Survey non-response risks introducing systematic bias to studies of childhood cancer survivors, potentially distorting estimates of late effects and health outcomes. We evaluated the ability of logistic regression and five machine learning (ML) algorithms to predict non-response in two independent population-based studies of childhood cancer survivors in Germany. We analysed data from the VIVE study (n = 10,125; 48.5% respondents) and the E-SURV study (n = 2,292; 29.8% respondents). Non-response was the binary outcome. Candidate predictors included demographic, clinical, and geographical characteristics, as well as the German Index of Socioeconomic Deprivation (GISD). We trained logistic regression, elastic net regularisation, random forest (RF), gradient boosting machine (GBM), support vector machine (SVM), and k-nearest neighbours (kNN) classifiers within a stratified 80/20 train-test framework using repeated ten-fold cross-validation. Prior registry engagement was the dominant predictor of survey participation in both cohorts, with implications for non-response adjustment in patient-reported outcome studies.