Breast-feeding and maternal risk of type 2 diabetes: a prospective study and meta-analysis

The European Prospective Investigation into Cancer and Nutrition (EPIC)-Potsdam Study has 27,548 participants (16,644 women and 10,904 men). The recruitment took place between 1994 and 1998 and focused on the general population in Potsdam and the surrounding area. All participants provided informed consent, and permission was given by the ethics committee of the State of Brandenburg, Germany. The baseline examination involved a personal interview including questions on prevalent diseases and women’s health, a self-administered questionnaire about socioeconomic and lifestyle characteristics and number of births including breast-feeding, interviewer-conducted anthropometric measurements and blood sample collection [11]. We used a case–cohort design (electronic supplementary material [ESM] Fig. 1) to evaluate biochemical risk factors for diabetes, described in detail previously [12]. From 26,444 participants who provided blood samples at baseline, 2,500 individuals were randomly selected. Of 1,430 women within this random sub-cohort, 1,301 reported having given birth. We excluded women with missing data on breast-feeding behaviour (n = 17), missing data on oral contraceptive use, implausible energy intake (<3,349 or >25,121 kJ/day) and prevalent diabetes including gestational diabetes mellitus (GDM), as well as women with missing data on BMI at the age of 25 years and missing biomarker measurements, leaving 1,059 women for analysis in the sub-cohort. After application of similar exclusion criteria, 226 incident cases identified among all childbearing mothers in the cohort remained for analysis (overlap of 23 women with the sub-cohort due to the case–cohort design).

ESM Fig. 2 illustrates the time points of data collection and exposures.

Women recalled their age at childbirth and whether and for how long they breast-fed their children, separately for their first, second, third and last child (women with more than three children) in a self-administered questionnaire at the baseline examination. Eleven categories were given: 1 week or less; 2–3 weeks; 4–5 weeks; 6–7 weeks; 2 months; 3 months; 4–5 months; 6–7 months; 8–9 months; 10–11 months; 12 months or more.

Age at baseline examination and socioeconomic and lifestyle factors such as marital status, level of education, occupation, smoking behaviour and physical activity were assessed by a self-administered questionnaire and a personal interview. Weight, height and waist circumference at baseline examination were measured by trained interviewers who followed standard protocols under strict quality control. BMI at the age of 25 years was calculated from body weight at the age of 25 years (self-reported at baseline) and measured height. Dietary intake during the preceding 12 months was assessed through a validated food frequency questionnaire.

We used a large set of biomarkers reflecting insulin sensitivity and lipid metabolism as well as markers of liver fat accumulation, such as γ-glutamyltransferase and fetuin-A, and C-reactive protein (CRP) as an inflammation-related marker.

Biomarkers were measured in blood samples collected at the baseline examination and stored at −80°C or lower until analysis. Plasma CRP concentrations were measured with a high-sensitivity latex-enhanced immunoturbidimetric assay on an automatic Advia 1650 Analyzer (Siemens Medical Solutions, Erlangen, Germany). Plasma adiponectin concentrations were determined by ELISA (Linco Research, St Charles, MO, USA). Plasma levels of triacylglycerols, HDL-cholesterol, γ-glutamyltransferase and fetuin-A were measured with the automatic Advia 1650 Analyzer. For determination of fetuin-A, an immunoturbidimetric method was used with specific polyclonal goat antibodies to human fetuin-A (BioVendor Laboratory Medicine, Modreci, Czech Republic) [12]. All assay procedures were performed according to the manufacturer’s description. LDL-cholesterol levels were calculated using the Friedewald equation [13].

Every 2–3 years, follow-up questionnaires were sent out to identify incident cases of diabetes. All incident cases were verified by treating physicians, who were asked in a questionnaire to provide data on the date and type of diagnosis, diagnostic tests and the treatment. Cases confirmed by a physician (ICD-10: E11) and a diagnosis date after the baseline examination were considered to be confirmed incident cases of type 2 diabetes. For the present analysis, we used data collected until August 2005. Women with missing follow-up questionnaires were excluded from the analysis. However, the follow-up rate was high, exceeding 95%, and was similar between breast-feeding categories (data not shown).

