Monday, Aug 3: 8:30 AM - 10:20 AM
6230
Contributed Papers
Thomas M. Menino Convention & Exhibition Center
Room: CC-256
Main Sponsor
Section on Nonparametric Statistics
Presentations
The coefficient of variation (CV) is a widely used dimensionless measure of relative variability and plays an important role in quantifying measurement error. Its inverse, the inverse coefficient of variation (ICV), though less commonly discussed, provides a meaningful alternative, particularly when the population mean is near zero. Despite its practical relevance, statistical inference for the ICV remains underdeveloped, especially in correlated and non-normal settings.
This study addresses inference for comparing two inverse coefficients of variation when the estimates are correlated, a situation that commonly arises in paired designs, such as measurements obtained from the same subjects using different devices or outcomes. We propose nonparametric methods and an adjusted asymptotic approach for constructing confidence intervals for two correlated ICVs.
Extensive simulation studies were conducted under normal and non-normal distributional settings using Google Colab (v6e-1 TPU). In addition, biometric and lifestyle data collected by the Institute of Clinical Bioethics at Saint Joseph's University from underrepresented communities in Philadelphia were used as an application.
Keywords
Inverse coefficient of variation
Confidence intervals
Skewed distributions
Nonparametric methods
Adjusted asymptotic approach
Likelihood-based methods
We propose a kernel-weighted deep neural network (K-DNN) for predicting an individual's future conditional mean response trajectory from a history of predictors in sparse and irregularly measured longitudinal data. To accommodate subject-specific and irregular observation times, we first reconstruct each subject's predictor trajectories nonparametrically, yielding predictor values on a common grid over the historical time window. The proposed architecture has two stages. In Stage 1, early layers form time-specific local subnetworks that learn representations of predictor information at each grid time point. In Stage 2, representations from multiple grid times are integrated in subsequent layers to capture complex temporal dependence and interactions across the history. To further mitigate the impact of data sparsity, we incorporate a kernel-weighting mechanism that prioritizes observations measured in close proximity to the target future time point. This hybrid approach combines the flexibility of deep learning with the local smoothing advantages of kernel methods, providing a robust predictive tool for complex longitudinal profiles.
Keywords
Deep neural networks
kernel weighting
Longitudinal data analysis
The Gaussian mixture model has long been a staple of nonparametric density estimation, but while these models are highly flexible, inference often poses a number of non-trivial challenges. The possibilistic inferential model (IM) overcomes these challenges by offering inference with exact frequentist coverage in all sample sizes while enjoying probability-like interpretation in the style of Bayesian methodologies. Recent computational advances allow us to efficiently evaluate this IM's output. We compare existing methods with a new, more flexible approach and highlight its superior performance in challenging models like the Gaussian mixture. This paper goes on to demonstrate that this exact coverage can be leveraged into desirable behavior in ``downstream'' tasks, such as prediction and model selection, through the use of Choquet integrals. A real-data application is offered in support of the methods using the Galaxy data, a common benchmark for work in Gaussian mixture models.
Keywords
Model Selection
Valid
Exact Coverage
Finite Sample
Density Estimation
The rapid growth of data has introduced considerable computational challenges in statistical analysis. This study addresses this issue in local linear regression through representative subdata selection to reduce the computational burden, then extends the method to partial linear models. For local linear regression, a residual-based subdata selection (RESS) method is introduced. RESS yields a lower asymptotic mean squared error than existing methods in a neighborhood where the absolute asymptotic bias is largest. For partial linear models, an integrated estimation approach, termed IBRESS, combines RESS for the nonlinear component with information-based optimal subdata selection (IBOSS) for the linear component. IBRESS leverages the strengths of both methods and satisfies two theoretical properties: (i) similar to IBOSS, the convergence rate of the linear component depends on the full data size; and (ii) the nonlinear component retains the asymptotic properties of RESS. Simulation studies demonstrate that IBRESS reduces computational cost while maintaining estimation accuracy.
Keywords
Local linear regression
Subdata selection
Partial linear model
Nonparametric regression
L1 convergence
Speaker
Chia-Wei Lin, Institute of Statistics and Data Science, National Tsing Hua University, Taiwan
Co-Author
Li-Shan Huang, National Tsing Hua University, Taiwan
We propose a flexible class of semiparametric regression models for count data based on zero-altered distributions (ZADs), estimated via the generalized method of moments (GMM). Unlike conventional parametric GLMs (e.g., Poisson or negative binomial), our approach accommodates both under- and over-dispersion while retaining full support on the nonnegative integers. By relying only on moment conditions rather than full likelihoods, the GMM framework offers robust inference even under distributional misspecification, and ensures consistent estimation when higher-order features are misspecified. Theoretical results establish consistency and asymptotic normality of the estimators under broad conditions. The robustness and flexibility of the proposed models are extensively demonstrated using three real-world data sets-covering diverse dispersion patterns and practical domains, which highlight superior performance relative to standard GLMs. Accompanying R functions facilitate straightforward implementation for applied researchers.
Keywords
Generalized method of moments
over-dispersion
under-dispersion
misspecification
excess zeroes
semiparametric methods
In many applications, the target parameter depends on a nuisance function defined by a conditional moment restriction, whose estimation often leads to an ill-posed inverse problem. Classical approaches, such as sieve-based GMM, approximate the restriction using a fixed set of test functions and may fail to capture important aspects of the solution. Adversarial estimators address this limitation by framing estimation as a game between an estimator and an adaptive critic. We study the class of Regularized Adversarial Stabilized (RAS) estimators that employ reproducing kernel Hilbert spaces (RKHSs) for both estimation and testing, with regularization via the RKHS norm. Our first contribution is a novel analysis that establishes finite-sample bounds for both the weak error and the root mean squared error (RMSE) of these estimators under interpretable source conditions, in contrast to existing results. Our second contribution is a detailed comparison of the assumptions underlying this RKHS-norm-regularized approach with those required for (i) RAS estimators using L2 penalties, and (ii) recently proposed, computationally stable Kernel Maximal Moment estimators.
Keywords
Finite Error Bounds
Integral Equations
Kernel Adversarial Estimators of Moment Equations
Mini-max Learning
Reproducing Kernel Hilbert Spaces
Source Condition
Rural STEM teacher shortages motivate survival models capturing early-career attrition and student achievement impacts. We adapt the starshaped mean equilibrium life (SMEL) framework-defined by nondecreasing mean residual life ratio m(t)/t-to model teacher careers as burn-in-to-equilibrium processes. We embed SMEL in proportional mean residual life (PMRL) regression and estimate a three-parameter Weibull baseline via Bayesian inference using the No-U-Turn Sampler. Monte Carlo studies under decreasing, increasing, and bathtub hazards with up to 40% censoring show SMEL-PMRL maintains minimal bias and reduces integrated Brier scores by ~20% relative to Cox, accelerated failure time, and standard Weibull models. Applied to 2018-2023 NSF Noyce data on 169 rural Texas STEM teachers and 3,214 students, the model reveals a 32% rural penalty in expected tenure, a 38\% drop in remaining tenure during years 1-3, and positive effects of Noyce scholarships and rural-origin backgrounds. Joint teacher--student modeling links persistence beyond year 3 to substantial achievement gains, illustrating how SMEL-PMRL yields interpretable, policy-relevant metrics for timing retention interventions.
Keywords
mean residual life
survival analysis
Bayesian methods
teacher retention
rural STEM education
PMRL regression