Talks and presentations

Open-ended survey items capture what closed-ended items cannot, and are routinely dropped from analysis because hand-coding them does not scale. catllm puts the whole coding loop inside Stata: extract/explore discover categories directly from the data, classify applies a category scheme and returns ordinary Stata indicator variables ready for tabulate or regress. It supports OpenAI, Anthropic, and Google models as well as local open-weight models through Ollama, so restricted-use data need never leave the analyst’s machine.

Large language models are increasingly used to classify open-ended survey responses, but they systematically over-classify, assigning categories too liberally on ambiguous cases and producing high sensitivity but low precision. Drawing on the established principle that aggregating multiple noisy annotators outperforms any single annotator, we test whether ensembles of LLMs can correct this problem. Using four open-ended survey questions with human-coded ground truth (3,208 responses, 6 categories per question), we evaluate ensemble configurations across 16 models spanning three cost tiers and six providers. Unanimous voting (requiring all models to agree before assigning a category) directly corrects over-classification by dramatically improving specificity: on the most ambiguous categories, the false positive rate drops from 50% to 3%, and precision triples. This advantage concentrates precisely where over-classification is worst, on subjectively ambiguous categories with fuzzy boundaries, while categories with clear criteria show no benefit. This pattern replicates across three independent datasets. Decorrelated errors are what drive the correction, and of the sources we test, cross-provider model access is the most reliable: models from different providers err differently on ambiguous cases, and consensus filters the idiosyncratic false positives. Neither temperature variation nor within-family size scaling reliably decorrelates errors. As few as three diverse lower-tier models suffice to reliably exceed GPT-5. For the ambiguous classification problems common in open-ended survey research, the well-established annotation principle of multi-coder agreement transfers directly to LLMs: investing in diverse perspectives is more effective than investing in a single expensive model.

Recent infectious disease outbreaks highlighted the critical role of human behavior and social processes in shaping infectious disease dynamics. Individual decisions—such as whether to wear a mask, get vaccinated, or work remotely—are not made in isolation, but are strongly influenced by the current state of the disease and individuals’ perceptions of risk. Such behavioral responses are also not fixed and change over the course of a pandemic. As these behavior changes accumulate across individuals, they scale up to shape population-level outcomes, including the timing, magnitude, and geographic spread of outbreaks. In order to curtail disease spread with minimal impact to a population, governmental interventions to control epidemics must therefore account for individual decision-making and, particularly, noncompliance or even defiance of recommended or even mandated policies. Modeling techniques developed to account for such dynamic human responses are therefore essential for understanding and managing infectious disease spread.

Presented as part of a Gateway/HCAP consensus session at AAIC on the Caribbean American Dementia and Aging Study (CADAS). Twenty-seven clinicians from five countries (Dominican Republic, Mexico, Puerto Rico, the United States, and India) independently classified 300 older adults using CADAS survey data, applying a structured 10-question NIA-AA diagnostic algorithm through a web-based consensus platform. Each case was reviewed by at least three clinicians per country team; cases with within-country disagreement underwent moderated consensus discussion, and international consensus was defined by majority rule across all five teams.

Contributed to the Sage Research Methods Video collection Doing Survey Research, produced by Sage Publishing as a teaching and training resource for students and researchers. In the interview I discussed dementia classification using survey data, the use of large language models for classifying open-ended survey responses, and approaches to quality control in survey data workflows.

Using four waves of the Health and Retirement Study (2016-2022), we document a sharp post-COVID cognitive decline in a nationally representative panel of older adults. Mean performance on the 27-point cognitive composite was flat across three pre-pandemic waves, then fell by approximately 0.7-0.9 points in 2022 — three to four times larger than any prior wave-to-wave fluctuation. This decline replicated across the full balanced panel and two non-overlapping Leave Behind cohorts, ruling out sample attrition and rotation-group assignment as explanations. Descriptively, both loneliness and structural social isolation were associated with lower baseline cognition and steeper decline. Loneliness showed an acute spike at COVID onset followed by partial recovery; structural isolation rose monotonically through 2022 with no sign of attenuation. Preliminary fixed-effects models provide initial causal evidence that within-person increases in structural isolation predict within-person cognitive decline (b = -0.30, p < .001). The UCLA loneliness scale adds an independent negative association (b = -0.47, p < .001). These estimates survive elimination of all stable individual traits via person fixed effects, strengthening causal inference. Whether isolation’s cognitive penalty was specifically amplified by COVID — beyond individual illness and mortality — awaits a fuller model with area-level contextual controls.

As large language model (LLM)–based text classification becomes routine in the social sciences, researchers confront dozens of competing models, inconsistent advice on prompting, and little standardized tooling with evidence-based defaults. CatLLM, an open-source Python and R package, addresses this gap with a three-stage pipeline — exploration, extraction, classification — for coding open-ended survey responses. The package supports multi-model ensembles, batch processing, and fully local deployment via open-weight models, allowing researchers working with sensitive data to avoid transmitting responses to external servers.

How effectively can Large Language Models (LLMs) approximate social scientist judgment in categorizing open-ended survey responses? This study compares eight contemporary LLMs—GPT-5, Claude Sonnet 4.5, Gemini 2.5 Flash, Grok 4 Fast, Qwen 3, DeepSeek v3.1, Llama 4, and Mistral Medium—to human annotators on 3,208 responses from the UC Berkeley Social Networks Study, spanning four question types and yielding 19,248 multi-label coding decisions. Models do not reach human-like inter-rater reliability in comparisons with individual coders (Krippendorff’s alpha 0.58–0.59 vs. 0.77 for humans), yet they achieve high accuracy rates of 82–97% relative to a human consensus standard, depending on task complexity. Accuracy declines for longer, more ambiguous responses and for rare thematic categories, indicating that model performance is sensitive to both response length and category prevalence. Demographic differences in performance are present—for example, responses from female respondents are classified less accurately—but much of this gap is associated with differences in response style, such as length and complexity, rather than clearly attributable to direct demographic targeting.

