimators
using bias, variance, mean absolute error (MAE), and mean squared error (MSE). Additionally, the influence of
different missing data patterns is analyzed to evaluate their impact on estimation accuracy. The findings highlight
the effectiveness and flexibility of the proposed methodology and lay the groundwork for future extensions to more
complex factorial designs and high-dimensional linear models.
From Planning Steps to Candidate Polymers: Completion-Based Process Rewards for Property-Constrained TSMP Generation
Borun Das, University of Louisiana at Lafayette Abstract:
Designing polymers for target physical properties is challenging because many chemically distinct monomer pairs
may satisfy the same glass-transition temperature and rubbery modulus, while others may be incompatible or
unlikely to achieve the target. Generative models can explore this combinatorial space, but feedback usually
arrives only after a complete candidate is produced. This outcome-level feedback makes it difficult to determine
which design decisions should be reinforced, revised, or avoided during generation. Process reward models (PRMs)
offer a way to provide intermediate feedback, but existing settings do not transfer directly. In mathematics
and code, intermediate reasoning can often be assessed relative to an exact final answer; in retrosynthesis,
meaningful reaction steps are already naturally defined. De novo polymer design has neither. We therefore
formulate the design of two-monomer thermoset shape-memory polymers as a sequence of mechanically checkable
chemistry decisions, including reaction family, structural rigidity, functionality, and molecular flexibility.
A Planner proposes these decisions from target properties, while a separate Completer generates candidate
monomer pairs from successive planning states. Each planning state is evaluated through completion-based rollouts.
Multiple candidate pairs are generated and checked for molecular validity, reaction compatibility,
plan consistency, and predicted property agreement. Their success rate provides an automated,
completion-based estimate of downstream feasibility and property success without requiring manual step-level
annotations. Using 1,119 reported two-monomer shape-memory polymer systems, our preliminary study examines
whether structured plans can represent multiple feasible molecular designs and whether completion-based scores
can distinguish more and less promising planning states. More broadly, the framework explores process
supervision for scientific design problems with many feasible solutions rather than a single correct answer.
This work is joint with Andrew Peters ( Louisiana Tech University ),
Guoqiang Li ( Louisiana State University ), and Xiali Hei (University of Louisiana at Lafayette)
How Emotional Intelligence and Positive Mental Health Drive Academic Performance: A Study of National University Undergraduates
Sajib Sarker, University of North Florida Abstract:
This study examined the relationships among emotional intelligence (EI), positive mental health (PMH), and academic
performance (AP) among undergraduate students of National University, Bangladesh, with particular attention to the
mediating role of PMH.
A quantitative cross-sectional study was conducted among 210 full-time undergraduate students aged 18–29 years from
seven colleges affiliated with National University. Participants had completed at least one academic year and had
an available official CGPA. EI and PMH were assessed using structured five-point Likert-scale questionnaires,
while AP was measured using official CGPA records. Descriptive analyses were conducted using IBM SPSS 29.0, and
Partial Least Squares Structural Equation Modeling (PLS-SEM) was performed using SmartPLS 4. The measurement model
was evaluated for reliability and validity, followed by structural model assessment using 5,000 bootstrap samples.
The mean participant age was 23.08 years; 51.0% were male and 49.0% female. Most participants (64.8%) were aged
22–25 years, and 86.2% reported a middle-income background. The measurement model demonstrated satisfactory
reliability and validity, with AVE values ranging from 0.708 to 0.844 and Cronbach’s alpha values from 0.890 to 0.962.
EI was positively associated with PMH (β=0.657, p<0.001) and AP (β=0.453, p<0.001). PMH was also positively
associated with AP (β=0.288, p<0.001). Moreover, PMH significantly mediated the relationship between EI and
AP (β=0.189, p<0.001; 95% BC CI: 0.107–0.296).
EI and PMH were significantly associated with academic performance among National University undergraduates.
The mediation finding suggests that stronger emotional intelligence may contribute to better academic outcomes
partly through improved positive mental health. Integrating EI development and mental health support into higher
education may therefore help promote students’ academic success and psychological well-being.
Modeling Kidney Health Biomarkers Using Blood Heavy Metal Levels From Complex Survey Data
Sarah Czuba, High Point University Abstract:
Kidney disease affects millions of adults worldwide, and environmental heavy metal exposure has been
documented to be associated with impaired kidney function. Predictive models based on National Health
and Nutrition Examination Survey (NHANES) data have often ignored the complex survey design or relied
solely on survey-weighted linear regression without assessing whether model assumptions are satisfied,
limiting the reliability and generalizability of their findings. In this study, NHANES 2021–2023 data from
adults aged 20 years and older were used to evaluate associations among demographic characteristics,
clinical risk factors, blood heavy metal concentrations, and four kidney health biomarkers: albumin-tocreatinine
ratio (ACR), blood urea nitrogen (BUN), uric acid (UA), and estimated glomerular filtration
rate (eGFR).
Survey-weighted linear regression models were compared with several weighted statistical and machine
learning approaches, including LASSO, Ridge Regression, Elastic Net, Random Forest, and Support
Vector Regression. Model performance was evaluated using survey-weighted root mean squared error
(WRMSE) estimated through k-fold cross-validation that accounted for the complex NHANES sampling
design. Our findings suggest that more flexible modeling approaches may improve predictive accuracy for
kidney health biomarkers while maintaining the possibility for population-representative inference.
This work is joint with Michael Larson and Pujita Sapra (High Point University).
Machine Learning Approaches for Early Prediction of Ventilator-Associated Pneumonia Using EHR Data
Bao Maddux, Winston-Salem State University Abstract:
Ventilator-associated pneumonia (VAP) is a leading cause of ICU morbidity, prolonged mechanical ventilation,
increased costs, and mortality. Standard diagnostic criteria suffer from diagnostic lag, delaying timely management.
We developed and evaluated supervised machine learning (ML) models using multicenter Electronic Health Record
(EHR) data to predict VAP within the first 48 hours of ventilation.
