Publications

Summary. This study asks whether fake news is equally detectable when it is written by humans versus generated by AI. Using the same feature representation and modeling pipeline for both tasks, the paper compares real news with human-written fake news and with AI-generated fake news using lexical-diversity, readability, and emotion features together with logistic regression, random forests, support vector machines, gradient boosting, neural networks, and an ensemble model.

Across all models, AI-generated fake news was substantially easier to distinguish from real news: most models achieved AUCs above 0.99 for the AI-generated task, compared with roughly 0.89–0.96 for human-written fake news. Feature analyses suggest that AI-generated misinformation currently shows more uniform patterns in readability and emotional expression, whereas human-written fake news more closely resembles legitimate journalism.

Summary. This paper develops a multivariate, distribution-free rank-based framework for comparing clustered and correlated outcomes across groups. The method adapts the Longitudinal Rank-Sum Test (LRST) to a clustered cross-sectional setting, treating states as independent units and counties as dependent observations within states, while combining several socioeconomic outcomes into a single omnibus comparison.

The framework is applied to pooled U.S. county data from 2019–2023 to compare states with and without refundable Earned Income Tax Credit (EITC) policies across poverty, income inequality, housing cost burden, medical care costs, and per-capita income. Across repeated county subsampling and different cluster sizes, counties in EITC states consistently showed more favorable joint socioeconomic rankings. The analysis is explicitly descriptive rather than causal and illustrates how rank-based information fusion can provide robust, interpretable comparisons when outcomes are multivariate and clustered.

Summary. This paper develops an interpretable machine-learning framework for distinguishing organic users from social-media bots using information extracted from users’ X/Twitter histories. Rather than relying only on profile metadata or opaque deep-learning representations, the method combines temporal behavior—such as periodicity, ARIMA-based dynamics, and periodogram structure—with semantic features including lexical diversity, hashtag use, word-frequency patterns, and sentiment.

Using verified organic and inorganic accounts, the study evaluates k-means clustering and support vector machines and finds that combining temporal and semantic features yields the strongest discrimination. The framework also adds split conformal prediction to quantify uncertainty, producing calibrated prediction sets alongside class labels; the conformal analysis reported approximately 94.7% test accuracy with an AUC of 0.978.

Summary. This paper develops a general inferential theory for polyspectral means—weighted averages of higher-order spectra designed to capture nonlinear and non-Gaussian dependence that ordinary second-order spectral methods can miss. It establishes asymptotic normality for polyspectral-mean estimators of arbitrary order and derives an explicit limiting variance involving higher-order polyspectra, providing a unified basis for large-sample inference.

The paper also proposes a new bispectrum-based test of the linear-process hypothesis. Simulations study finite-sample behavior, while applications illustrate the methodology through linearity testing for sunspot data and clustering GDP growth trajectories across countries using bispectral features.

Summary. This review organizes the major methodological families used for efficacy analysis in clinical trials, spanning classical parametric procedures, nonparametric methods, Bayesian approaches, and modern machine-learning techniques. It emphasizes when each framework is most appropriate, the assumptions each requires, and how the choice of method changes with longitudinal, skewed, high-dimensional, or otherwise complex trial data.

The paper also discusses missing-data mechanisms, adaptive and personalized trial settings, deep learning, reinforcement learning, and emerging computational approaches. Its central goal is to connect traditional statistical inference with newer predictive methodology while making the tradeoffs in robustness, interpretability, uncertainty quantification, and scalability explicit.

Summary. This paper introduces DISTINCT (Distributional Subsampling for Covariate-Targeted Alignment), a framework for selecting a subsample from a large real-world cohort whose joint demographic distribution matches that of a virtual phantom population. The method aligns continuous and categorical covariates together and evaluates distributional similarity using Wasserstein and Kolmogorov–Smirnov distances.

Applied to the National Lung Screening Trial and the Virtual Lung Screening Trial, DISTINCT identified an aligned NLST cohort of 9,974 participants. Downstream lung-cancer risk-score performance stabilized once aligned samples exceeded roughly 6,000 participants, illustrating how demographic alignment can support more representative comparisons between real and virtual imaging studies.

Summary. This paper generalizes classical spectral factorization from second-order spectra to higher-order polyspectral densities, developing a mathematical framework for nonlinear and non-Gaussian time series. It establishes new mappings between higher-dimensional cepstral coefficients and autocumulants while respecting the symmetry structure of polyspectra.

The resulting representation provides tools for nonlinear filtering and forecasting and offers a flexible semiparametric route to modeling stationary nonlinear processes without imposing a specific nonlinear equation in advance.

Summary. This paper introduces Causal-GCN, a graph convolutional framework designed to move beyond purely associational Alzheimer’s disease prediction. Each subject is represented as a brain graph, while age, sex, and APOE4 information are incorporated into a back-door adjustment strategy so that ROI-level interventions can be evaluated within the trained network.

Applied to 484 ADNI participants, the framework achieves classification performance comparable to standard graph-neural-network baselines while producing interpretable intervention-based rankings of brain regions. The analysis highlights posterior, cingulate, and insular regions as influential hubs and uses bootstrap resampling to assess the stability of the estimated effects.

Summary. This paper uses extreme-value methods to study severe Air Quality Index episodes rather than focusing only on average pollution levels. Daily maximum AQI values are analyzed for five major Indian cities and eight metropolitan areas elsewhere in the world, with separate attention to seasonal behavior and the upper tail of the pollution distribution.

