Low Religion or No Religion? Improving Research with the Logic of Machine Learning

Abstract

Social science research is interested in the growing number of Americans who express no religious affiliation on surveys, but concerns about underreporting, misreporting, and inconsistency in lived religion raises a question about our most common survey measure of nonreligious self-identification. What is the predictive validity of our current explanations for why people disaffiliate? I advance the current literature using a logistic regression model for no religious affiliation fit on eleven samples from the General Social Survey (1988-2014) to predict respondents’ affiliation in the 2016 and 2018 samples. Results show our explanations can yield a fairly accurate predictive model, but errors are important and informative. The model is more likely to misclassify religiously unaffiliated respondents as affiliated. Analysis using model estimates shows that selection effects into non-affiliation explain differences in political views on culture wars issues. These findings challenge the use of categorical measures for nonreligion alone, because they suggest that measures of “low religion,” rather than “no religion,” are more useful for researchers seeking to overcome survey measurement error in studying this group.