Using the NIH AIM-AHEAD Bridge2AI CHORUS dataset, adult ICU patients (age >18 years) mechanically ventilated for
>48 hours were identified. Patients with baseline pneumonia, pre-existing tracheostomy, early broad-spectrum
antibiotics, or ECMO were excluded. VAP was defined via standardized SNOMED codes (Concept ID: 259992) or
post-48-hour targeted antibiotic initiation within 30 days. Demographics, chronic comorbidities, and
biophysical/lab values from the initial 48 hours of ventilation were extracted. Eight ML models across regression,
tree-based, and neural architectures were trained on an 80/20 train-test split. Models were calibrated using
isotonic regression; discrimination was evaluated by AUC-ROC, optimal thresholds by Youden's J statistic,
clinical net benefit by Decision Curve Analysis (DCA), and feature importance by SHapley Additive exPlanations
(SHAP).
Of 3,660 included patients, 510 (13.9%) developed VAP. VAP patients showed higher baseline rates of neurological
(41.2% vs. 15.3%), respiratory (69.8% vs. 45.1%), and hematologic (73.5% vs. 49.4%) comorbidities. AUC-ROC ranged
from 0.732 to 0.776 across all models. Tuned Random Forest achieved the highest discrimination (AUC-ROC 0.776,
sensitivity 72.5%, specificity 71.4%, NPV 94.1%), followed by XGBoost (AUC 0.752) and SVM (AUC 0.749). DCA confirmed
net clinical benefit across threshold probabilities of 5–35%. SHAP analysis identified subacute neurological
conditions, hematologic and respiratory comorbidities, blood urea nitrogen, hemoglobin, chloride, and core
temperature dynamics as key predictors.
Data Subsampling Strategies
John Stufken, George Mason University Abstract:
Subsampling from large datasets has attracted growing attention over the past decade in both the computer science and
statistics literature, leading to the development of a wide range of methods. After a brief high-level overview, this
presentation will examine two representative approaches in greater detail: one model-based and one model-free.
The model-based method is designed to achieve efficient estimation of model parameters, whereas the model-free
approach emphasizes accurate prediction of future observations.
The Rainbow Number for x + y = z2 in Zp2
Lucas Danburg and Ian Lundy, Fairfield University Abstract:
The rainbow number rb(S, eq) for an equation eq defined on a set S is the minimum number
r such that any partitioning of S into r non-empty sets (an “r-coloring”) contains a solution to
eq with each of its variables in a different set. We give the rainbow number for x + y = z^2 in
Z_{p^2} and Z_{p^3} for prime p. We additionally provide lower bounds for this rainbow number in Z_{p^k}
and Z_{p_1···p_n} for distinct primes p_1, . . . , p_n, and show that these results generalize to x+y = az^2
for certain a.
This work is joint with Zhanar Berikkyzy (Fairfield University), Emily Codere (Westminster College), Sycamore Herlihy (Iowa State University),
Austin Longin (Iowa State University), Ian Lundy (Fairfield University), and Amy Wang (University of Maryland)
The Sperner Scoring Game: From Parities to Probabilities
Simon Harnik, Bronx High School of Science Abstract:
We introduce the Sperner Scoring Game, a two-player game played on the triangulation of a d dimensional
simplex subject to Sperner boundary conditions (each outer vertex is a different color, boundary vertices
color must be one of the colors from the vertices of the [d-1] dimensional simplex containing it, while
interior vertices can have any color). Players alternate coloring vertices, with one seeking to maximize
the number of fully-colored (FC) d dimensional simplices, and the other seeking to minimize that quantity.
We derive optimal strategies for the one-dimensional game. The parity arguments there apply but not fully
to larger d; in those cases we study random play through the expected value and variance of the number
of FC simplices. These probabilistic methods suggest scoring choices to make the games interesting, and
if time permits we discuss progress towards optimal strategies in 2 dimensions. This work is joint with
Steven J. Miller (Williams College)
Binomial Edge Ideals and the Lefschetz properties for certain graph families
Gasper Ndokaj, Vassar College Abstract:
Given a graph G = (V,E) with vertices V = {1,2, \ldots, n} and edge set E, the binomial edge ideal J_G is
defined in the polynomial ring K[x_1, ..., x_n, y_1, ... y_n] as:
J_G = (x_i y_j-x_j y_i | (i,j) \in E).
We studied how certain combinatorial features of a graph G determine the weak and strong Lefschetz properties of
graded algebras defined by J_G. We proved that the graded algebras defined by the binomial edge ideals of path
graphs, cycle graphs, and complete graphs with one or two attached paths have the strong Lefschetz property. We also
proved that the weak Lefschetz property holds for graded algebras associated with the binomial edge ideals of two
complete graphs joined by a path, and possibly with additional paths attached, when the total length of all paths
is odd. From these results, we showed how specific combinatorial conditions and operations on graphs govern the weak
and strong Lefschetz properties of graded algebras defined by the binomial edge ideals of graphs. This work is joint
with the 2026 Polymath Jr Binomial Edge Ideals Group.
The double descent behavior in two layer neural network for binary classification
Hailin Sang, University of Mississippi Abstract:
Recent studies observed a surprising concept about test error called the double descent phenomenon where the
increasing model complexity decreases the test error first and then the error increases and decreases again.
To observe this, we worked on a two-layer neural network model with a ReLU activation function designed for
binary classification under supervised learning. Our aim was to observe and find the mathematical theory behind
the double descent behavior of the test error in the model for varying over-parameterization and under-parameterization
ratios. We have been able to derive a closed-form solution for the test error of the model and a theorem to find
the parameters with optimal empirical loss when model complexity increases. We proved the existence of the double
descent phenomenon in our model for square loss function using the theorems derived.
This talk is based on the joint work with Chathurika Abeykoon and Aleksandr Beknazaryan.
