Asian Journal of Probability and Statistics

COVID-19 vaccine hesitancy remains a public health concern in Nigeria, and understanding both predictive factors and plausible causal pathways may inform more focused interventions. This secondary analysis integrated machine-learning prediction with causal-inference methods using cross-sectional survey data from Oshodi/Isolo Local Government Area, Lagos State. The full descriptive, causal and predictive modelling sample comprised 651 respondents.  Least absolute shrinkage and selection operator (LASSO) logistic regression identified vaccine safety perception, perceived vaccine effectiveness, side-effect perception, travel history, and age as nonzero predictors. Binary logistic regression (BLR) and support vector machines (SVMs) with linear, polynomial, and radial-basis-function kernels were evaluated. The polynomial SVM had the highest test area under the receiver operating characteristic curve (AUC = 0.913) and specificity (0.922), whereas BLR had the highest sensitivity (0.769) and accuracy (0.845). For causal estimation, affirmative perception that the vaccine was safe was treated as the exposure and vaccine acceptance as the outcome, with unsafe and unsure perceptions combined as the comparison exposure category. Marginal structural modelling (MSM), inverse probability weighting (IPW), and augmented IPW (AIPW) yielded estimated risk-difference effects of 0.595, 0.589, and 0.590, respectively. The numerical agreement across estimators is consistent with a large adjusted contrast in this dataset, but causal interpretation remains conditional on exchangeability, positivity, consistency, correct model specification where required, and an assumed temporal ordering that cannot be established from cross-sectional data. These findings support vaccine-safety confidence as an important target for further longitudinal investigation and carefully evaluated communication interventions.