50107 research outputs found
Sort by
Idiosyncratic spatial scaling of biodiversity–disease relationships
No full textHigh host biodiversity is hypothesized to dilute the risk of vector-borne diseases if many host species are ‘dead ends’ that cannot effectively transmit the disease and low-diversity areas tend to be dominated by competent host species. However, many studies on biodiversity–disease relationships characterize host biodiversity at single, local spatial scales, which complicates efforts to forecast disease risk if associations between host biodiversity and disease change with spatial scale. Here, our objective is to evaluate the spatial scaling of relationships between host biodiversity and Borrelia (the bacterial taxon which causes Lyme disease) infection prevalence in small mammals. We compared the associations between infection prevalence and small mammal host diversity for local communities (individual plots) and metacommunities (multiple plots aggregated within a landscape) sampled by the National Ecological Observatory Network (NEON), an emerging continental-scale environmental monitoring program with a hierarchical sampling design. We applied a multispecies, spatially-stratified capture–recapture model to a trapping dataset to estimate five small mammal biodiversity metrics, which we used to predict infection status for a subset of trapped individuals. We found that relationships between Borrelia infection prevalence and biodiversity did indeed vary when biodiversity was quantified at different spatial scales but that these scaling behaviors were idiosyncratic among the five-biodiversity metrics. For example, species richness of local communities showed a negative (dilution) effect on infection prevalence, while species richness of the small mammal metacommunity showed appositive (amplification) effect on infection prevalence. Our modeling approach can inform future analyses as data from similar monitoring programs accumulate and become increasingly available through time. Our results indicate that a focus on single spatial scales when assessing the influence of biodiversity on disease risk provides an incomplete picture of the complexity of disease dynamics in ecosystems
The Jwst Excels Survey: Tracing The Chemical Enrichment Pathways Of High-redshift Star-forming Galaxies With O, Ar, And Ne Abundances
No full textWe present an analysis of eight star-forming galaxies with from the JWST Early extragalactic Continuum and Emission Line Survey for which we obtain robust chemical abundance estimates for the -elements O, Ne, and Ar. The -elements are primarily produced via core-collapse supernovae (CCSNe) which should result in-element abundance ratios that do not vary significantly across cosmic time. However, Type Ia supernovae (SNe Ia) models predict an excess production of Ar relative to O and Ne. The abundance ratio can therefore be used as a tracer of the relative enrichment of CCSNe and SNe Ia in galaxies. Our sample significantly increases the number of sources with measurements of at, and we find that our sample exhibits subsolar Ar/O ratios on average, with. In contrast, the average Ne/O abundance is fully consistent with the solar ratio, with . Our results support a scenario in which Ar has not had time to build up in the interstellar medium of young high-redshift galaxies, which are dominated by CCSNe enrichment. We show that these abundance estimates are in good agreement with recent Milky Way chemical evolution models, and with Ar/O trends observed for planetary nebulae in the Andromeda galaxy. These results highlight the potential for using multiple element abundance ratios to constrain the chemical enrichment pathways of early galaxies with JWST
Measurements Of Stratospheric Clo From Mauna Kea: 1992–2023
No full textWe present a reanalysis of the ground-based microwave measurements of upper stratospheric ClO from Mauna Kea over 1992–2023 made by the Chlorine Oxide Experiment (ChlOE) instrument. In order to reduce instrumental baseline artifacts, the retrieval makes use of the difference of daytime and nighttime spectra (the nighttime ClO spectra are much smaller) to produce a day-minus-night ClO mixing ratio, which is the fundamental quantity analyzed throughout this study. Upper stratospheric ClO values peaked in ∼1997, and the trend in the upper stratospheric ChlOE measurements from 1997 to 2023 was found to be −0.4 ± 0.3%/yr (2σ). Comparisons of ChlOE measurements with a combined coincident Upper Atmosphere Research Satellite (UARS) and Aura Microwave Limb Sounder (MLS) data set showed a relative trend of +0.3% ± 0.3%/yr (2σ) over 1992–2023, where a positive trend indicates that the ChlOE ClO measurements are increasing relative to those from MLS. Restricting the comparisons to ChlOE and Aura MLS over 2004–2023, resulted in relative trends that varied by pressure level, from +0.15%/yr to +0.42%/yr, all agreeing to within the 2σ uncertainty. Over the period 2004–2021, the average ChlOE trend for the three levels was −0.3 ± 0.4%/yr, but, with the addition of the unusually high ClO measured in 2022 and 2023 the trend from 2004 to 2023 became 0.0 ± 0.4%/yr. The increase in ClO during 2022 is, at least in part, caused by dynamical variations. The higher ClO in 2023 is closely correlated with the presence of increased H2O from the Hunga eruption, which affects the ClO chemistry
