A novel sparse model-based algorithm to cluster categorical data for improved health screening and public health promotion

Abstract Screening for interpersonal violence is critical to mitigate the consequences of violence and improve women’s health. Current guidelines recommend that health care providers screen all women for experiences of violence. Despite these recommendations, studies have noted a large variation in provider-reported interpersonal violence screening rates ranging from 10% to 90%. Given the disparity in screening rates, identifying variables correlated with providers’ screening practices is an important contribution. A survey of healthcare providers previously collected was utilized for this analysis and consisted of the providers’ socio-demographics, attitudes and beliefs, practice environment characteristics as well as self-reported screening practices. The objective of the study was to stratify healthcare providers into relatively homogeneous clusters based on mixed types of categorical nominal and ordinal variables and correlate the identified clusters with the violence screening rates. This paper proposes a sparse categorical Factor Mixture Model (sc-FMM) to cluster a large number of categorical variables, in which an norm was used for variable selection. An Expectation Maximization framework integrated with Gauss-Hermite approximation was developed for model estimation. Simulation studies show significantly better performance of sc-FMM than competing methods. sc-FMM was applied to identify clusters/subgroups of healthcare providers. The identified clusters were further correlated with interpersonal violence screening rates. The findings reveal how the providers’ screening rate for interpersonal violence are associated with multi-source impacting factors which inform the formation of policy and intervention development to promote the uptake of routine screening for interpersonal violence in women.