Developing and Validating Predictive Models of Vaccine Hesitancy Among Parents in the United States

OBJECTIVES: We used machine learning to better understand the factors that drive parental decision-making with respect to vaccinating their children.

METHODS: We conducted a cross-sectional survey with parents (N=692) of children under 18 years old in the US in 2022. Using the survey responses, several machine learning algorithms were implemented to predict parental hesitancy towards vaccinating their children, including logistic regression, decision tree, random forest, XGBoost, support vector machine, and neural network. Predictors included factors related to health literacy, information seeking behavior, and attitudes and beliefs. Performance was evaluated by F1 score and area under the receiver operating characteristic curve (ROC-AUC) in the testing set. Important features were extracted from the model with highest performance based on SHAP (SHapley Additive exPlanations) values.

RESULTS: Overall, more hesitant parents were more likely to be younger, male, and Hispanic/Latino. Random forest performed the best in predicting vaccine hesitancy (F1 score= 0.86, ROC-AUC= 93.00%), followed by XGBoost (F1 score= 0.84, ROC-AUC= 92.75%). Based on the random forest model, the belief in “no need for their children to get vaccinated because everybody else does” factor contributed significantly to higher hesitancy, followed by beliefs related to a “lack of trust in vaccines,” children “getting too many vaccines,” and “healthy children don’t need vaccines.” The model reflected that information-seeking challenges and concerns over safety, efficacy, and side effects were strong predictors of attitudinal shaping.

CONCLUSIONS: This study introduced an effective machine learning approach to help providers and policy makers understand and monitor factors that shape attitudes and influence behaviors towards vaccination, and disentangle how parents interpret information discussed in shared clinical decision-making.