Approaches to causal inference using Bayesian machine learning and nonparametric methods with applications to the assessment of effects of family planning programming
No full textImproving access to Family Planning (FP) is a global priority. However, high-quality evidence on the long-term effects of FP programs is limited. This dissertation focuses on how to assess the causal effect of FP programs and interventions on outcomes of interest in various settings. In Chapter 1, we aim to compare FP-related performance outcomes across facilities providing FP services. The longstanding issue of how to implement a fair comparison is framed in terms of a causal inference problem: what would be the outcome had all facilities served a population with the same characteristics? We introduce an approach based on balancing weights to estimate the counterfactual. In Chapters 2 and 3, we estimate the effect of FP (adoption of modern contraception) on empowerment related outcomes (employment). Because adoption of modern contraception and employment share common causes, we focus on a subsample of women who changed their behavior due to an “external” reason, namely, the implementation of a FP program. Our approach, referred to as Prince BART, combines principal stratification to account for differences in FP behavior based on exposure to a FP program, with Bayesian Additive Regression Trees (BART), to non-parametrically model stratum membership and the outcome within each stratum. Prince BART allows for estimation of effect heterogeneity. Compared with chapter 2, chapter 3 examines the application more extensively, focusing on communicating the approach to a broader audience of applied researchers. In Chapter 4, we consider how to generalize findings from a study to a population of interest, when provided with a large probabilistic sample of the population of interest. We develop an approach to average conditional treatment effects over an estimated covariate distribution in the population of interest to obtain the population average treatment effect. Bayesian bootstrap is used to assess the additional uncertainty and to account for the complex sampling design.Chapter 1
Funding for this study was provided by the Bill & Melinda Gates Foundation under grant numbers OPP10709004 and INV-00844. Funders were not involved in any aspect of the study design, data collection, and analysis, nor the interpretation and writing of the manuscript.
A version of this chapter is under review at Health Services and Outcomes Research Methodology. I gratefully acknowledge my coauthors:
Godoy Garraza, L., Cardona, C., Gichangi, P., Thiongo, M., Anglewic, P., Alkema, L. (2024). Using balancing weights to compare performance across facilities providing family planning services in Kenya, Health Services and Outcomes Research Methodology, under review.
Chapters 2, 3 and 4
These papers were made possible by grants from the Bill & Melinda Gates Foundation and the Children’s Investment Fund Foundation, who support the work of the Family Planning Impact Consortium, under award numbers INV-018349 and 2012-05769. The findings and conclusions contained within do not necessarily reflect the positions or policies of the donors.
Versions of chapters 2 and 3 are under review at Observational Studies and Gates Open, respectively. I gratefully acknowledge my coauthors:
Godoy Garraza, L., Speizer, I., & Alkema, L. (2024). Combining BART and Principal Stratification to estimate the effect of intermediate on primary outcomes with application to estimating the effect of family planning on employment in sub-Saharan Africa (arXiv:2408.03777). arXiv., http://arxiv.org/abs/2408.03777, Observational Studies, under review.
Godoy Garraza, L., Speizer, I. S., & Alkema, L. (2024). How to estimate causal effects associated with family planning? An introduction to Prince BART, a new approach to effect estimation based on principal stratification and Bayesian non-parametric models. VeriXiv, 1, 5., https://doi.org/10.12688/verixiv.31.2, Gates Open, under review.
