Montana ASA Chapter Meeting, 2026

This presentation introduces the current Vice Chair of ASA District 6 (2026-2028 term), and provides an overview of the district’s structure, chapters, and current ASA initiatives. It highlights opportunities such as DataFest, traveling courses, poster competitions, and the visitation program. The presentation also outlines proposed District 6 initiatives designed to connect students and early-career professionals with statisticians across a range of industries. These include a career-focused symposium, a statistics-themed game show, a student mentorship program, a research paper competition, and a research opportunity for high-achieving high school students. The presentation concludes by recognizing the Montana Chapter’s 50-year legacy.

Network data often consists of underlying continuous time interactions between the nodes in the network. Representing the true underlying dynamics of such interactions is challenging, and often instead utilises an aggregate summary of these events. We propose a continuous time functional latent space model for the nodes in a network, based directly on these underlying interactions, using modern computation methods to scale inference to large scale networks commonly seen in practice. We validate our proposed model and interpret the results for real interactions on networks, identifying interpretable latent structure.

As the hype of GenAI becomes pervasive, I would like to take a step back and discuss some statistical concepts that need to be thought about. Where does stochasticity, statistical parity, unbiasedness, fairness (which are hallmarks of statistical thinking) fit in GenAI. The bigger question I am working on is how to fit in variability along with accuracy in algorithms. In this short talk I will discuss a couple of recent case studies and discuss the very simple statistical reasons why sometimes these GenAI models fail catastrophically. I will end with the accuracy, fairness conundrum.

There are many opportunities for people with a statistics background to pursue careers in government and health policy. Drawing on the speaker's experience working in a local government setting, for a federal government contractor, and in an academic health services research setting, this talk gives an overview of how statisticians can successfully collaborate with interdisciplinary teams to contribute to improvements in health programs and policy. This talk will also discuss factors to consider when deciding whether to pursue a PhD for students interested in health policy and public sector careers.

Statistics rarely arrive as neatly defined textbook problems. This presentation uses three real-world examples from the Montana Legislative Audit Division (gravel pits, affordable housing, and the state Motor Pool) to show how quantitative methods can help answer practical questions facing state government. Through examples involving causal inference, spatial analysis, and optimization, the presentation explores how statisticians can translate messy policy questions into defensible evidence, turning analysis into something policymakers can actually use.

State-space models are widely used for analyzing multivariate time series with missing observations. In many longitudinal studies, the pattern of missingness is nonmonotone, and the probability of missingness depends directly on unobserved signal values, leading to nonmonotone Missing Not At Random (MNAR) mechanisms that induce systematic bias under the widely popular MAR-based state-space inferences developed in the literature so far. We propose linear-Gaussian state-space models with explicit MNAR modeling for such data, where the missingness process is governed by a logistic regression depending on latent states, current and lagged emissions, and historical missingness. By leveraging Polya–Gamma augmentation, the MNAR mechanism is rendered conditionally conjugate. This enables closed-form posterior updates via Gibbs sampling and variational inference via Kalman filtering and Rauch–Tung–Striebel smoothing within a coordinate-ascent algorithm. Through simulation studies and an application on oceanographic dataset, we demonstrate the advantages of our method over existing state-of-the-art approaches.

Recurrent neural networks have shown promising performance in the modelling of real-world spatiotemporal data. Among the many deep learning approaches to this problem, SHallow REcurrent Decoder (SHRED) networks use spatially sparse sensor data to reconstruct full system states across time, motivated by Takens’ embedding theorem. Prior SHRED networks have used long-short-term-memory (LSTM) and gated recurrent units (GRU), which are best suited to complete, synchronously observed samples on a regular time grid. Environmental monitoring data, however, are frequently incomplete and irregularly sampled. In this work, we explore the use of continuous recurrent units (CRU) in a SHRED network. CRUs use Kalman-style predict and update steps to assimilate observations, allowing the model to account explicitly for variable time intervals and missing measurements. We evaluate the resulting CRU-SHRED architecture on global sea surface temperature data and compare its reconstruction with conventional LSTM- and GRU-based SHRED models under regular, irregular, and missing-data scenarios.

The validity of statistical inferences drawn from observational field studies, experimental designs, or probability sampling depends upon the decisions made prior, during, and after data collection. To support defensible scientific conclusions, analysts must distinguish between what the data can reliably tell us from what we hope they might reveal. We present a generalizable framework for evaluating datasets, which emphasizes four foundational design components: Spatial, Temporal, Response, and Inference Design (aka “STRIDe”). By applying an iterative, five step workflow grounded in clearly articulated objectives, STRIDe helps users diagnose misalignment between desired statistical inferences and the design architecture underlying any dataset. The framework uses structured critical thinking—supported by tools such as Ishikawa (fishbone) diagrams and targeted empirical or simulation based follow up explorations—to identify factors that may undermine defensible inference. Although applicable to any ecological dataset, we illustrate the approach in the context of long term monitoring, where evolving objectives, data integration challenges, and shifting analytical methods frequently introduce design–inference disconnects. Our workflow anchored by an easy to recall acronym, STRIDe, provides a transferable foundation for exploring the alignment between data structures and statistical inference goals—closing the gap between evidence and aspiration.

