Methodological issues in studies comparing breast-feeding with not breast-feeding

There are specific difficulties when comparing outcomes in infants who are breast-fed v. those who are not, in particular that it is unethical and unfeasible to randomly assign healthy term infants to their mode of feeding. This precludes the use of the ‘gold-standard’ randomised trial approach for the majority of the population. In both randomised trials and observational studies, there is the additional problem that neither ‘breast-feeding’ nor the alternative (‘not breast-feeding’) are precise or constant entities; breast-feeding varies in duration and exclusivity, and nutrient intakes vary even within the same infant over the course of a day, while the alternative, ‘not breast-feeding’, can also vary in terms of the type of milk or infant formula used. Thus, in contrast to the situation in a pharmaceutical trial that compares fixed doses of drug and placebo for a fixed time period, there is much more variability or ‘noise’ in any comparison of breast-feeding v. not breast-feeding.

Other issues that are important when comparing the findings of the many different studies in this field include the classification and definition of infant feeding and the source of data on infant feeding. At one extreme, some studies use a crude measure such as ‘ever’ v. ‘never’ breast-fed, whereas at the other extreme, studies include much more detailed data on breast-feeding duration and exclusivity. Some studies obtain infant feeding data prospectively from the mother or medical records, whereas others rely on maternal recall, sometimes years after the event. It is well recognised that the apparent incidence of breast-feeding is influenced by the method used to obtain the information⁽Reference Flaherman, Chien and McCulloch⁶⁾. This problem could be addressed by adopting more standardised methods for collecting and classifying infant feeding behaviour in prospective studies. However, so far this has not been achieved.

The type of data collected on potential confounding variables also varies in quantity and quality between studies, and this is recognised to be a significant problem when performing systematic reviews and meta-analyses⁽Reference Horta, Bahl and Martinés^1, Reference Ip, Chung and Raman³⁾. It is particularly relevant in studies examining the relationship between breast-feeding and later cognitive outcome, where studies that adjust for maternal IQ are less likely to report a significant beneficial effect of breast-feeding on cognitive outcome than those adjusting for more proxy measures such as maternal educational level or social status⁽Reference Ip, Chung and Raman³⁾. However, this issue is also important for studies examining effects on later fatness or obesity risk, where parental BMI is a significant, but inconsistently measured, confounder⁽Reference Ip, Chung and Raman³⁾. Although confounding is a major issue for observational studies, it is also relevant for randomised trials where later follow-up is incomplete⁽Reference Fewtrell, Kennedy and Singhal⁴⁾.

Specific difficulties in conducting long-term follow-up studies

In order to examine effects of infant nutrition on later health outcomes, long-term follow-up is required. A major problem experienced to some degree by all studies in this field is that of attrition or loss-to-follow-up. The degree of attrition varies depending on the length of follow-up, nature of the investigations proposed and the inconvenience involved for the subject. The problem of attrition in long-term follow-up studies was recently discussed in detail in a paper that addressed the statistical implications of attrition for the interpretation of data⁽Reference Fewtrell, Kennedy and Singhal⁴⁾. Attrition can lead to selection bias, since subjects who are lost to follow-up are rarely identical to those who agree to participate, and also to a loss of power to detect a hypothesised effect. It may also affect the generalisability of the findings. Rather than imposing arbitrary figures for what constitutes an ‘acceptable’ follow-up rate, for example, 80% as used to separate ‘high’- and ‘low’-quality randomised control trials (RCT) in evidence-based medicine⁽Reference Sackett, Strauss and Richardson^{7, 8)}, it has been suggested that investigators should explicitly address the potential effects of attrition when analysing and reporting follow-up studies, so that readers are able to consider the likely impact when interpreting findings⁽Reference Fewtrell, Kennedy and Singhal⁴⁾.

Randomised trial of infant feeding in preterm infants⁽Reference Lucas, Gore and Cole⁹⁾

This randomised trial was initiated in 1982, at a time when there was considerable uncertainty about the best way to feed preterm infants. The diets commonly used to feed preterm infants at the time included unsupplemented banked donor breast milk and infant formulae designed to meet the nutritional requirements for healthy term infants. It was therefore considered both acceptable and feasible to randomise these vulnerable infants to their diet during the neonatal period to test the hypothesis that improving their nutrition would have benefits for both short- and long-term health outcomes. A total of 926 preterm infants (birthweight <1850 g and gestational age <37 weeks) were recruited from five UK neonatal units, and randomised to the milk received for, on average, the first month of life. In three of the hospitals, where there was a breast milk bank, the assignment was either to unsupplemented banked donor breast milk or to a newly developed nutrient-enriched preterm infant formula. In the remaining two centres, which did not have a breast milk bank, the assignment was to standard term infant formula or to preterm infant formula. In all cases, the randomised assignment could be fed as the infant's sole diet if the mother chose not to express any of her own breast milk, or given as a supplement to maternal breast milk if the latter was available. The comparison of donor breast milk v. preterm infant formula represents the only trial in which human infants have been randomised to breast milk v. infant formula without confounding by the mother's decision to breast-feed.

Infants enrolled in this trial were intensively monitored while they were in the neonatal unit, with daily records of nutritional intake, regular anthropometry and blood sampling. They were subsequently followed-up at intervals throughout childhood and into early adult life to examine effects of infant diet on a variety of later health outcomes, including neurodevelopment, bone health and cardiovascular health. It is important to note that many of the preterm infants in the study were never actually breast-fed; human milk was typically fed via a naso-gastric tube during the period of hospitalisation and many infants did not breast-feed after discharge. This is therefore primarily a study of ‘breast milk feeding’ rather than ‘breast-feeding’.

