Predicting cervical cancer screening among sexual minority women using Classification and Regression Tree analysis

Abstract

Cervical cancer screening is a critical preventive healthcare service for all women. Sexual minority women (SMW) in the United States experience multiple health disparities including decreased access to and use of cervical cancer screening. The mechanisms driving these disparities are not clear and SMW with multiple marginalized identities may be more likely to miss recommended cervical cancer screening. This study aimed to identify subgroups of SMW that are more and less likely to be screened for cervical cancer according to American Cancer Society guidelines. We used cross-sectional data from the latest (2010-2012) wave of the Chicago Health and Life Experiences of Women (CHLEW) Study (N = 691). Informed by intersectionality theory, we performed classification and regression tree (CART) modeling to construct a data-driven, predictive model of subgroups of SMW who were more and less likely to receive guideline-recommended screening. Notably, the CART model did not include commonly tested variables such as race/ethnicity or level of income or education. The model did identify subgroups with low likelihood of receiving screening and several novel variables that may be important in understanding SMW's use of cervical cancer screening; lifetime number of sexual partners, age at drinking onset, childhood physical abuse, and internalized homonegativity. Our results point to the importance of early life experiences and identity development processes in shaping patterns of preventive healthcare use among adult SMW. Our analysis also demonstrated the potential value of CART modeling techniques for evaluating how multiple variables interact in complex ways to predict cervical cancer screening.

Fig. 1

CART model predicting past-year Pap testing with complexity parameter set to 0.011 (Chicago, 2010–2012). Note: Decision tree models are interpreted based on both their overall performance in predicting the outcome accurately as well as individual terminal nodes that predict the outcome for specific subgroups of data. The “root node” of the decision tree displays the distribution of the outcome variable in the entire data set. Each subsequent “node” displays the next splitting variable, the number of participants represented by that node, and the percent of those participants with the outcome of interest. “Terminal nodes” display the outcome distribution (in this case, past-year Pap test) in final subgroups for which further splits would not improve prediction.