The methods for this study have been detailed previously [3
]. Briefly, a total of 84 member institutions of the COG provided cases for this study. Cases were eligible if they had a newly diagnosed GCT under the age of 15 years from January 1993 to December of 2002. Specific GCT diagnoses included germinoma, embryonal carcinoma, yolk sac tumor, choriocarcinoma, immature teratoma, and mixed GCT. Patients with tumors located in the brain were excluded. Cases were required to have a biological mother available who spoke English and have a telephone in their residence. The treating physician gave approval for study coordinators to contact the case’s mother.
Controls were selected through random digit dialing and were frequency matched to cases based on sex, age, and telephone exchange. Using methods similar to those of Waksberg [23
], potential control phone numbers were generated from case phone numbers. The area code and exchange of the case phone number were retained and the last four digits were randomly selected in order to obtain a control number. If a residential number was reached, screening questions were asked to determine whether or not a child under the age of 15 was present in the household. Case-control ratios were different based on sex; for males the ratio was 1:2 and females 1:1 since GCT in children is more common in females.
Approval for this study was granted by the institutional review boards at each participating institution as well as from the University of Minnesota.
Information on potential exposures was collected for cases and controls through maternal interview. The interview included questions about pregnancy history, maternal exposures during pregnancy with the index child, family history of cancer and other diseases, and information about the medical history of the mother. In the maternal questionnaire, several questions about infertility and infertility treatment were asked including length of time to index pregnancy, history of infertility (more than one year of trying without becoming pregnant), visit to a doctor by mother or index biological father due to non-pregnancy, specific infertility treatment, use of ovulating stimulating drugs in the year before or during pregnancy, history of multiple birth, and history of recurrent pregnancy loss.
Unconditional logistic regression was used to explore the association between infertility and infertility treatment and GCT. Exposures were included in the logistic analysis if at least four cases and controls were represented in all exposure categories. Matching factors sex and year of birth were included in all analyses. In addition, after exploration of potential confounding variables, the following variables were included in all models: maternal age, maternal race, household income, maternal education, and child’s gestational age. Results are reported as odds ratios (OR) and 95% confidence intervals (CI). Subgroup analysis was performed for age at diagnosis (<5, 5+) and tumor location (non-gonadal, gonadal), and gender. A sensitivity analysis was performed in subgroups limited to children with white, non-Hispanic mothers to assess the influence of race/ethnicity on results. Gender was analyzed in subgroups rather than through interactions since the etiology of GCT may differ by gender. All analysis was performed using SAS 9.1 (SAS institute, Inc, Cary, NC).
An additional method for assessing infertility was explored based on latent class analysis (LCA) [25
]. The variables used in the latent class analysis were maternal age, history of infertility, use of female hormones before or during pregnancy, history of multiple birth, and history of recurrent pregnancy loss. While each variable individually has the potential for misclassification of the true exposure of interest, using all variables together may mitigate this misclassification. Models with and without maternal age were explored to determine if the effect of infertility was only through maternal age or if there was an independent risk factor for infertility apart from age. LCA was conducted using M-Plus software [26
]. Predicted class membership was dichotomized and used as a predictor in a logistic regression model along with potential confounders.