Objectives: Perceptions of health risks can guide individuals’ judgments, inform decisions about protective behaviors, and are often targeted in health behavior change intervention. We applied recursive partitioning analysis to determine the relationship between behavioral measures and self-perceived and Latent Class Analysis (LCA)-derived COVID-19 risks among residents of Onitsha city, Nigeria.

Methods: This study used data obtained from a cross-sectional survey conducted in March 2020 among residents (n=140) of Onitsha City, Nigeria. Participants assessed their self-perceived risk of COVID-19 infection, preparedness, knowledge, attitude, preventive practice, misconceptions, and information gap. We performed the recursive partitioning analysis using selected demographic characteristics and behavioral measures to develop decision-tree models for self-perceived (subjective) and LCA-derived risks (objective) prediction. The final model was selected based on the cross-validation R-square. The sensitivity and specificity of the outcomes’ probabilities were determined using receiver operating characteristic curve.

Results: The decision-tree model for self-perceived risk yielded 17 subgroups with 9 predictors (Entropy R²=0.342) and a misclassification rate of 0.2029. In comparison, the LCA-derived risk produced 5 subgroups with 4 predictors (Entropy R²=0.913) and a misclassification rate of 0.0221. Education (G^2=15.14, 24.7%), occupation (G^2=14.44, 23.6%), misconception (G^2=10.63, 17.4%), knowledge (G^2=7.03, 11.5%) and preventive practice (G^2=6.90, 11.3%) were significant (p<0.01) contributors to self-perceived risk. In contrast, LCA-derived risk was significantly (p<0.01) predicted by knowledge (G^2=81.73, 46.5%), attitude (G^2=62.81, 35.7%), misconception (G^2=24.26, 13.8%), and preventive practice (G^2=6.95, 3.95%). The pattern of response probabilities varied widely across the subgroups for “at-risk of COVID-19” and “not at-risk of COVID-19” categories for the self-perceived and LCA-derived risks. The area under the curve for the self-perceived risk and LCA-derived risk models were 86.9% and 99.8%, respectively.

Conclusions: LCA-derived risk produced a person-centered mixture model with parsimonious subgroups that offers a more reliable framework for identifying at-risk group, and designing effective and targeted interventions for the city residents.