Personal networks influence health and mortality at the individual level, but less is known about how population-scale social network structure relates to mortality. This study examines how US county-level social network structure relates to mortality disparities. Using measures from 21 billion Facebook friendships, we investigate how two structural features of population social networks — cohesiveness and diversity — are associated with age-standardized and age-specific mortality rates. Bivariate results show that measures of social network structure rival smoking rates, median income, and educational attainment in their association with mortality rates. Social network structure remains predictive of mortality even after controlling for traditional measures like socioeconomic status and rural/urban classification. Network diversity is associated with lower mortality in both bivariate and multivariate analyses. Network clustering is associated with higher mortality bivariately, but this association reverses after controlling for county-level demographic and socioeconomic factors, revealing a protective effect masked by confounding. Age-stratified analyses further complicate this picture, showing that clustering predicts lower mortality among adults aged 15-64 but higher mortality among those 70 and older. These findings highlight social network structure as an important dimension of place-based health disparities, one not fully captured by conventional measures of socioeconomic composition or spatial segregation.

Invited talk as part of the session “Understanding NIH-Funded Training Opportunities for Early-Career Population Scientists,” chaired by Sarah Burgard (University of Michigan). Spoke about the benefits and challenges of applying for NIH F31 fellowship funding as an early-career population scientist.

Hands-on workshop on computational tools that augment social science research using large language models, presented as part of Session 103, “Innovative Data Methods and AI Applications,” at the 2026 Population Association of America (PAA) Applied Demography Conference. Introduced CatLLM—an open-source Python package that enables researchers to apply language and vision models to survey coding, image analysis, and data categorization without machine learning expertise. The workshop covered practical applications in demographic research, including automated coding of open-ended survey responses and analysis of cognitive health data.

Presented on computational tools that augment social science research using large language models. Introduced CatLLM—an open-source Python package that enables researchers to apply language and vision models to survey coding, image analysis, and data categorization without machine learning expertise. The talk covered practical applications in demographic research, including automated coding of open-ended survey responses and analysis of cognitive health data.

Joint presentation with Dr. William H. Dow examining the Hispanic health paradox in cognitive aging. The talk explored the contradiction in Hispanic aging where Mexican migrants, despite being more socioeconomically disadvantaged, demonstrate better cognitive aging outcomes than other Hispanic migrant groups. This research contributes to understanding health disparities and the complex relationship between socioeconomic status and health outcomes in aging populations.

Presented research with William H. Dow and Henry Dow comparing sociodemographic characteristics of Caribbean Hispanic older adult immigrants in the United States with those in their origin countries. Session chaired by Marc A. Garcia.

Participated in a panel discussion on the benefits and challenges of AI in demography at the 2025 PAA Annual Meeting. Discussed use-cases and benefits/limitations of large language models (LLMs) for survey research alongside other panelists.

Presented as a Wonderfest “Science Envoy” at “Ask a Science Envoy: Anthropocene Alarm; Partisan Contagion,” a public science event at HopMonk Tavern co-hosted with Marin Science Seminar, alongside UC Berkeley biologist Kristy Mualim (Genetic Biodiversity Loss in the Anthropocene). The talk explored how political partisanship significantly influences how different groups respond to public health guidance, affecting their adherence to protective measures against infectious diseases. Understanding partisan-based disparities in acceptance of scientific evidence helps us to predict the population-wide spread of diseases and to improve public health outcomes.

Radio interview as a Wonderfest Science Envoy on KPOO’s “Let Me Touch Your Mind” show, hosted by Marilynn Fowler. Discussed research on partisan-influenced health outcomes and how political partisanship shapes public health behaviors and disease spread.

Joint presentation with Dr. Audrey Dorélien examining partisan differences in the adoption of protective health behaviors during COVID-19 and modeling how these differences impact disease dynamics. Using detailed survey data on partisanship, contact rates, mask usage, and vaccination rates, we demonstrated that partisan differences in health behaviors exceed racial and gender differences. We incorporated these observations into a Susceptible-Infected-Recovered (SIR) model framework that explicitly incorporates partisanship to identify the most significant mechanisms driving disease spread, emphasizing the importance of considering partisan identification in public health policy and pandemic preparedness.

Democrats tended to have lower contacts, increased mask-usage, and a higher probability of vaccination during the COVID-19 pandemic. But, what happens when Democrats live in counties where they are the minority? How does this impact their behavior?

This talk examined how accurately leading large language models classify open-ended survey responses compared to human coders. Findings suggest that LLMs are best used to augment human coders in a loop rather than as a sole source of classification.

Despite rapidly expanding interest in fathers, scholars know little about the impact of fathers on adult health and well-being. How does the positive presence – or lack thereof - of fathers in adult children’s lives affect their social networks? Drawing on attachment theory and social capital theory to examine novel UC Berkeley Social Networks Study (UC Nets) data, I seek to extend understanding of how father attachment and socialization can influence adult social well-being. I find that individuals who name a father in their social network have significantly more social ties. Those with a “close” father have larger social networks than those who name a father who is not “close.” This has consequences for individuals’ broader networks: respondents with a “close” father report more males – but not females – in social activities networks. In contrast, having a “close” mother was associated with more females to confide in, but not males.