Thanks to the generous support of the Graduate School at the University of Massachusetts Amherst for awarding me the School of Public Health and Health Sciences Dean’s Ph.D. Fellowship and Dissertation Completion Fellowship.Doctor of Philosophy (Ph.D.
Variation and Processing of Subject Contact Relatives: A Comparison of AAL and MAE
No full textSubject contact relatives (SCRs) are relative clauses that do not require an overt complementizer (e.g., that, which) and where the relative clause is a subject-extracted relative clause. For example, “The girl bought the leash is looking for the dogs” would be a subject contact relative, whereas “The dogs (that) the girl found are happy” would be an object contact relative. In African American Language (AAL) both subject and object contact relatives are present, but Mainstream American English (MAE) only the object contact relative is allowed.
The central question of this dissertation is how speakers of comprehend complex structures that are present or not present in their dialect. To do this, I complete three experiments. In Chapter 2, I complete a judgment study to establish how the SCR structure is distributed in the community of AAL speakers that I work with. SCRs can occur in a variety of different forms that are not consistent cross-linguistically. Therefore, it was important to determine what types of SCRs the community of speakers allowed and disallowed. In this chapter, I find that there is a unimodal rather than a bimodal distribution in acceptance of the different types of SCRs. It could have been the case that members of the community do allow SCRs, but they could have accepted different types of SCRs. However, what I observed is that of the three different SCR classes tested, AAL comprehenders found them all to be acceptable.
In Chapter 3, I turn to computational modeling to investigate how predictive lan- guage modeling might account for SCRs when the language models are trained on mostly MAE language data. I complete two experiments. First, modeled after work done by Aina and Linzen (2021), I manipulate the length of the text preamble to un- derstand whether, and if so how, a off-the-shelf language model can make predicted continuations for SCRs. For example, and ambiguous preamble would be “Alisha was the one handed the chair ” because it could continue as a SCR (e.g., Alisha was the one handed the chair to the magician) or as a reduced relative clause (e.g., Alisha was the one handed the chair by the magician). What I find is that the language models are capable of producing text continuations in line with an SCR structure, but do so at a low and inconsistent rate, especially compared to other the unambiguous pream- bles. The second experiment in this chapter focuses on computational parsing, and I complete several parsing studies to investigate how a computational parser might as- sign structure to the SCR as well as whether augmentation of representative samples (i.e., including toy SCR examples) in the training data influence performance. Once again, the models are capable of assigning relations in line with an SCR structure, but do so at a low rate. This rate does improve when toy examples are included in the training data.
Lastly, I move onto human studies in Chapter 4. This chapter focuses on how speakers of AAL and MAE comprehend the SCR. I use a visual world paradigm study and manipulate the ambiguity of the SCR before a dismabiguating word. For example, “Alisha was the one handed the chair” is ambiguous for an AAL speaker between a SCR and a reduced relative clause, as discussed above. The presence of a preposition such as “to” or “by” disambiguates as it indicates if Alisha is receiving or giving the chair. With this design I test both AAL and MAE speakers. Special focus is given to the MAE speaking participants as the SCR is not present in their dialect. Using the Noisy Channel Model proposed by Levy (2008), I explore how these speakers comprehend a unfamiliar, ungrammatical, and “noisy” structure. The results suggest that MAE comprehenders prefer the rare parse, the reduced relative, over the “noisy” and ungrammatical option, the SCR. This is an interesting result as it contrasts with similar work done by Keshev and Meltzer-Asscher (2021) and their work on Hebrew. I conclude with a discussion surrounding how the incorporation of dialects can further inform linguistic study surrounding the Noisy Channel model, as well as ensure a more diverse representation of languages and dialects in the literature.
The key contributions of this dissertation are as follows: (i) it establishes the distribution of SCRs in an AAL speaking community, (ii) finds that off-the-shelf predictive language models can exhibit behavior that would allow for the presence of an SCR, but it does so inconsistently and at a low likelihood, and (iii) finds that MAE comprehenders prefer the uncommon reduced relative parse over the subject contact relative, following a Rare over Noisy pattern.Doctor of Philosophy (Ph.D.