Health and Human Services organizations generate vast amounts of data – from patient records and service utilization to program performance and community-level outcomes. Yet, more data does not necessarily lead to better decision-making. The real challenge is transforming complex information into insights that are understandable, timely, and actionable. This talk demonstrates how the Montana Department of Public Health and Human Services is integrating data into its decision-making process. Using a practical example, the presentation will walk through strategies for simplifying a complex dataset while preserving the context needed to understand it. The result is an analysis that turns data into meaningful insights and supports informed, impactful program decisions.

This research addresses a critical limitation in the standard Permutational Multivariate Analysis of Variance (PERMANOVA), specifically its sensitivity to unequal group variances (heteroscedasticity). In high-dimensional distance- based situations, test statistics such as the standard pseudo-F (F1) can produce misleading results, with inflated Type I error rates. This is a challenge known as the Generalized Multivariate Behrens-Fisher problem. This study investigates the effectiveness of the recently proposed statistic, F2, as a solution to this problem in distance-based hypothesis testing. We utilized a simulation-based approach to compare the performance of the statistics F1 and F2. Synthetic high-dimensional datasets were generated with varying degrees of heteroscedasticity. Our findings demonstrate that F1 is consistently misled by heteroscedasticity, yielding smaller p-values (p < 0.05) even when group locations are identical. Conversely, F2 maintained nominal Type I error rates reasonably well by correctly partitioning location differences from dispersion effects. We developed the F2Adonis2 R package, providing a user-friendly tool for researchers to evaluate shifts in group centers in the presence of unequal noise levels. To demonstrate the use of F2 and our R package, the results are applied to real-world biological datasets.

Existing global tests for functional ANOVA perform well under homoskedasticity but many can be problematic when variability differs across groups. This issue arises in our motivating functional dataset, where heat flow curves from a biomedical experiment show differences in variance across treatment groups as well as possible differences in the means of the groups. The presence of heteroskedasticity can lead to inflated Type I error rates and reduced power to detect differences in group mean functions. To address this, simulation studies are conducted to assess the performance of existing global tests for functional ANOVA under conditions such as non-constant variance, unbalanced designs, and small sample sizes. We explore a combination of the best-performing global tests with a two-level within-time testing framework to identify specific times or regions where group mean curves differ and to conduct pairwise comparisons within those identified regions in the presence of heteroskedasticity.

Protecting freshwater habitats is vital for supporting diverse fish, wildlife, and surrounding human communities. To accurately monitor these ecosystems, Spatial Stream Network (SSN) models were developed to account for the complex topology of river networks. Scalable Spatial Stream Network (S3N) models improve computational efficiency by utilizing a Nearest Neighbor Gaussian Process (NNGP) to optimize storage requirements for estimation and prediction. S3N and NNGP performance depend on observation point order, and Guinness (2018) demonstrated the choice of permutation can enhance model fit on rectangular planes. This study extends these findings to river networks by evaluating four distinct point-ordering permutations to determine their influence on S3N model goodness of fit and computational efficiency. Using simulated data, we investigate breadth-first or depth-first searches, originating from the most upstream or most downstream observation point. As this research is ongoing, preliminary findings on the impact of systematic versus random ordering will be presented.

Vector-borne and vaccine-preventable diseases such as malaria, dengue, chikungunya, and measles remain major public health challenges and a leading cause of morbidity and mortality in Africa, accounting for a significant proportion of the region’s disease burden. Factors including fragile health systems, rapid urbanization, and climate variability create an environment conducive to disease outbreaks and hinder socio-economic development. Despite the interconnected nature of these diseases, cross-disease vulnerability modeling using integrated environmental and social determinants remains limited in the African context.

In many domains, practitioners seek models that produce accurate forecasts while faithfully capturing latent system dynamics. Existing approaches typically sacrifice one of these goals: deep state space models often assume Gaussian latent transitions, limiting fit and forecasting, while diffusion models are highly expressive but lack principled inference for the underlying dynamics. To combine the strengths of both, we introduce the Diffusion-Driven State Space Model (DDSSM), which replaces the conventional Gaussian transition distribution with a diffusion model. Our DDSSM resolves the open problem of how to jointly train an autoencoder and a diffusion model on sequential data, thereby extending the literature on latent diffusion models for time series. Moreover, we find that the DDSSM empirically outperforms a state-of-the-art deep SSM at fitting and forecasting a simulated time series with multimodal transitions.