Functional canonical correlation for discriminant analysis
Xin Dang, University of Mississippi Abstract:
Canonical correlation analysis (CCA) is a classical tool for quantifying the dependence between two random vectors,
X in Rp and Y in Rq. When X and Y are functional observations, functional canonical correlation analysis (FCCA)
extends this framework to characterize the linear association between random functions. In this paper, we develop
a novel FCCA framework tailored for discriminant analysis with a categorical response variable Y. The proposed
dependence measure is constructed by extending the Gini distance correlation to the functional setting, enabling
a rigorous quantification of the association between a functional predictor and a categorical outcome. We establish
key theoretical properties of the proposed estimator, including its consistency. Extensive simulation studies
demonstrate that the proposed method provides accurate and stable estimation of the underlying dependence structure
and exhibits strong performance in discriminant analysis tasks. The practical utility of the approach is further
illustrated through a real data application, where a K-sample testing problem is considered.
Grouped feature screening for ultrahigh-dimensional classification via Gini distance correlation
Yongli Sang, University of Louisiana at Lafayette Abstract:
Gini distance correlation (GDC) was recently proposed to measure the dependence between a categorical variable, Y, and a numerical random vector,
X. It mutually characterizes independence between X and Y. In this work, we
utilize the GDC to establish a feature screening for ultrahigh-dimensional discriminant analysis where the
response variable is categorical. It can be used for screening individual features as well as grouped features.
The proposed procedure possesses several appealing properties. It is model-free. No model specification is
needed. It holds the sure independence screening property and the ranking consistency property. The proposed
screening method can also deal with the case that the response has divergent number of categories. We conduct
several Monte Carlo simulation studies to examine the finite sample performance of the proposed screening procedure.
Real data analysis for two real life datasets are illustrated.
Patients Satisfaction and Quality of Services at Community Clinics: An in-depth Study in Haor Region
Md. Shahjahan, Daffodil International University, Dhaka Abstract:
Community clinics (CCs) were established to ensure easy access to health services in underserved areas in Bangladesh.
This study employed a cross-sectional study with mixed-method to explore satisfaction level of patients and
investigates healthcare services quality in Haor region of Sunamganj district of Bangladesh. PSQ-18 framework
was used to measure - general satisfaction, technical quality, communication, and accessibility. Qualitative data
was gathered through 16 KIIs and 4 FGDs to explore challenges and opportunities of services. KIIs were conducted
with individuals holding strategic positions in healthcare. Quantitative data were analyzed using SPSS for descriptive
and inferential statistics, while qualitative data were thematically analyzed to capture contextual insights into
service quality and user experiences. Quantitative findings showed moderate overall satisfaction. Only 18.8%
respondents reported consistent emergency service availability, and nearly 70% had to purchase medicines from
pharmacies due to stock-outs. Both community members and health officials emphasized that effectiveness of community
clinics depends on accessibility, affordability, and dedication of healthcare providers.
From qualitative data, vital role of primary healthcare struggle to provide services due to limited manpower, weak
supervision, inadequate infrastructure, insufficient community and administrative support, which requires a coordinated
approach including strengthening staffing structures, enhancing training and supervision, improving infrastructure, and
promoting local ownership. Under-utilization of some community clinics was observed due to a combination of geographic
isolation, weak infrastructure, staff shortages, inconsistent operations, and limited community engagement. Findings
underscore need for policy reforms and targeted interventions to ensure equitable and high-quality primary healthcare
in geographically vulnerable regions of Bangladesh.
Artificial Intelligence and Disability: A Path toward Inclusive Innovation
Mohd. Arifuzzaman, North Western University Abstract:
Artificial Intelligence (AI) is opening an opportunities for persons with disabilities to enhance their
accessibility, independence and social inclusion. Communication, mobility, education, health care and everyday living
will be supported by new technologies like natural language processing, computer vision, robotics, adaptive learning
and smart environments. Yet, concerns expresses about bias, privacy, unequal access and limited participation of
persons with disabilities in development of AI. To raise awareness about the importance of AI for
disability inclusion, to identify opportunities and challenges, and suggest priorities for policy and research
Narrative review based on evidence from recent literature, global reports, policy documents and case
examples. Topics ranged from communication, vision, mobility, learning and daily life to ethical and governance
issues in AI applications. People with hearing or speech impairments can communicate better with the help
of AI-powered speech-to-text, text-to-speech, translation and voice interfaces. The computer vision applications
for object recognition and image understanding are useful to the visually impaired users. Smart wheelchairs,
exoskeletons and navigation devices can be used to improve mobility, and adaptive learning systems can provide
individualized educational help. Independence can also be supported by smart homes and virtual assistants.
Major challenges include algorithmic bias, lack of accessibility standards, privacy risks, digital inequality
and underrepresentation of persons with disabilities in design and decision-making. AI has the
potential to greatly boost disability inclusion, but it must be developed ethically, accessibly and inclusively.
Policymakers, academics, developers and health professionals must promote equal access, strong accessibility
requirements and the active participation of persons with disabilities in the development of AI.
Integration and Implementation Challenges of Shared Health Records in Bangladesh
Mohammad Rashid, Directorate General of Health Services, DGHS Abstract:
Shared health records (SHRs) are a key component of integrated patient-centered digital health systems
that facilitate the sharing and collection of health information across different systems. Bangladesh has multiple
digital health efforts like DHIS2, EMID, and e-IMS, but these systems are being established independently, limiting
interoperability and the development of a comprehensive national SHR system. Considerate the barriers to SHR
integration are important for digital health progress and for the attainment of UHC. To identify and
analyze the main issues in the integration and use of SHRs in the public health system of Bangladesh. A
descriptive cross-sectional study was conducted from July to December 2023. A total of 120 individuals selected health
managers, clinicians, ICT professionals, and policy officials. Data were collected by semi-structured surveys, KIIs,
and FGDs. All the participants had previous training in digital health, but only 8.3% had an understanding
of the topics of SHR. The primary issues noted were lack of interoperability (83.3%), unstable internet connectivity
(83.3%), usability concerns (57.5%), limited inter-agency cooperation (56.7%), and lack of leadership support (55%).
74.2% said they do not have ICT training. The lack of data protection regulations seriously hindered 74.2% of the
participants. However, 75.8% were in favor of mandatory SHR deployment in public health facilities.
