Machine learning approach to examine the intersectional association of social identities and circumstance with current cigarette smoking among US adults

Little is known about how the intersections of social identities and circumstances exacerbate cigarette smoking disparities among US adults. We analyzed data from the 1995-2019 Tobacco Use Supplement to the Current Population Survey (n = 1 496 458). Participants reported current cigarette smoking status (smoking cigarettes some days or every day vs not smoking at all) and 13 social identities (eg, race, ethnicity, Hispanic heritage) and circumstances (eg, education, marital status). We applied a statistical-learning boosting algorithm that allows interactions of these identities and circumstances to identify a minimal set of social identities and circumstances within each race/ethnicity with maximum predictive accuracy for current smoking. We then used weighted logistic regression models with interaction terms to estimate predicted marginal probabilities by 3-way combinations of these identities and circumstances. We found that social identities and circumstances used in this study predicted current cigarette smoking with varying degrees of accuracy by race/ethnicity, with highest accuracy among White adults and lowest accuracy among American Indian adults. Social identities and circumstances associated with current cigarette smoking differed somewhat by race/ethnicity (eg, citizen status was an important variable only among Hispanic and Black/African American adults). Prevalence of current cigarette smoking varied greatly by combinations of these identities and circumstances within each race/ethnicity (eg, 73.4% among 31-45-year-old American Indian adults in the Midwest whose spouse was absent vs 6.7% among American Indian adults in the South with bachelor's degrees and >$75 000 annual household income). These findings allow tobacco control researchers and practitioners to develop and deliver tailored interventions to reduce cigarette smoking disparities.