Lifetime breast-feeding duration was calculated as the sum of breast-feeding periods for each child and was stratified into five categories: no breast-feeding; ≤3 weeks; >3 weeks to <2 months; ≥2 months to <6 months; ≥6 months.

To evaluate the association between single biomarkers and breast-feeding duration as a continuous variable (per additional 6 months), we used multivariate linear regression models. Biomarkers were not normally distributed after log-transformation, therefore Box–Cox transformation was used. Associations between breast-feeding and diabetes risk were evaluated using Cox regression modified for the case–cohort design according to the Prentice method [14]. The proportional hazards assumption was tested by plotting the Schönfeld residuals [15]. Age was used as the primary time-dependent variable in all models, with entry time defined as the participant’s age at recruitment, and exit time as the date of diagnosis, death or return of the last follow-up questionnaire. Cox models were stratified for age at baseline and further adjusted for marital status (unmarried, married, divorced, widowed), education (no vocational training or in training, vocational training, technical school, technical college or university), occupation (sedentary, standing, or [heavy] manual work), smoking behaviour (never smoker, ex-smoker, current smoker <20 units/day, current smoker ≥20 units/day), sporting activities (no sport, ≤4 h/week, >4 h/week), biking (no biking, <2.5 h/week, 2.5–4.9 h/week, ≥5 h/week), alcohol intake (no alcohol intake, 0 to ≤5 g/day, >5 to ≤10 g/day, >10 to ≤20 g/day, >20 to ≤40 g/day, >40 g/day), coffee consumption (ml/day), intake of red meat (g/day), intake of whole-grain bread (g/day), age at birth of last child, number of children, duration of oral contraceptive use (no use, ≤5 years, 6–10 years, >10 years), as well as BMI at the age of 25 years, baseline BMI and waist circumference. To evaluate potential biochemical mediators, we adjusted for different biomarkers determined in blood samples collected at baseline. Attenuation of the association indicates a mediator effect. We conducted several sensitivity analyses. We thereby stratified the analysis for number of children, educational level of the mothers, and time since last birth. Stratification for the age at first birth with a cut-off of 25 years was used to evaluate if BMI in young adulthood acts as both a confounder and a mediator. All data analyses were performed using the software package SAS Enterprise Guide 4.3 (SAS Institute, Cary, NC, USA).

We searched the PubMed and Web of Science databases for published studies on the association between breast-feeding and maternal risk of type 2 diabetes. A total of 300 references were identified from the two databases by combining text words and medical subject heading (MESH) terms in PubMed (the search strategy in the ESM Methods and the flow diagram in ESM Fig. 3). Eight additional references were identified by the Web of Science ‘Times cited’ function. The search was completed on 27 March 2014. Reference lists of retrieved studies provided no additional articles. Our inclusion criteria were: prospective cohort study; type 2 diabetes as outcome; description of breast-feeding assessment; presentation of relative risks with 95% CI; description of adjustment for potential confounders. We excluded animal studies and human studies that focused on children’s health or other outcomes such as weight change, metabolic changes, cardiovascular diseases or GDM. Unpublished material was not considered. The literature review was performed by two authors (S. Jäger, S. Jacobs), and data were extracted for multivariate-adjusted models (with and without adjustment for BMI). To evaluate the quality of the included studies, we adapted a score derived from Hu et al [16], which summarizes 14 aspects of each study (ESM Table 1). Meta-analysis was performed with small Stata, version 12.0 (Stata Corp, College Station, TX, USA) using fixed-effects models. Degree of heterogeneity was expressed as an I ² statistic, and Cochran’s Q test of heterogeneity (α = 0.05) was performed [17]. We assessed potential publication bias by regressing the standard normal deviate (HR/SE) against precision (1/SE) with α = 0.1 [18].