Study population

This cross sectional study was conducted based on the Japan Society of Obstetrics and Gynecology Perinatal Database (JSOG-DB), an ongoing registry based currently on 149 Japanese tertiary hospitals and covers nearly a tenth of all births in Japan, with over a hundred thousand births registered each year [24]. For this database, maternal demographics, pregnancy complications and birth outcomes were transcribed from medical charts in each hospital using a standardized format.

Multiple pregnancies are at higher risk of adverse outcomes compared to singleton pregnancy, and conception with ART is associated with both older age and multiple pregnancies. [25] Therefore, to differentiate the direct effect of maternal age on birth outcome from any indirect effect mediated by multiple pregnancies, [26] we included only women with singleton pregnancies. Similarly, we excluded women carrying a fetus with congenital abnormalities, as these women have a higher risk of adverse outcomes. Also, as the risk of adverse pregnancy outcomes in women of younger age is strongly related to social risk factors, [27] we restricted our sample to 370,964 women aged 30 years or older who gave birth to singletons with no congenital anomaly between April 2005 and December 2011. From this population, we excluded 5547 women with missing data on either gestational age (n = 207), birthweight (n = 2023), mode of delivery (n = 2210), and those with unreliable combination of birthweight and gestational age using the criteria proposed by Alexander et al. [28] (n = 1107). Among the other variables, smoking status, maternal height, pre-pregnancy body mass index (BMI) and gestational weight gain were missing in a large number of women. An additional 4393 had extreme values (> + 4SD or <−4SD) of height, BMI or gestational weight gain, thus we considered these data to be unreliable. To address these issues while maximizing our sample size to maintain the potential for a generalizable and robust analysis, we performed multiple imputation on the missing and unreliable data and pursued the main analysis on 365,417 women. These results were subsequently confirmed in a sensitivity analysis on the subset 183,084 women after excluding those with missing or unreliable data on height, BMI or gestational weight gain, and including “missing” as a smoking status (yes, no, missing).

For multiple imputation, we replaced missing or unreliable data with 30 sets of imputations for the following variables: maternal height (n = 157,767), maternal BMI (n = 120,257), maternal gestational weight gain during pregnancy (n = 134,122) and smoking (n = 153,652). For imputation, we used multivariate imputation by chained equations, which does not require the assumption of a multivariate normal distribution, and uses a series of regression models where each variable with missing data is modeled conditional upon the other variables in the data.