. 2017 Jan 20;45(6):1866–1886. doi: 10.1093/ije/dyw314

Table 4.

Different approaches used in triangulation to determine the effect of having been breastfed on later body mass index

Approach	Brief description of approach	Sources of and potential direction of bias^a
Multivariable regression⁶¹	Systematic review and meta-analysis of prospective cohort studies in participants largely of European origin, including up to 355 301 participants in different analyses. BMI assessed across ages from 1 to 70 years	Residual confounding by maternal SEP and BMI (the majority of studies did not adjust for these) which would produce an effect estimate that is an exaggeration of any real effect
Cross-context comparison²¹	Multivariable regression in a UK cohort (BMI assessed at mean age 9; n = 4852) compared with similar multivariable regression analyses pooled across five studies from low- and middle-income countries (LMIC). BMI was assessed mean age 15–41 years; n = 10 912	It was demonstrated that SEP did not relate to breastfeeding in the LMICs or was in the opposite direction (more affluent and educated women being less likely to breastfeed) to that seen in the UK (breastfeeding more common in the more affluent and educated women). If an association in UK participants was due to SEP confounding, we would expect a null or opposite direction association in those form LMICs
Within-sibship comparisons^62–64	Three studies were identified which examined associations of being breastfed with BMI. Two of these used the same data from the US National Longitudinal Study of Adolescent Health (Add Health) and had similar results;⁶²^,⁶³ we present results from the study that focused on mean BMI (2734 sib-pairs; ages 12–18).⁶² The third study was also US-based and included 488 sib-pairs whose BMI was assessed at ages 9–19 years⁶⁴	In both studies breastfeeding was retrospectively reported by mothers when the children were adolescents, and misclassification is potentially the key source of bias here, which would be likely to result in an attenuated estimate of any true causal effect. Neither study had adequate statistical power to detect a difference between the effect estimates from the within-sibship analyses and that of unrelated participants in the cohorts
RCT⁶⁵^,⁶⁶	17 046 Belarusian women with healthy singleton births were randomized to a breastfeeding promotion intervention or usual care. The intervention resulted in marked differences in ever breastfeeding, duration of breastfeeding and whether breastfeeding was exclusive. An intention- to-treat analysis was used to assess the impact of these differences on BMI at age 6.5 (n = 13 879) and on BMI and fat mass index (FMI) at age 11.5 (n = 13 866).	Because the intervention affected multiple aspects of breastfeeding, it was not possible to do a formal IV analysis, but the intention-to-treat analyses would have similar assumptions to an IV analysis to test an intermediate. There would be violation of the exclusion restriction criteria if the intervention, in addition to influencing breastfeeding, also affected the mothers, such that they had a tendency to a healthier lifestyle more generally, including encouraging their child to have a healthier diet and be more active postnatally. That violation would result in an effect estimate that was an exaggeration of any true causal benefit of breastfeeding
Negative control study⁶⁷^,⁶⁸	We tried to identify outcomes in the ALSPAC cohort⁶⁷^,⁶⁸ that would be affected by confounders that are important in the relation of breastfeeding to later offspring outcomes but for which a biological/causal effect of breastfeeding was unlikely (home invasion by mice and by pigeons)	If the negative control outcomes are biologically influenced by the exposure, the interpretation of this approach would be biased. However, we consider it unlikely that house infestation by pigeons or mice would be affect being breast-fed, other than through confounding or other sources of bias. We explored associations of observed confounders with these outcomes and these verified that these were appropriate negative controls (Table S3)

In Supplementary text (available at IJE online) we provide full details of how we assessed a range of potential key sources of bias; here we describe the ones that we concluded were the key sources. As the length of exposure (to breastfeeding) varied across the studies or was not measured in some, we have not commented on duration of exposure; timing will be similar in all approaches.