Records-based Childhood Cancer Research Applications and Methods

Abstract

Due to the rarity of childhood cancer, researchers often rely on methods such as registry linkage studies in order to gather sufficient sample size for analysis. These methods have been successfully used to establish associations between childhood cancer risk and birth certificate reported variables in the past. However, these methods have limitations. For example, variables such as maternal smoking may be misreported on birth records. Additionally, missing data, especially missing paternal data, is a concern. In manuscript 1, we investigated the association between maternal smoking and childhood cancer risk by pooling several existing registry linkage studies. We also applied a probabilistic bias adjustment in order to address misclassified smoking status. After adjustment, only other gliomas and other and unspecified malignant tumors had evidence of an association with maternal smoking. This finding falls in line with previous studies that have largely reported null associations, and greatly strengthens this finding through a strong study design and the use of probabilistic bias adjustment. In manuscript 2, we compared the performance of three different methods for handling the occurrence of impossible values (such as negative 2x2 table cell counts) in probabilistic bias analysis using simulation methods. We found that even in cases of a moderate probability of impossible values, all three methods—the removal of impossible iterations, the selection of a new parameter distribution, and Bayesian methods—performed similarly. However, further analysis is needed. In manuscript 3 we investigated the use of genotype data from newborn dried blood spots and other birth certificate derived variables in the imputation of missing paternal race/ethnicity. This analysis provides evidence that random forest methods can reliably impute missing paternal race/ethnicity using a child’s global ancestry and maternal race/ethnicity as predictors. We also demonstrated how these methods better situate prediction in specific populations when compared to ancestry cut-off-based methods.