The analysis documents pronounced city-to-city and seasonal differences in extreme pollution risk, with especially severe wintertime behavior in cities such as Delhi and Dhaka. The study illustrates how extreme-value analysis and bootstrap-based uncertainty assessment can help quantify rare but high-impact environmental events.

Summary. This paper develops a time-series clustering approach that uses spectral and bispectral means as features, allowing the clustering procedure to incorporate higher-order dependence and nonlinear structure that ordinary second-order spectral summaries can miss.

The method is applied to 49 NIFTY 50 stocks using a 1,000-day window of value-weighted average prices. K-means clustering of the higher-order spectral features identifies five distinct groups, revealing market structure that cuts across conventional sector labels and illustrating the use of polyspectral features for nonlinear financial time-series analysis.

Summary. This paper develops two statistical forecasting frameworks that combine traditional opinion polls with social-media information. THOS uses Twitter hashtag frequencies, while THANOS augments those signals with network-structure features such as centrality measures, with the goal of using fast-moving online information alongside slower but more representative polling data.

The models are evaluated on Ireland’s 2018 referendum on the 36th constitutional amendment and the 2018 U.S. Senate elections. In those studied cases, the paper reports that the simpler THOS model worked well when margins were relatively clear, whereas network information became more useful in closely contested races. The results are presented as an empirical forecasting study rather than as a general guarantee about future elections.

Summary. This paper extends the Longitudinal Rank-Sum Test from two-arm trials to multi-arm randomized clinical trials with multiple longitudinal outcomes. The resulting U-statistic framework provides a global, distribution-free comparison across treatment arms while retaining the temporal and multivariate structure of the data.

Simulation studies show strong Type I error control and improved power relative to applying separate two-arm LRSTs with multiplicity correction. An application to the Bapineuzumab 301 Alzheimer’s disease trial demonstrates how the method can be used to evaluate several dose groups against a common control and then identify promising treatment arms after a global rejection.

Summary. This study develops an ensemble survival framework for predicting progression from cognitively normal status to mild cognitive impairment or Alzheimer’s disease using longitudinal ADNI biomarkers. The approach combines penalized Cox models with Random Survival Forest, DeepSurv, and XGBoost, and then aggregates predictions using ensemble averaging and Bayesian model averaging.

Among 721 initially cognitively normal participants, the ensemble achieved a peak C-index of 0.907 and integrated time-dependent AUC of 0.904, compared with a C-index of 0.608 for baseline-only models. A notable practical finding is that adding one follow-up visit produced most of the predictive gain, while a second follow-up added relatively little.

Summary. This paper develops the power and sample-size theory needed to use the Longitudinal Rank-Sum Test prospectively when designing multivariate longitudinal clinical trials. It derives the LRST distribution under both null and alternative hypotheses, obtains closed-form expressions for power and sample size, and develops consistent estimators with large-sample justification.

Simulation studies validate the calculations, and applications to Alzheimer’s and Parkinson’s disease trials illustrate how the framework can be used to plan adequately powered studies when several outcomes are followed repeatedly over time.

Summary. This paper develops Penalized Fast Causal Inference (PFCI) for high-dimensional causal discovery when the underlying biomarker network is sparse and latent confounding may be present. The method first uses penalized neighborhood selection to reduce the dependency graph and then applies FCI to the smaller structure, preserving the ability to represent hidden confounders while making causal discovery more scalable.

The approach is applied to PPMI proteomic data containing 2,924 plasma and CSF biomarkers. The analysis identifies a compact set of candidate biomarkers connected to Parkinson’s disease, including the plasma biomarker Myocilin together with several CSF markers, illustrating how sparsity-aware causal discovery can prioritize potentially informative biomarker pathways in very high-dimensional biological data.

Summary. This paper introduces the Longitudinal Rank-Sum Test (LRST), a nonparametric U-statistic-based global test for clinical trials with multiple primary endpoints measured repeatedly over time. The method aggregates evidence across outcomes and visits into a single omnibus test, avoiding separate multiplicity adjustment while accommodating non-normal and heterogeneous outcome distributions.

Simulations examine a range of realistic trial settings, including conflicting effects across endpoints, and demonstrate Type I error control together with competitive power. Applications to neurodegenerative-disease trial data illustrate how LRST can summarize treatment efficacy across cognitive, functional, and other longitudinal domains.

Summary. This paper develops a general nonlinear forecasting framework based on quadratic predictors for stationary time series. The predictor extends the classical linear Yule–Walker equations by incorporating second-, third-, and fourth-order auto-cumulants, allowing nonlinear and non-Gaussian dependence to contribute directly to prediction.

The theory characterizes when a process is second-order forecastable—that is, when quadratic prediction can improve on the best linear forecast—and connects that condition to spectral and bispectral structure. Numerical studies show substantial prediction gains in nonlinear settings, while applications to U.S. unemployment and Wolfer sunspot data reduce mean-squared prediction error by 16.5% and 29.2%, respectively.

Summary. This study evaluates photon-counting CT for quantitative lung-density measurement and compares it with conventional energy-integrating CT. It combines retrospective clinical data with scanner-specific virtual imaging experiments, allowing reconstruction parameters and radiation dose to be varied systematically against known virtual ground truth.

The analysis identifies imaging settings associated with lower lung-density error and reports that, under the optimized settings studied, photon-counting CT achieved lower lung and emphysema density error than the compared energy-integrating CT scanners. The work illustrates how virtual imaging trials can complement clinical data when controlled repeated imaging experiments would be impractical in patients.