Computational Investigations of Catalytic and Separation Approaches for Plastics Upcycling
No full textGlobal plastic production reaches 400 MM tons each year, and the amount is projected to increase to 700 MM tons by 2030. Efficient polymer deconstruction/upcycling strategies are lacking and reflected in how nearly 80% of used plastics go to landfills. In this dissertation, computational methods were applied to explore catalytic and separation approaches for plastics upcycling. First, ab-initio density-functional theory calculations were used to model olefin cross-metathesis, used in tandem chemistry with alkane dehydrogenation, to convert polyethylene into fuel- and lubricant- range hydrocarbons. The metathesis cycle and several reactions for active-site formation were examined on a silica-supported tungsten-oxide catalyst. To understand the behavior of polyolefin/hydrogen mixtures confined in catalyst pores for hydrogenolysis processes, atomistic simulations were used to calculate the solubility of hydrogen in both bulk and confined polyethylene. For separations, specifically process modeling of adsorption-based separation, the Real Adsorbed Solution Theory was systematically evaluated using binary adsorption of ethanol/water in hundreds of zeolites. Here, ethanol/water is studied as a representative system to demonstrate how zeolites can separate azeotropic mixtures when traditional distillation is not efficient. Another separation process is examined for antioxidant/polymer mixtures using molecular dynamics where the objective is to reduce catalyst poisoning by removing antioxidants from melt-polymer streams. Overall, these studies contribute to an improved understanding of catalytic and separation-based methods to address the growing levels of used plastics while highlighting the broad capability of computational tools for modeling complex systems.Doctor of Philosophy (Ph.D.)2026-02-0
TOWARD UNIFIED EXPERTISE: ONE MODEL FOR ALL TASKS
No full textUnderstanding the real visual world involves processing diverse forms of perception and learning the intrinsic connections among different perceptions. Humans exhibit a remarkable ability to adapt and respond appropriately to various types of visual stimuli, whether it's a glimpse of the 3D real world, perceiving a 2D black and white image, or watching a blurry video clip. In contrast, visual recognition systems often encounter challenges when learning from multiple sources. One such challenge is gradient conflict, where gradients from different tasks contradict each other. This conflict can lead to breakdowns in the systems' ability to learn across multiple tasks simultaneously. Another challenge is catastrophic forgetting, where a neural network trained sequentially on different tasks overwrites what it previously learned, and this can occur at any point between training iterations.
This dissertation aims to endow visual recognition systems with multi-task learning (MTL) ability. The aim is to enable these systems to transfer knowledge inductively between tasks. Deep learning naturally clusters similar concepts while maintaining separation between unrelated ones in the data and feature space. Here, the objective is to replicate this effect but optimize within the parameter and task spaces. Gradient conflict and catastrophic forgetting could be alleviated if parameters are carefully assigned to the best set of tasks. Challenges such as gradient conflict and catastrophic forgetting can be mitigated by strategically assigning parameters to the most suitable set of tasks. This allows for better performance across tasks without one task interfering with another. Along with this motivation, this dissertation seeks to identify the most effective neural network architectures for MTL. These architectures should support a scaling law, where increasing the model size, the amount of data, and the number of tasks leads to improved performance across a broad range of tasks—though with diminishing returns as the scale continues to grow.
We begin by addressing fundamental visual tasks such as object localization and object categorization. In the initial step, a unified framework was designed to incorporate these basic perceptual capabilities and enable knowledge transfer between tasks. The transformative dynamics between localization and categorization were parameterized and directly modeled to achieve this. This approach involves designing architecture and allocating the model parameters for various purposes, relying on human comprehension.
Beyond manual efforts in architecture design, we explore methods for automatically allocating model parameters to specific tasks. This involves creating a framework where different parts of the model can specialize in learning distinct tasks. To achieve this, we introduce the concept of the mixture of experts (MoE), where each expert represents a fundamental building block of the model. These experts can either be shared across a set of tasks or dedicated to a single task, depending on what the system needs. By structuring the model in this way, we avoid the limitations of sharing the entire backbone for every task, while still enabling knowledge transfer between them. We further extend this approach to manage a large number of tasks efficiently. Our strategy focuses on dynamically allocating resources to ensure that as the system scales with more tasks, it maintains high performance and efficiency, allowing the model to grow gracefully without overwhelming computational resources.
