Determination of pregravid BMI

Pregravid BMI was calculated based on self-reported pregravid weight and height at the first prenatal visit. Self-reported pregravid weights were examined for biological plausibility and imputed if deemed appropriate (<5% of weights were imputed). This imputed weight was calculated using the measured weight at the first prenatal visit (if taken <15 weeks) minus the recommended amount of weight to be gained in the first and second trimesters as defined by the IOM.

Study covariates

The PIN datasets include information from telephone interviews, self-administered questionnaires, medical chart abstraction, and biological specimen collection. Information on race/ethnicity (non-Hispanic Caucasian, non-Hispanic African American, and other) and maternal age was self-reported by the mother. Gestational age was estimated based on an algorithm that combined ultrasound dating with last menstrual period. If an ultrasound was done <22 weeks’ gestation, it was used to date the pregnancy. If no ultrasound was done or it was done later in pregnancy, last menstrual period was used to date the pregnancy. In the PIN cohorts, 90.7% had an ultrasound that was used to date the pregnancy, with the remaining 9.3% based on last menstrual period.

Outcome assessment

Clinically obtained weights were recorded at each prenatal visit. We calculated gestational weight gain as the difference between pregravid self-reported weight and the last weight prior to delivery. We defined excessive or inadequate weight gain based on the 1990 IOM recommendations ; these were as follows: 28-40 lb for low BMI (<19.8), 25-35 lb for normal BMI (19.8-26.0), 15-25 lb for overweight BMI (>26-29), and at least 15 lb of gain for obese BMI (>29). The IOM did not specify an upper limit for this group. For purposes of this analysis, excessive gain in the obese group was defined as >18 lb of gain, consistent with other analyses in the PIN cohorts. To calculate adequacy of gain for any given time point in pregnancy, the upper and lower limits of the weight gain intervals were extrapolated based on IOM-recommended rates of weight gain for the second and third trimesters, consistent with earlier studies in this cohort.

Genotyping

The Sequenom iPLEX platform (Sequenom, San Diego, CA) was used to genotype 27 SNPs associated with obesity and diabetes. For the purpose of quality control, 6 SNPs that had been previously assessed in the PIN cohorts were also genotyped. All SNPs were tested for Hardy-Weinberg equilibrium (HWE) among self-identified Caucasian participants.

Population stratification

In genetic association studies, differences in allele frequency among ethnic groups can confound relationships between genotype and disease outcome. To address population stratification in this cohort, genotyping was performed for 37 ancestry-informative markers that have been successfully used in other genetic association studies. STRUCTURE ( http://pritch.bsd.uchicago.edu/structure.html ) was used to infer population substructure and assign individuals to populations using probabilistic clustering methods. We analyzed self-identified Caucasian and African American participants separately, and we included probability of Yoruban ancestry as a covariate among self-identified African American women.

Statistical analysis

The DNA used for this study had been extracted for previous case-control studies of small-for-gestational age, preterm birth, and placental vascular disease, so prevalence of these outcomes was high. To produce estimates of the association between genotype and outcome that would approximate what we would have observed for the full study cohort, we calculated the probability of each participant’s inclusion in our study population. We used inverse probability weights to adjust our findings in all regression analyses. We used the SAS 9.2 surveylogistic and surveyreg (SAS Institute, Cary, NC) for these analyses.

We used linear regression to model associations between maternal genotype and total gestational weight gain, adjusting for maternal age, linear and quadratic gestational age at birth, as well as probability of Yoruban ancestry among self-identified African American women. We similarly used logistic regression to model associations between maternal genotype and probability of excessive gestational weight gain. We did not include in our models reproductive and obstetric factors that may be affected by genotype and may also affect weight gain, such as gestational diabetes and preeclampsia. These factors are potential intermediates on the causal pathway between genotype and weight gain, and including them in our models would attenuate the true association between maternal genotype and the outcome of interest. Moreover, because genotype may impact parity, we did not include parity as a covariate in our models.

We next considered models incorporating pregravid BMI in addition to gestational age and maternal age. Because the association between pregravid BMI and gestational weight gain is nonlinear, both linear and quadratic terms were included. We then used hierarchical selection to model: (1) quadratic models including the joint effects of SNP allele carriage and interactions between SNP carriage and both log BMI and log BMI squared; (2) linear models including the effects of SNP carriage and interactions between SNP carriage and log BMI; and (3) main effect models including only SNP allele carriage. If the Wald χ ² P value for model 1 was < .05, the Wald χ ² P value for the quadratic interaction term was determined. If this was < .1, then the quadratic interaction term was retained. We similarly evaluated the linear model, retaining the interaction term if the SNP and SNP * log BMI P was < .05 and the SNP * log BMI term was < .1. Finally, we retained the main effects model if the P for the SNP term was < .05. To avoid false-positive findings due to small cell sizes, we excluded SNPs with <5 homozygous low- or high-risk allele participants from these interaction models.

We did not adjust alpha levels for multiple comparisons in this analysis. We recognize that this approach may produce false-positive associations. However, the purpose of our pilot study was to investigate the strength and direction of associations between these diabetes and obesity SNPs and gestational weight gain. All results should be viewed as exploratory findings pending confirmation in larger cohorts.

We next considered the simultaneous effects of multiple SNPs, using Bayesian models. For these analyses, we analyzed the data for subjects with complete information on genotypes using additive parameterization for SNPs. We used linear regression with gestational weight gain as the response variable, and SNP carriage, obesity, and their interactions as covariates, adjusting for maternal age, linear and quadratic terms for gestational age at birth, and probability of Yoruban ancestry among self-identified African American women. We avoided using a model with interactions between SNPs and logBMI, logBMI squared as it would lead to a much larger model space and induce high correlations in the design matrix. The Bayesian variable selection framework allows each covariate to be either included or excluded from the model with a preassigned prior probability, which we chose as 0.5. After observing the data, the idea is to search over the list of all models, which includes models with no covariates, 1 covariate, 2 covariates, …, all covariates, to identify the models that explain the observed data the best. With our choice of prior distributions, the prior odds ratio (OR) of including a covariate vs excluding is 1. We report the covariates with posterior (after observing the data) ORs >3.