Bangladesh has major constraints in infrastructure, governance, human resources, and legal systems in
the introduction of SHR. The government has to invest in improving the digital infrastructure, developing
interoperability standards such as HL7/FHIR, enhancing worker competencies, and creating robust data governance.
Addressing these systemic barriers and scaling up SHRs will effectively pave the way to universal health care and a
patient-centric digital health ecosystem in Bangladesh.
Quality of Emergency Care in Selected Hospitals: A Cross-Sectional Study
Mohammad Islam, Ministry of Health and Family Welfare Abstract:
Bangladesh has made impressive progress in improving maternal and child health. However, the quality
of care in public health facilities remains a major concern and continues to limit progress toward universal health
coverage, particularly for disadvantaged communities. This study assessed the quality of emergency
healthcare services in government hospitals and health centers, with a focus on patients’ expectations,
experiences, satisfaction, and the main factors influencing service quality. A mixed-method
cross-sectional study was carried out in two Medical College Hospitals, two District Hospitals, and four Upazila
Health Complexes. Multistage and purposive sampling were used. Data were collected through 62 in-depth interviews
with healthcare providers and policymakers and exit interviews with 260 patients or attendants. There
were significant gaps in staffing, equipment, medicines, oxygen supply, diagnostic services, hygiene, ambulance
support and referral systems. Morning hours were frequently the only time X-ray and lab services were available,
and there were common shortages of reagents and other supplies. Healthcare providers said they had heavy
workloads, poor working conditions, staff shortages and problems referring patients. 76% of service recipients
rated emergency services poorly and 85% reported low satisfaction. Overall trust was low although 71% were
somewhat satisfied with emergency department management. Shortages of doctors, medicines, equipment, and
diagnostic services were common reasons for patients to seek care elsewhere. Communication, confidentiality
and behavior of staff were generally acceptable. Improving emergency care will require adequate
staffing and funding, reliable supplies of medicines and oxygen, stronger diagnostic and ambulance services,
better referral systems, faster response, shorter waiting times, and greater public awareness.
Local and Global Optimization – A Decision Tree Perspective
MB Rao, University of Cincinnati Abstract:
Local and global optimization problems crop up in many statistical paradigms. Prominent areas are likelihood estimation, cluster analysis; and model selection.
The presentation begins with a panoramic vista of the problem and then sails into data mining. Decision trees carved a niche into machine learning lore. Optimal trees
are built following Breiman, Friedman, Olshen, and Stone CART greedy algorithm. This is local optimization. We dare to enumerate all possible trees in a simple
environment and determine a globally optimal tree. The greedy algorithm should not get a pass.
This work is joint with Mehdi Fazel Najafabadi (University of Cincinnati College of Medicine), Sarmistha Das (University of Cincinnati College of Medicine and Cancer Center)
and Shesh Nath Rai (University of Cincinnati College of Medicine and Cancer Center).
Sampling With or Without Replacement: A Comparative Study
MB Rao, University of Cincinnati, and Tianyuan Guan, Kent State University Abstract:
Mathematical Statistics is taught at many undergraduate and graduate programs across the United States. This note celebrates the breadth and
depth of the subject. A typical course begins with probability theory, including discrete and continuous distributions, and then progresses to multivariate,
marginal, and conditional distributions, order statistics, and transformations. These topics provide the foundation for statistical inference, including
likelihood, unbiased estimation, sufficiency, completeness, and minimum variance of unbiased estimation. Key results such as the Rao–Blackwell theorem,
the Cramér–Rao inequality, and the Lehmann–Scheffé theorem help identify efficient estimators, while hypothesis testing completes the core sequence.
Numerous textbooks present these topics clearly, often supported by exercises that reinforce difficult concepts. In this note, we focus on a single
example that brings together several important ideas in Mathematical Statistics and illustrates how they interact in one unified problem.
Designing and Analyzing for Valid Inferences of Large Databases involving High-throughput Observations – Part I
Shesh Rai, University of Cincinnati Abstract:
Bioinformatics databases are essential for storing and sharing proteomics, genomics, transcriptomics, metabolomics data along with clinical data.
In this talk, we will highlight the issue of data volume, data integration, data quality and curation, data standardization, database maintenance and updating,
efficient extraction of data, and multi-omics data integration. We will discuss how to design studies utilizing large publicly and commercially available
databases such as BEAT Acute Myeloid Leukemia (AML), Molecular Taxonomy of Breast Cancer International Consortium (METABRIC), Database for Genotype and
Phenotype (dbGAP), Gene Expression Omnibus (GEO), Genomic Data Commons (GDC), GenBank, Library of Integrated Network-based Cellular Signatures (iLINCS),
and ARCHS4 plus others. Designing studies using each of these databases present unique challenges. We will discuss inferential aspects of data extracted
from these databases and provide recommendations for investigators working with molecular data. We will illustrate the issues using examples.
This work is joint with Shesh Rai, Sarmistha Das, Erin K. Hertlein, Sarah J. Skuli, Bipul Das, Jianmin Pan, Davendra P. S. Sohal, MB Rao, and Anand Seth.
Designing and Analyzing for Valid Inferences of Large Databases involving non-High-throughput Observations – Part II
Anand Seth, University of Cincinnati Abstract:
In this talk, we will discuss how to design studies utilizing large publicly and commercially available databases such as EPIC Cosmos, TriNetX,
The Surveillance, Epidemiology, and End Results (SEER), National Cancer Database (NCDB), Nationwide Readmissions Database (NRD), Veterans Affairs (VA)
Informatics and Computing Infrastructure (VINCI), and National Health and Nutrition Examination Survey (NHANES) among others. Designing studies using
each of these databases present unique challenges specific to a given database. We will discuss inferential aspects of such studies and provide recommendations
for investigators who want to utilize these databases for generating plausible hypotheses. We will illustrate the issues using examples.
This work is joint with Anand Seth, Bipul Das, Jianmin Pan, Davendra P. S. Sohal, MB Rao, and Shesh Rai.
