The American Sociological Association's Methodology Section convened its 2026 Midyear Meeting on April 24 and 25 at the University of Toronto, hosted by the Data Sciences Institute. The conference brought scholars of all ranks from five different continents for two days of presentations on causal inference, computational methods, longitudinal analysis, and the role of large language models in social research, among other topics. The meeting was organized by Ethan Fosse, Nicholas Spence, and Angelina Grigoryeva of the University of Toronto, together with Xi Song of Columbia University.
"This year's program reflects where the section is heading," said Ethan Fosse, Associate Professor of Sociology and Associate Director of the Data Sciences Institute at the University of Toronto. "We have papers on foundational identification problems sitting alongside papers asking whether a language model can stand in for a survey respondent."
The keynote, "Multiverse Analysis: Toward Transparent and Robust Results," was delivered by Professor Cristobal Young of Cornell University and moderated by Fosse. Young argued that every empirical analysis rests on auxiliary assumptions about how variables are defined, which controls are included, and how missing data and outliers are handled. A point estimate bundles those choices together, and reporting a single preferred result creates an asymmetric information problem between author and reader. Multiverse analysis addresses this by estimating all combinations of plausible model ingredients and reporting the full distribution of estimates. Young illustrated the approach with applications, including the effect of education on voting for Trump in 2016 and a comparison with a "many analysts" crowdsourcing project on red cards in European football. He closed by suggesting that multiverse analysis will eventually sit alongside statistical significance as a standard tool, since significance asks what happens in repeated sampling while multiverse analysis asks what happens in repeated modelling.
The opening session featured Professor Christopher Winship of Harvard University, presenting joint work with Fosse on age-period-cohort (APC) analysis. The two argued that the field's standard estimators rest on identification assumptions that are rarely justified substantively, and that the same data can support meaningfully different conclusions depending on which assumptions an analyst is willing to defend. Instead of treating a single estimate as definitive, they urged researchers to use methods based on explicit, theoretically formed assumptions and to move toward more descriptive, data-driven approaches. The talk framed APC analysis as a case study in a broader problem the meeting kept returning to, namely, how empirical results depend on choices the analyst makes before any data are seen.
That theme surfaced again in Professor Sarah Mustillo's paper on growth mixture models, which examined how functional form, fit statistics, and starting values shape class enumeration when the data depart from normality. Mustillo, of the University of Notre Dame, was joined in this thread by Professor Anders Holm of Western University, whose paper on intergenerational mobility showed that commonly-used measures of mobility can disagree about what the same data say about college attainment, and by Samuel R. Lucas of the University of Toronto and the University of California, Berkeley, who introduced a margin-free correlation coefficient and argued that a more intuitive measure of association can change how researchers read familiar tables.
A second through-line ran through papers on large language models, where the methodological question is what these tools can credibly do for sociology and where they fall short. Professor Xi Song of Columbia University presented work using large language models to code job descriptions into occupational categories, a task that has long been a bottleneck in stratification research. Professor Lai Wei of the University of Hong Kong argued that AI predictions cannot replace surveys but may augment them, while Del Coburn of the University of Toronto examined alignment-induced measurement error when open-source models are used to simulate respondents with real demographic profiles. Other papers in the same vein treated LLM reasoning itself as data for social science inquiry and examined how transformer-based models can improve sociological theory. Together the papers suggested a section taking the new tools seriously without taking them on faith.
"These are not separate conversations," Fosse added. "The question of what we can credibly infer from our data is the same question whether the tool is a regression or a transformer, and the section is a key place where it gets argued out."
Co-organizer Nicholas Spence, Associate Professor of Health and Society, Sociology, and Physical and Environmental Sciences at the University of Toronto, emphasized the breadth of participation. "What stood out to us in assembling the program was the range of voices that wanted to be part of this conversation, from graduate students presenting their first methodological paper to senior scholars revisiting problems they have worked on for years." Fosse concluded: "Hosting these exchanges in Toronto, with the Data Sciences Institute as a partner, is exactly the kind of meeting the section should be holding."