Further, the strategy of optimizing the parameter and task spaces can extend beyond efficient upstream pre-training to accommodate the diverse needs of downstream applications. In line with this approach, we delve into Dynamic Structured Optimization techniques for adaptable and efficient downstream learning. We explore how the adaptive nature of MoE layers can enable fine-tuning, support continual learning, and provide effective control over model capacity and computational cost.
Just as humans have specialized body parts—hands, brains, and more—each suited for specific functions, neural networks are composed of parameters that serve as the AI’s specialized components for different tasks. Our goal is to teach AI how to coordinate these diverse elements, much like the human body seamlessly orchestrates its parts, allowing it to manage a wide range of tasks. By optimizing how these components are allocated and adapt to different tasks, we aim to build AI systems that can handle complex and varied applications efficiently, while scaling gracefully.Doctor of Philosophy (Ph.D.
EVALUATING THE IMPACTS OF PRE-OXIDATION AND INTERMEDIATE OXIDATION ON DUAL MEDIA FILTRATION
No full textThe integration of oxidation processes in drinking water treatment has been observed to enhance coagulation, flocculation, and subsequent particle removal processes, effectively addressing dissolved metals, algae cells, and their by-products. Despite extensive research on the impact of various oxidants on coagulation and flocculation, the interaction between oxidation processes and particle filtration remains underexplored. This study involved analysis of full-scale filtration performance data as well as results from pilot-scale studies.
Full-scale operational data from the Putnam Water Treatment Plant (Greenwich, CT, Aquarion Water Company) spanning 2017-2023 was analyzed to assess how seasonal pre-chlorine dioxide (ClO2 ) addition influences physical and chemical treatment operations. Given the complex cause-effect relationship between ClO2 pre-oxidation and filter performance, the study also examined indirect impacts on media filter operation, considering variables such as raw water quality, chemical dosing, and pretreatment effectiveness. The full-scale data analysis highlights several challenges in correlating changes in unit filter run volume (UFRV) directly with pre-ClO2 treatment due to influential factors such as seasonal variations, coagulation effectiveness (linked to alum and caustic doses), and sedimentation effectiveness (related to floc-aid dose and indicated by filter influent turbidity). Moreover, pre-ClO2 affects natural organic matter (NOM) and manganese removal, modifying coagulant demands and pre-filter chlorine requirements. While pre-ClO2 impacts UFRV dynamics, these effects are embedded in a multifaceted framework where various other factors exert more significant influence. Nonetheless, pre-ClO2 tended to have a neutral to slightly beneficial effect on UFRV.
In the first pilot study evaluating the impact of pre-oxidation with chlorine dioxide and ozone on dual media filter performance, two experimental rounds were conducted: pre-chlorine dioxide before coagulation in round one and pre-ozone before coagulation in round two. Each round included a control train without pre-oxidants. The study monitored head loss development at four depths along the dual media filter as well as water quality parameters after each treatment process during a filter run. The results indicated no significant difference in overall head loss development between filters with or without pre-ClO2 or pre-ozonation. However, pre-oxidation altered head loss distribution among different media filter layers, more notably with ozone than ClO2. The pre-oxidants potentially facilitated firmer floc formation, better retained in the media filter, possibly due to the destruction of organic coatings on particles, enhancing floc aggregation.
The second pilot study assessed the impact of intermediate ozonation on dual media filter performance. Conducted with two parallel trains (one with post-clarification ozone and one control), the study monitored head loss and water quality parameters. Results showed that intermediate ozonation resulted in a lower terminal head loss development slope with the same alum dose, similar effluent quality in terms of UV254 and TOC/DOC, and higher turbidity and particle counts for particles under 3 µm in the ozonated water. The top 9.5 inches of the anthracite filter exhibited consistent head loss levels with and without ozone treatment, while the deeper anthracite and sand layers experienced slightly to significantly lower head loss when ozone was applied. These findings suggest that pre-oxidation and intermediate ozonation can influence filter performance, particularly in terms of head loss distribution and particle retention on the media surface.Doctor of Philosophy (Ph.D.