The Future of Drug Development: AI as a Strategic Enabler
Hiya Banerjee, Eli Lilly and Company Abstract:
In large outcomes trials, independent, blinded adjudication of endpoints is an important and expected methodology to ensure
standardization of endpoint assessment, particularly for complex or subjective events where the subjectivity often relates to
the assessment of the cause of the event (e.g. hospitalization due to heart failure, CV death, MI, stroke). Adjudication is
performed by a Clinical Event Committee (CEC) consisting of impartial, clinical experts who review trial events in a blinded
fashion (i.e. without knowledge of the treatment allocation) using prespecified clinical definitions which ensure cohesive
outcome interpretation across varying local practices and investigator opinions. While critical to study conduct and interpretation,
CEC processes rely on manual review which is time consuming, prone to subjectivity, and resource intensive, often accounting for lot of money.
This presentation explores the use of natural language processing (NLP) and artificial intelligence to support and potentially
automate cardiovascular endpoint adjudication. We will discuss how AI-enabled adjudication can improve efficiency across clinical
trials, observational research, and quality improvement efforts, while also addressing current limitations of automated systems.
A novel multi-agent AI framework will be introduced, designed to mirror the structured, stepwise reasoning of human adjudicators
by breaking the adjudication process into specialized intermediate tasks.
IoT-Enabled Lean Six Sigma for Cleaner Production in Solar Photovoltaic Manufacturing: A Conceptual Framework for Sustainable Process Improvement
Ramisa Nawar Promee, First Solar, Inc. Abstract:
Lean Six Sigma (LSS) depends heavily on statistical process control, but its effectiveness is limited by how often
and reliably data can be collected. This paper proposes a conceptual framework that pairs Internet of Things (IoT)
sensing with the DMAIC methodology in solar photovoltaic (PV) manufacturing, using continuous sensor data instead
of periodic sampling to support real-time statistical quality management.
The framework was developed through a systematic literature review of 46 peer-reviewed studies published between
2018 and 2024, drawn from ScienceDirect, IEEE Xplore, SpringerLink, and Taylor and Francis Online. Drawing on
integration patterns from the automotive, semiconductor, and electronics industries, four components emerged as
central: IoT-enabled DMAIC integration, where sensor streams replace manual sampling; digital twin simulation for
virtual parameter optimization; IoT-driven predictive maintenance using multivariate time-series data; and
blockchain-based supply chain traceability for material and compliance tracking.
A cross-industry simulation applies documented improvement rates to First Solar's publicly reported Series 7
baseline metrics. Projected outcomes under full implementation include a 25% reduction in material waste rate
(6.2% to ~4.7%), a 15% decrease in energy intensity (0.24 to ~0.20 kWh/Wp), a 30% reduction in defect rate
(~320 to ~224 ppm), and roughly a 10% reduction in greenhouse gas emissions. Five falsifiable propositions
are proposed to guide future empirical testing. The paper argues that data collection frequency changes what
DMAIC can actually detect and control, with measurable consequences for environmental performance.
This work is joint with Tamjid Shafiq (Energy Institute of Louisiana, University of Louisiana at Lafayette) and Jim Lee ( University of Louisiana at Lafayette).
Recurrent Myocardial Infarction-Heart Failure Trajectories and Terminal Mortality
Mohammad Tariquel Islam, University of South Alabama Abstract:
Myocardial infarction (MI) and heart failure (HF) hospitalization may recur, alternate, and precede death,
but conventional time-to-first-event analyses do not capture these clinically important trajectories.
In this study, we aim to characterize recurrent MI–HF patterns and the timing of subsequent MI, HF hospitalization,
and death in the REasons for Geographic And Racial Differences in Stroke (REGARDS) cohort. Each qualifying MI or HF
hospitalization initiates a gap-time episode ending at the next MI, HF hospitalization, death, or non-death
censoring. Origin-specific Aalen–Johansen cumulative incidence functions estimate next-outcome probabilities
at 30 days, 1 year, and 3 years. Trajectories following first and later events are compared using participant-level
bootstrap contrasts. Cause-specific gap-time Cox models with participant-clustered robust variance assess
associations of the preceding event type and prespecified baseline characteristics with subsequent outcomes.
Preliminary data assessment identified substantial support for all six transitions: 4,749 MI→MI, 2,220 MI→HF, 3,
856 MI→death, 2,007 HF→MI, 1,532 HF→HF, and 1,573 HF→death. This next-event framework distinguishes recurrent
from cross-event progression and provides a clinically interpretable assessment of cardiovascular trajectories
and mortality beyond the first event.
Two Lenses on the Same Problem Space: Contrasting Predictive and Explanatory Statistical Approaches to E-Commerce Customer Behavior
Asadullah Ahmad, University of Louisiana at Lafayette Abstract:
Customer analytics in e-commerce is frequently approached from one of two directions: predictive modeling,
which identifies individual customers likely to churn, or exploratory and inferential statistical analysis, which
characterizes population-level behavioral patterns. Recent literature notes that these two directions are typically
studied in isolation, with limited work connecting individual-level prediction to population-level explanation.
This presentation merges two independently developed graduate projects to examine both directions applied to the
same underlying problem space.
The first project applies supervised classification methods (logistic regression, decision tree, and random forest)
to a simulated 300-customer e-commerce dataset with a 35% churn rate, using engineered behavioral and transactional
features and SMOTE to address class imbalance. Logistic regression achieved the strongest performance (F1 = 0.667),
and SHAP-based explainability was applied to support individual-level interpretation of model outputs.
The second project applies an exploratory and inferential statistical framework (OSEMN) to a simulated
15,000-observation behavioral dataset, using hypothesis testing (t-tests, ANOVA) across demographic segments,
correlation analysis, funnel/drop-off analysis across the purchase pipeline, principal component analysis
(reducing 26 behavioral variables to four components explaining 96% of variance), and k-means clustering to
identify four distinct buyer personas.
Presented together, these projects illustrate a methodological contrast between individual-level predictive
inference and population-level explanatory inference applied to the same domain, and motivate a discussion
of practical integration — specifically, whether combining churn prediction with persona-based behavioral
segmentation improves the actionability of retention-focused analytics beyond either approach in isolation.
The talk closes with a proposed direction for empirically testing this integration on real transactional data.
