This is a secondary analysis using a cross sectional data set collected by Dozier and colleagues in order to evaluate breastfeeding outcomes. The data were obtained from two linked sources; a single-time point mailed survey based on the Pregnancy Risk Assessment Monitoring System methodology (PRAMS) that contained questions regarding maternal beliefs, behaviors and experiences during the pre-pregnancy through postpartum period (data source one) (15). As described in Dozier et al. (16), a random sample of all mothers residing in Monroe County, New York with a live birth were mailed the survey instrument at approximately 4 months postpartum to women who had given birth between May 2009 and April 2012. The survey responses were then matched with the birth certificate (data source two), which contained additional information including demographics, medical information and history, delivery characteristics and birth hospitalization information for both mother and baby.
Birth certificate data were abstracted after delivery by trained birth certificate registrars based on the mothers’ prenatal and delivery records and infants’ medical records. This process is a standard part of state-mandated reporting for all live births, and is completely independent from the study protocol.
Inclusion criteria for this secondary analysis were: initiation of breastfeeding, complete responses to infant feeding questions allowing for determination of breastfeeding initiation, duration and exclusivity, complete height and pre-pregnancy weight for calculation of pre-pregnancy BMI category, non-underweight pre-pregnancy BMI, data on total weight gain during pregnancy, and reporting that the infant was alive at the time of survey completion. The latter criteria would not contribute to selection bias as infant mortality in this community is <7 per 1000 births.
Breastfeeding initiation (any and exclusive) was defined using data from the birth certificate in combination with survey responses. The birth certificate indicates whether mothers during the delivery hospitalization initiated exclusive or non-exclusive (i.e. was formula administered in the hospital along with breastfeeding) or did not initiate breastfeeding. The timing of breastfeeding initiation is assumed to be within the first 24 hours of life. In the rare cases where mothers initiated formula feeding only during the hospital and subsequently began breast feeding after they had been discharged from the hospital, their duration of any breastfeeding reflected this difference in our analyses. Survey responses were used to categorize the duration of any breastfeeding among all women who initiated breastfeeding, and the duration of exclusive breastfeeding among initiators of exclusive breastfeeding. Exclusive breastfeeding categorization considered not only the addition of formula, but also complementary foods and liquids. A mother’s duration of exclusive breastfeeding was categorized as the number of days of providing her baby only human milk, without the use of infant formula or other foods or liquids. The duration of any breastfeeding was categorized as the number of days a mother provided human milk to their baby, regardless of other sources of nutrition. Mothers who were still providing human milk (exclusively, or in addition to other foods) at the time of survey completion were censored at 3 months for these analyses.
Pre-pregnancy BMI was categorized into underweight (excluded), normal, overweight and obese based on the categories listed in Table 1, using the height and pre-pregnancy weight data abstracted by trained hospital birth registrars from the mothers’ prenatal records and entered into the corresponding fields of the electronic birth certificate. The final pregnancy weight was similarly abstracted from the hospital/labor medical record for entry into the birth certificate. This data source was utilized for weight and height, to minimize social desirability bias commonly associated with self-reported data.
Pre-pregnancy weight is what is documented on the prenatal record by the provider (and subsequently entered into the birth certificated (by the hospital birth registrars). This may be based on weight at the first prenatal visit or at the most recent pre-pregnancy visit. The weight at the end of pregnancy is abstracted from the documentation at the time of the mother’s admission when in labor. This is most commonly an actual weight. If that is not feasible then the last prenatal weight may be used or the field is left blank.
Pregnancy weight gain was categorized based on the Institute of Medicine’s 2009 recommendations was determined by subtracting pre-pregnancy weight from final pregnancy weight of each mother An interaction term including these two variables was created to allow investigation of the impact of pregnancy weight gain at each level of pre-pregnancy BMI. All combinations of categories of pregnancy weight gain and pre-pregnancy BMI were compared to women who gained as recommended and had a normal pre-pregnancy BMI; this served as the reference group for all other comparisons. Women who were recorded as having an underweight pre-pregnancy BMI were excluded given that the number of women in this category was small, and precluded their inclusion in statistical modeling.
Additional covariates considered in this analysis included: maternal age (continuous) maternal education (less than Bachelor’s degree/Bachelor’s degree or higher), race and ethnicity (white non-Hispanic/other), income status (low and non-low income; low income defined as enrollment in prenatal WIC and/or Medicaid funded delivery), marital status (married/not married), parity (continuous), smoking in previous two years (yes/no), infant sex (female/male) and vaginal delivery (yes/no). Education and race and ethnicity categories are expanded in Table 1 for descriptive purposes, but these variables were operationalized as above for all statistical analyses. All covariates evaluated utilized data from the birth certificate. To categorize income status and smoking in the last two years, responses from the survey were used when birth certificate information was missing.
Descriptive statistics (mean and standard deviation for continuous variables, number and percent for categorical variables) were used to characterize the sample. Bivariate associations of each outcome (time to any and exclusive breastfeeding cessation) and each covariate were evaluated using univariate Cox proportional hazard models. Bivariate associations between the exposure (pregnancy weight gain category) and each covariate were evaluated using Chi square tests (categorical), ANOVA (continuous, normal) or Kruskal-Wallis test (continuous, non-normal). Directed acyclic graphs were used to evaluate the potential confounding structure of the relationships of interest. Covariates were considered confounders of the association of interest if they were associated with both the outcome and exposure variables, and not on the proposed causal pathway.
Confounders were included in adjusted Cox-Proportional Hazards models of cessation of any breastfeeding among all women who initiated breastfeeding, and cessation of exclusive breastfeeding among only those mothers who initiated exclusive breastfeeding. For both models, pregnancy weight gain was the exposure of interest, and this term was interacted with pre-pregnancy BMI. Mothers with missing values on included confounders were excluded from these analyses. Unfortunately, due to a small number of mothers in the underweight BMI category, it was necessary to exclude these subjects from analyses. Continuous categorizations of BMI and pregnancy weight gain were not utilized given evidence of a non-linear relationship with the outcome of interest. Cox-proportional hazards model assumptions were checked using graphical assessment as well as interacting each covariate with time (PHREG procedure, SAS). SAS software version 9.4 and R version 1.1.453 were used for analysis (17, 18).