Research

Chris Soria is a PhD candidate in Demography at the University of California, Berkeley, where he studies how social networks shape cognitive health and dementia in aging populations. His NIA F31-funded research uses causal inference methods to examine how personal relationships and network structures influence cognitive aging and health disparities across diverse social contexts. His work has been presented at the Population Association of America (PAA), the American Public Health Association (APHA), and the Pacific Sociological Association (PSA), and published in the American Journal of Epidemiology and The International Journal of Aging and Human Development.

Soria also develops computational tools that augment social science research using large language models. After encountering challenges analyzing open-ended survey responses in his research, he created CatLLM—an open-source Python package and web app that enables researchers to apply language and vision models to survey coding, image analysis, and data categorization without machine learning expertise.

Social Network Determinants of Health and Cognitive Aging

My research examines how social network structure and quality influence health and cognitive aging across individual and population scales. At the individual level, I use longitudinal data from the Health and Retirement Study to disentangle the effects of objective social isolation from subjective loneliness on cognitive trajectories. Building on the concept of Social Network Cognitive Buffers, I investigate whether restoring lost connections or forming new ones can slow—or even reverse—cognitive decline.

In collaboration with Dennis Feehan, I extend this work to population-level network structure. Using measures from 21 billion Facebook friendships, we examine how county-level network cohesiveness and diversity relate to US mortality disparities. Network structure rivals traditional predictors like smoking and income in its association with mortality, and age-stratified analyses reveal that clustering is protective for working-age adults but associated with higher mortality among older populations.

This research aims to inform interventions that strengthen social networks to promote health across the life course.

Partisanship, Health, and Mortality

Our goal is to understand how political networks shape disease transmission during pandemics. We investigate how partisan affiliation influences both health behaviors and contact patterns, and how social network structure—specifically the tendency for partisans to interact more frequently with politically similar others—amplifies or dampens epidemic dynamics across communities.​​

A collaboration with Audrey Dorlien, Ayesha Mahmud, and Dennis Feehan.

We ask: How do estimates of prevalence and mortality change when we acknowledge that Democrats and Republicans not only behave differently during pandemics, but also interact within distinct social networks? Our research reveals that Republicans report 20% more daily contacts than Democrats, adopt protective behaviors at lower rates, and exhibit strong political homophily—preferentially mixing with other Republicans. These network structures create insular transmission pathways where infection dynamics diverge sharply between partisan groups.​​

Using detailed contact survey data from the Berkeley Interpersonal Contacts Study, we model disease transmission across partisan networks with a three-group framework that incorporates group-specific contact matrices and varying levels of political homophily. Our simulations show that network structure can more than double the infection gap between Republicans and Democrats, shift epidemic peaks by up to 46 days, and substantially alter cumulative mortality. Models that ignore partisan network structure systematically mispredict both epidemic timing and magnitude, revealing that political identity operates not just through individual health decisions, but through the structure of social mixing itself.​​

Utlizing Large Language Models for Assisting the Research Process

Large Language Models (LLMs) are transforming social science research by automating time-intensive tasks that have historically created barriers to qualitative analysis. Open-ended survey questions provide rich insights but have been underutilized due to the weeks or months required for manual coding. LLMs can process thousands of responses within hours while achieving human-level accuracy, democratizing sophisticated text analysis for researchers with limited resources.​

My research demonstrates that leading models like GPT-4o and Claude 3.7 Sonnet achieve 88-97% similarity to expert human coders across various categorization tasks. These models excel at straightforward classification while maintaining reliability on complex interpretive challenges, and they serve as quality control tools that can identify inconsistencies in human annotation. The CatLLM Python package and web app operationalize these capabilities by providing a standardized framework for applying language and vision models to common research workflows.​

Despite their potential, LLMs function best as supplements to human judgment rather than replacements. Models struggle with vague language, implicit meanings, and very long responses, and researchers should remain vigilant about potential demographic biases. When implemented thoughtfully with clear category definitions and human oversight for complex tasks, LLMs can dramatically expand the scale and scope of qualitative research.

Other Research

My research portfolio spans several collaborative projects examining social networks, mobility, and inequality across diverse populations. I am currently assisting Professor Claude Fischer on a project using open-ended survey data from the UC Berkeley Social Networks Study (UCNets) to better understand the causal predictors of internal US mobility. In a separate collaboration with William Dow, I examine how migration and network patterns help explain old age disability outcomes in migrant populations. I have also contributed to research assessing the 10/66 algorithm for dementia classification across diverse international settings.

Beyond demographic and health research, I support work examining systemic inequalities in higher education. This project investigates biases in graduate school admissions that may contribute to broader educational inequality. Using application data from over 100 graduate programs, we analyze how public school applicants fare relative to private school applicants in the admissions process. We further examine whether the educational backgrounds of decision-making committee members influence admission outcomes, specifically testing whether public school applicants receive more favorable consideration when committee members share similar educational backgrounds.

This project is a collaboration with Matthew Stenberg, Sara Quigley, Catherine Madsen, and Lisa Bedolla.