Toward Statistically Rigorous Benchmarking of Stochastic and Hyperdimensional Computing Methods for Resource-Constrained Edge Vision Abstract:
Mohammad Abdul Hannan, Independent Researcher
As promising alternative to conventional deep learning pipelines for edge vision, Stochastic computing (SC) and
hyperdimensional computing (HDC) offer substantial reductions in memory footprint, energy consumption, and
compute complexity. By encoding visual features as low-precision bitstreams or high-dimensional binary/bipolar
vectors, both paradigms introduce distinctive sources of statistical variability: bitstream correlation noise
in SC circuits and similarity-distribution behavior of random hypervectors in HDC. They are rarely addressed
with the same statistical rigor as their reported accuracy and efficiency gains. Most contemporary research
perceives these methods through single-run point estimates, and they leave open questions about the reliability,
reproducibility, and generalizability of reported performance improvements.
This proposal offers a statistical validation framework specifically designed for SC- and HDC-based edge vision
methods. It characterizes the variance introduced by SC bitstream correlation and by HDC’s random projection/encoding
process across repeated trials, and apply significance testing (for example, paired t-tests, bootstrap confidence
intervals). It determines which reported performance gains are statistically robust versus attributable to
run-to-run variability. I further examine how important design parameters impact the trade-off between statistical
reliability and resource efficiency, such as bitstream length in stochastic computing (SC) and hypervector
dimensionality in hyperdimensional computing (HDC). By adjusting these parameters systematically, I provide
practical guidance on how to choose parameters that produce stable performance which can be defended and
repeated rather than favorable but unstable results.
Grounding on SC and HDC evaluation in principled statistical methodology, this article supports more reliable
adoption of these techniques in real-world resource-constrained vision systems. Finally, it aims to close a
methodological gap between emerging low-power computing paradigms and the standards of rigor expected in the
broader statistics and machine learning communities.
Quantification of joint and conditional climatic stress dependence on surface soil profile: Bayesian-copula approach on predictive drought index
Kushani De Silva, Lamar University Abstract:
Indices are widely used for monitoring extreme events such as droughts. However, most existing indices rely
only on historical observations where marginal behavior of observations are transformed into index values.
These traditional techniques lack a mathematical model where future values can be predicted before observing
data. Thus lack of predictive capability and proper uncertainty quantification are major drawbacks of the
current index constructing approaches. To address this, we propose a consistent logical approach by training
a suitable mathematical model. Specifically for droughts, we use copula models to represent extreme behaviors.
In particular, the droughts are a consequence of climate pressure observed on surface soil layers. We identify
the distinct nature of two ecosystems and connect them via conditional probabilities using Bayesian methods.
The full inference model is trained for future predictions and accounts for uncertainty quantification. We
demonstrate the superiority of the new method by comparing it to the traditional index of soil moisture by
analyzing climate data in selected geographical regions.
Computational Modeling of Ion Transport Dynamics in Ion Exchange Membranes for Improved Efficiency
Afeez O. Jimoh, University of Louisiana at Lafayette Abstract:
Ion-exchange membranes govern the efficiency of fuel cells and redox flow batteries, yet the heat dissipated by ion transport through them remains poorly understood.
We simulate single-ion trajectories through model ion-exchange membranes, and apply stochastic thermodynamics to quantify the heat emitted along each path.
The long-term goal of this project is to propose a method for designing membranes that minimize heat emission while preserving high ionic conductivity.
AI Tools Indicate Repeated Hurricanes Reshape Coastal Biodiversity
Jacob Badcock, University of Louisiana at Lafayette Abstract:
Gulf of Mexico macrobenthic biodiversity is highly dynamic, with high levels of abundance variability, influenced by environmental variability,
anthropogenic pressures, and extreme climatic events. However, high dynamicity comes with increased risks of irreversible community alterations,
resulting in the occurrence of a regime shift. Regime shifts are persistent changes in biodiversity that can occur either abruptly or slowly over time.
Unfortunately, the definition, detection methodology, and causes of regime shifts remain heavily debated. The identification of factors that induce
regime shifts is critical for understanding ecosystem resilience and guiding conservation efforts on how to prevent or handle these events.
Here, we explore temporal dynamics and drivers of community change in a long-term dataset of macrobenthic communities of Terrebonne
Bay (Gulf of Mexico, Louisiana USA). By applying machine learning techniques, latent Dirichlet allocation (a method for the detection of
distinct recurring biodiversity patterns) and Bayesian changepoint models (a method for identifying large pattern changes within a dataset),
we aim to detect potential regime shifts and assess key environmental drivers within the system over a 5 year span. In total, we quantify
12,341 individuals across 114 species and 25 sampling events. The utilized models have indicated that shifts in biodiversity may correspond
with major hydrological and climatic disturbances (e.g. drought and hurricanes) and a regime shift may have occurred following a major
hydrological disturbance. The application of machine learning techniques to a long-standing marine data set holds great promise in
furthering the understanding of not only the marine environment, but of what factors are needed to induce a new regime.
Characterizing Seasonal Differences in Sperm Whale Activity in the Gulf Using a Generalized Linear Modeling Framework
Phillip Whitman, University of Louisiana at Lafayette Abstract:
We analyzed two years of passive acoustic data to characterize seasonal patterns in sperm whale presence and behavior
in the northern Gulf of Mexico. Using a generalized linear modeling framework, we compared candidate models (including
Poisson, zero-inflated Poisson, and zero-inflated negative binomial) to account for temporally correlated,
over-dispersed, and zero-inflated click-count data, selecting among them using information criteria. Wald tests
indicated that baseline activity levels differed significantly among the seasons. Acoustic activity was highest
in winter and fall and lowest in spring. These results establish a seasonal baseline for future work linking whale
habitat use to environmental and anthropogenic drivers in the Gulf.