Adult women’s age differences in links between behavioral and physiological indicators of cognitive regulation.
No full textThe Neurovisceral Integration (NVI) model has demonstrated that there are many associations between physiological and behavioral indicators of cognitive regulation including among respiratory sinus arrythmia (RSA), frontoparietal coherence (FPc), and executive function (EF). EF is associated with both RSA and FPc, and each shows a developmental pattern that could be explained within the Selection, Optimization, and Compensation (SOC) model. Specifically, all develop in childhood, become more efficient in adulthood, and decline in old age. But how all three of these variables interact – especially in adult women – is still unknown. This dissertation consisted of three studies that tested the additive and interactive effects of age, RSA, and FPc on EF in adult women from both a variable- and person-centered approach. We expected to see age moderated interactions between RSA and FPc that predicted EF. Indeed, the Study 1 results indicated age moderated an RSA and FPc interaction that predicted EF. And those results seemed to follow a SOC type pattern. However, Study 2 did not generally replicate those findings. In addition, there were no significant person-centered results from Study 3. Despite the non-replication and lack of person-centered significant results, we caution against concluding with confidence that there are no developmentally based patterns between these physiological and behavioral indicators of cognitive regulation. We acknowledge that there are limitations in this series of studies that might account for the findings.Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) Grants HD57319 and HD60110, and National Science Foundation Grant BCS-1917857.
Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), R01HD049878.Doctor of Philosophy (Ph.D.
CASCADING CONSERVATION: ASSESSING THE REPEATED SPAWNING MIGRATIONS AND FRESHWATER MUSSEL HOST POTENTIAL OF BLUEBACK HERRING (ALOSA AESTIVALIS)
No full textAnadromous blueback herring (Alosa aestivalis) interact with numerous species during their spawning migrations between marine and freshwater systems. As such, the drastic declines in blueback herring populations and altered population dynamics over the last centuries may be affecting not only the stability of blueback herring, but also the species they interact with, including freshwater mussels. Effectively evaluating the impacts of blueback herring population changes on freshwater mussels requires unbiased demographic assessments of blueback herring data, such as age and number of spawning events, and a robust understanding of blueback herring as hosts for mussels. To assess the quality of spawning mark data, I examined the precision and bias between paired spawning estimates of 8,698 blueback herring collected by the U.S. Fish and Wildlife Service (USFWS) over the last decade. Systematic bias was found in 70% of years, and 30% showed coefficient of variation (CV) values >10% indicating imprecision. Bias was absent and precision was highest in most recent years where new training requirements and reference collections were used. These findings support future spawning estimation improvements through standardized precision thresholds and uniform training. To explore freshwater mussel use of blueback herring as hosts within the Connecticut River, the frontmost gills of 1,011 fish from the 2023 USFWS survey were examined for mussel glochidia (larvae). I observed 18,738 glochidia from four genera, including a species not known to inhabit the watershed (Utterbackia imbecillis), and a 64% infection prevalence (i.e. # of infected fish / # of total fish examined). Water temperature significantly affected the prevalence on blueback herring, with prevalence peaking at 14.2 ºC for alewife floater (Utterbackiana implicata) and 18.8 ºC for eastern elliptio (Elliptio complanata). For U. implicata, higher infection prevalence was related to higher densities of blueback herring, highlighting the linkage between these two species. Maximum infection intensities were predicted for smaller fish that were female and virgin spawners, which may be primarily explained by temporal patterns in migrating blueback herring and correspondence to glochidia release by mussels. These results highlight the value of understanding and monitoring organisms with unique, coupled life cycles toward developing effective, comprehensive conservation strategies.United States Fish and Wildlife Service Pathways ProgramMaster of Science (M.S.