Andrei Boutyline
Sociological Science, May 29, 2017
DOI 10.15195/v4.a15
Abstract
Measurement of shared cultural schemas is a central methodological challenge for the sociology of culture. Relational Class Analysis (RCA) is a recently developed technique for identifying such schemas in survey data. However, existing work lacks a clear definition of such schemas, which leaves RCA’s accuracy largely unknown. Here, I build on the theoretical intuitions behind RCA to arrive at this definition. I demonstrate that shared schemas should result in linear dependencies between survey rows—the relationship usually measured with Pearson’s correlation. I thus modify RCA into a “Correlational Class Analysis” (CCA). When I compare the methods using a broad set of simulations, results show that CCA is reliably more accurate at detecting shared schemas than RCA, even in scenarios that substantially violate CCA’s assumptions. I find no evidence of theoretical settings where RCA is more accurate. I then revisit a previous RCA analysis of the 1993 General Social Survey musical tastes module. Whereas RCA partitioned these data into three schematic classes, CCA partitions them into four. I compare these results with a multiple-groups analysis in structural equation modeling and find that CCA’s partition yields greatly improved model fit over RCA. I conclude with a parsimonious framework for future work.
Measurement of shared cultural schemas is a central methodological challenge for the sociology of culture. Relational Class Analysis (RCA) is a recently developed technique for identifying such schemas in survey data. However, existing work lacks a clear definition of such schemas, which leaves RCA’s accuracy largely unknown. Here, I build on the theoretical intuitions behind RCA to arrive at this definition. I demonstrate that shared schemas should result in linear dependencies between survey rows—the relationship usually measured with Pearson’s correlation. I thus modify RCA into a “Correlational Class Analysis” (CCA). When I compare the methods using a broad set of simulations, results show that CCA is reliably more accurate at detecting shared schemas than RCA, even in scenarios that substantially violate CCA’s assumptions. I find no evidence of theoretical settings where RCA is more accurate. I then revisit a previous RCA analysis of the 1993 General Social Survey musical tastes module. Whereas RCA partitioned these data into three schematic classes, CCA partitions them into four. I compare these results with a multiple-groups analysis in structural equation modeling and find that CCA’s partition yields greatly improved model fit over RCA. I conclude with a parsimonious framework for future work.
 | This work is licensed under a Creative Commons Attribution 4.0 International License. |