Multivariate Functional Inference and Gaussian Mixture Modeling of Joint Temperature-Precipitation Patterns in the U.S. South Central Region
Gihani Vidumini Wickramasinghe Wickramasinghe Disanayakage, Texas Tech University Abstract:
Joint variability in temperature and precipitation is central in characterizing seasonal climate structure and
associated environmental processes, yet many regional analyses rely on marginal or univariate summaries. We
analyze seasonal temperature–precipitation patterns across the Southern United States using two complementary
multivariate statistical approaches. First, functional multivariate analysis of variance (FMANOVA) is employed
to test the equality of state‑level bivariate mean functions, with the permutation‑based Wilks' lambda and Pillai's
trace statistics. Second, Gaussian mixture models are applied to station‑level seasonal summaries to identify
latent climate regimes based on the joint distribution of temperature and precipitation. The FMANOVA results
indicate statistically significant differences in bivariate mean trajectories between states in both winter and
summer, with seasonal contrasts reflecting differing contributions of temperature and precipitation. Clustering
analysis indicates more clearly defined and spatially coherent winter regimes than summer regimes, with summer
regimes exhibiting greater variability and a stronger role for precipitation.
Keywords: Joint temperature and precipitation dynamics; Functional multivariate analysis of variance; Clustering;
Gaussian mixture model; Spatiotemporal climate clusters.
This work is joint with Asim Dey (Texas Tech University) and Kumer Das (University of Louisiana at Lafayette)
Understanding Digital Banking Adoption in Bangladesh: An Empirical Study of Retail Banking Customers
Sujan Kumar, Sam Houston State University Abstract:
This research investigates the factors influencing retail banking customers’ adoption of digital banking services in Bangladesh.
To conceptualize the relationships among factors affecting digital banking adoption, the study employs a theoretical framework
based on an extended Technology Acceptance Model (TAM). A cross-sectional survey using snowball sampling was conducted
among 580 banking service recipients in the Motijheel Commercial region from early February through late March 2026.
Data were collected using a structured questionnaire covering demographic characteristics, perceived usefulness (PU),
perceived ease of use (PEU), security, privacy, and trust (SPT), transaction cost (TC), awareness (AW), and website
features (WF). Binary logistic regression was used to examine the relationships between these factors and the
adoption of digital banking services. Participants ranged in age from 18 to 50 years, with a mean age of 29.58
years (SD = 5.44), and 65.0% were male. Approximately 74.0% of participants preferred using digital banking services.
App-based banking (61.6%) and mobile banking (21.4%) were the most frequently used platforms, and 44.3% of respondents
had used digital banking services for one to three years. The results indicate that perceived usefulness, security, privacy and
trust, and website features have significant positive marginal effects on digital banking adoption. Perceived ease of use shows
a significant negative marginal effect, moderated by participants’ gender. Transaction cost and awareness, however, do not
have a statistically significant effect on digital banking adoption. The findings provide useful insights for financial institutions,
customers, businesses, and researchers interested in understanding and promoting the continued development of digital
banking in Bangladesh.
Confidence intervals for normal quantiles: one- and two-sample problems
Justin Dunnam, University of Louisiana at Lafayette Abstract:
The problems of interval estimating the quantiles of normal distributions in one- and
two-sample cases are considered. For the one-sample case, we show that the classical confidence
interval (CI) based on the noncentral t (NCT) distribution and the
CI (Chakraborti and Li 2007) based on the uniformly minimum variance unbiased
estimator of the population quantile are the same. We also propose simple closed-form alternative CI
for a quantile on the basis of a normal approximation to the NCT
distribution. Furthermore, we develop fiducial distributions for population quantiles
using the NCT distribution as well as the approximate normal distribution. These
fiducial distributions are used to find fiducial and approximate closed-form CIs for
a ratio/difference of normal quantiles. The properties of the CIs are evaluated and
compared using Monte Carlo simulation. Three examples are given to illustrate the
methods of finding confidence intervals for quantiles and ratio/difference of quantiles.
Automating Specific Leaf Area, SLA, Measurements Through Image Processing
Harold Forbis, University of Louisiana at Lafayette Abstract:
Specific leaf area (SLA), defined as leaf area per unit of dry leaf mass, is an important metric in plant biology
and ecology. However, manually measuring leaf area can be tedious and time-consuming, particularly for
inexperienced researchers. We present an image-processing tool that automates area measurements from photographs
of leaf samples containing a 1 cm × 1 cm calibration square. The tool converts each image to HSV color space and
creates binary masks for the leaf and calibration square using OpenCV. It then applies a projective transformation
and morphological methods to improve the masks before estimating leaf area through pixel counting. A logistic
regression model identifies potentially unusable masks based on properties such as contour complexity and contact
with the image border. The resulting measurements are saved in a CSV file for subsequent SLA calculations. This
workflow is intended to reduce the time required for data collection and support the processing of larger sample
sizes. Future development will focus on automated image labeling through optical character recognition, support
for additional calibration formats, more robust image validity checks, and public web access.
Integrating Persistent Homology and Multilevel Functional Data Analysis for Brain MRI
Asim Dey, Texas Tech University Abstract:
Persistent homology provides a multiscale representation of structural features in biomedical images,
but its resulting summaries are often used primarily as predictors or features for classification and
group comparison. We construct a generalized multilevel functional framework for analyzing persistent-homology
summaries as longitudinal functional outcomes in repeated three-dimensional structural magnetic resonance imaging
(MRI). Specifically, we represent topological features using Betti curves and model these curves as count-valued
functional responses indexed by the filtration parameter. A negative-binomial distribution accommodates the
discrete and potentially overdispersed nature of Betti counts, while the multilevel formulation accounts for
the dependence induced by repeated measurements and separates between-subject and within-subject sources of
functional variation. Functional principal component analysis further characterizes the dominant modes of
variation at each level. We employ Bayesian estimation to jointly estimate the functional regression and
multilevel functional principal component models. We apply the proposed framework to longitudinal structural
MRI data from the OASIS-2 study to investigate associations between brain topology and demographic and
clinical characteristics, including age, gender, follow-up time, and dementia severity. The results demonstrate
that the framework can capture covariate-associated variation across the filtration continuum while
characterizing distinct sources and patterns of longitudinal variation across homology dimensions.
This work is joint with Shashipraba Rajakaruna ( Texas Tech University ) and A. Alexandre Trindade ( Texas Tech University )
Machine Learning in Stata via H2O
Zhao Xu, Stata Corp., Inc. Abstract:
Stata 19 introduces the h2oml suite of commands, which brings machine learning to Stata
through H2O, a scalable, open-source platform. Users can fit gradient boosting machines and random
forests for regression and classification, tune hyperparameters, evaluate model performance,
and make predictions on unseen data. The suite also provides explainability tools—variable importance,
partial dependence plots, ICE curves, and Shapley values—so predictive power comes
without sacrificing interpretability. This talk introduces the workflow and shows how users can
apply these methods from within Stata.
Resource-Adaptive Machine Learning for Micronutrient Screening: Temporal Evaluation Across Repeated NHANES Cycles
Edidiong Akpan, University of Louisiana at Lafayette Abstract:
Micronutrient deficiencies remain an important public health concern and may affect health before
becoming clinically obvious. Confirming deficiencies such as iron and vitamin D deficiency requires nutrient-specific
laboratory testing, creating a need for efficient ways to identify people who may benefit most from confirmatory
testing. Machine-learning methods have shown promise for micronutrient risk prediction, but important questions
remain: How much clinical information is needed for useful screening? Do these models remain reliable across
changing populations and over time? Can they help target confirmatory testing more efficiently? We developed a
resource-adaptive machine-learning approach to address these questions.
We assembled adults aged 18–59 years from repeated U.S. National Health and Nutrition Examination Survey
(NHANES) cycles spanning 2009–2023 and studied vitamin D insufficiency and pre-anemic iron deficiency. We defined
two information tiers designed to prevent leakage. Tier 1 used non-invasive demographic, socioeconomic,
healthcare-access, anthropometric, lifestyle, and dietary information; Tier 2 added routine complete blood
count variables. We trained six model families and evaluated them under three temporal designs: fixed-origin,
rolling-expanding, and backward/forward transport. We also assessed calibration, distribution shift, and ferritin
eligibility-matched performance. Decision-curve net benefit, confirmatory-testing burden, survey-weighted and
age-restricted sensitivity analyses, and subgroup transportability were examined as secondary checks.
Vitamin-D discrimination was nearly identical across tiers (mean AUROC 0.736 vs 0.735), showing negligible
added value from CBC information. Iron showed a different pattern: AUROC increased from 0.548 to 0.840 in the
observed cohort and from 0.534 to 0.826 in the eligibility-matched cohort. All six model families showed the
same Tier-2 advantage for iron. At the 90\% training-sensitivity operating target, Tier-2 iron models avoided
approximately 27–30\% of confirmatory tests. Eligibility matching substantially reduced the distributional shift
associated with changes in ferritin eligibility but did not consistently restore discrimination. Recalibration
improved probability reliability. Survey-weighted, age-restricted, and subgroup analyses did not change the main
tier conclusions.
The information needed for useful micronutrient triage is nutrient-specific. Routine hematologic
information substantially improved iron-deficiency screening but added little to vitamin D screening.
Resource-adaptive screening may help target confirmatory testing, but decisions about deployment should consider
temporal transportability, population and eligibility shifts, calibration, and subgroup performance, not just
overall discrimination.
This work is joint with Joy Ojuro ( University of Louisiana at Lafayette )
A Machine Learning Pre-Audit Screening Tool for Industrial Energy Efficiency Assessment
Tamjeed Ahmed, University of Louisiana at Lafayette Abstract:
Small and medium-sized manufacturers (SMMs) account for a disproportionate share of industrial energy waste, yet
most never receive professional energy audits due to financial constraints, limited awareness of savings potential,
and difficulty justifying audit commitments without preliminary estimates. This study develops a two-stage machine
learning framework that provides pre-audit screening using only basic facility characteristics available before an
on-site visit. The framework trains on 225,123 energy conservation recommendations from 22,636 facilities across the
U.S. Department of Energy Industrial Assessment Center (IAC) database spanning 1980–2026. Stage one employs 24 binary
LightGBM classifiers to predict applicable Assessment Recommendation Code (ARC) categories for a facility, achieving
a mean Average Precision of 0.341 with Motor Systems reaching 0.954. Stage two estimates energy and dollar savings
per recommendation using LightGBM regressors with 24 engineered features including intensity ratios, sector z-scores,
and ARC target encoding, attaining R2 = 0.550 for energy savings and R2 = 0.485 for dollar savings with five-fold
cross-validation stability (σ < 0.005). A composite scoring formula ranks recommendations by probability, historical
implementation rate, and predicted savings magnitude. Validation on 50 held-out Louisiana facilities using only
sector-median inputs as a conservative scenario yields Recall@3 = 82% and Recall@6 = 92%, demonstrating that six
ranked recommendations capture nearly all categories that auditors would identify on-site. Applied to Louisiana’s
580 IAC-assessed facilities, the tool identifies 188,920 metric tons of CO₂ reduction potential annually at a 48.4%
implementation rate. The framework addresses a critical gap between high-level benchmarking tools and full energy
audits, enabling facilities to understand their savings potential in under ten minutes with no specialized data
collection.
This work is joint with Md Mainuddin Khaled ( University of Louisiana at Lafayette ),
Tamjid Shafiq ( University of Louisiana at Lafayette ),
Md Rakibul Islam ( University of Louisiana at Lafayette ),
and Peng Yin ( University of Louisiana at Lafayette )
Testing multivariate normality of a large number of populations
Nurudeen Ajadi, University of Louisiana at Lafayette Abstract:
We provide a new approach to test multivariate
normality for K independent samples simultaneously
when the parameters are unknown and estimated from
sample. The number of samples, K, is allowed to be
arbitrarily large and independent of sample size. The
proposed test is based on the energy test for multivariate
normality in (Szekely 2005). Asymptotic normality
of the test statistics has been established to perform
the test as K tends to infinity. The proposed test is shown to
be consistent against fixed nonnormal alternatives under
mild regularity conditions. Simulations studies are
conducted to examine the performance of the new K sample
test.