Randomised trial of a breast-feeding promotion intervention in healthy term infants (the Promotion of Breast-feeding Intervention Trial trial⁽Reference Kramer, Chalmers and Hodnett¹⁰⁾)

This cluster-randomised trial, conducted in the Republic of Belarus, compared outcomes in healthy term infants born in clinics that were randomly assigned either to a breast-feeding promotion intervention (modelled on the UNICEF ‘Ten steps to successful breast-feeding’⁽¹¹⁾) v. standard management. The study enrolled 17 046 mother–infant pairs from thirty-one clinics (the unit of randomisation). The intervention resulted in a significant increase in both the incidence and duration of breast-feeding and exclusive breast-feeding up to 1 year of age. For example, rates of ‘any breast-feeding’ for the intervention and control groups were 72·7% v. 60% at 3 months, 36·1% v. 24·4% at 6 months and 19·7% v. 11·4% at 12 months, respectively. Rates of exclusive breast-feeding were also significantly higher in the intervention group (43·3% v. 6·4% at 3 months and 7·9% v. 0·6% at 6 months).

This was not a trial of breast-feeding v. formula-feeding, but rather compared groups of infants who received ‘some’ breast-feeding’ v. those receiving ‘more’ and ‘more exclusive’ breast-feeding. Children from this study are being followed up throughout childhood, with 6-year follow-up data already published (see later).

Blood pressure

Two hundred and sixteen subjects (of 926 subjects originally enrolled and 831 survivors) from the preterm RCT were studied at 13–15 years of age. Those who had been randomised to receive human milk, either as sole diet or as a supplement to maternal breast milk, during the neonatal period had significantly lower mean and diastolic BP, with a mean difference of 4·1 and 3·2 mmHg, respectively, between groups (P<0·05 for both). This difference remained after adjusting for sex, neonatal Na intake and current BMI⁽Reference Singhal, Cole and Lucas¹²⁾. Furthermore, there was a ‘dose–response’ effect, with decreasing BP as the proportion of human milk received in the neonatal period increased (P=0·02 for trend). Although the observed effect size is small in terms of the effect at an individual level, it would be expected to have a significant impact at a population level in terms of reducing morbidity and mortality from CVD. For example, lowering population-wide diastolic BP by only 2 mmHg would be expected to prevent approximately 100 000 myocardial and cerebrovascular events per year in the United States alone⁽Reference Cook, Cohen and Herbert¹³⁾.

In contrast to these findings of an apparent beneficial effect of breast milk on later BP in subjects born preterm, at 6-year follow-up of the Promotion of Breastfeeding Intervention Trial (PROBIT) cohort, no significant effect of the intervention on BP was found⁽Reference Kramer, Matush and Vanilovich¹⁴⁾. One interpretation of this finding is that breast milk does not affect later BP in healthy term infants and that the results from previous observational studies suggesting a beneficial effect of breast milk on later BP in this population reflect residual confounding. However, interestingly, other studies in animal models (including baboons) and human subjects have found that effects of early nutrition and growth on later BP, plasma lipids and fatness may not ‘emerge’ until adolescence⁽Reference Lewis, Bertrand and McMahan^15–Reference Owen, Whincup and Odoki¹⁷⁾, and an earlier follow-up of the preterm RCT cohort found no effect of early nutrition on BP at 8 years of age⁽Reference Lucas and Morley¹⁸⁾. A third possibility, since the PROBIT study was essentially a trial of ‘some’ v. ‘more’ breast-feeding, is that the separation between the intervention groups in this study in terms of breast-feeding behaviour was insufficient to demonstrate an effect. This possibility is discussed further later in this paper.

Plasma lipid profile

Plasma lipids were measured at 13–15-year follow-up of the preterm RCT⁽Reference Singhal, Cole and Fewtrell¹⁹⁾. The group randomised to human milk either as sole diet or as a supplement to maternal breast milk during the neonatal period had significantly lower concentrations of LDL-cholesterol, apo-B, and a lower LDL:HDL ratio during adolescence, all likely to be favourable changes in terms of later cardiovascular risk. These subjects also had significantly lower levels of the inflammatory marker, C-reactive protein CRP. Once again, there was a dose–response effect with more favourable plasma lipid concentrations in those with the highest neonatal intakes of human milk (P=0·04 for trend for LDL:HDL). Plasma cholesterol is known to track with age⁽Reference Juhola, Magnussen and Viikari⁵⁾. The effect size of about a 10% reduction in plasma cholesterol found in this study is very similar to that reported from a meta-analysis of breast-feeding in term infants⁽Reference Owen, Whincup and Odoki¹⁷⁾. In terms of the likely impact on later health, systematic reviews in adults have estimated that a 10% reduction in the concentration of plasma cholesterol in plasma would reduce the incidence of CVD by 25%⁽Reference Holme^20, Reference Law, Wald and Thompson²¹⁾ and mortality by 13–14%⁽Reference Gould, Rossouw and Santanello²²⁾.

No plasma lipid results are yet available from the PROBIT cohort, although the 10-year follow-up study (currently underway) includes measurement of blood lipids.

Fatness and obesity

Follow-up of the preterm RCT cohort at ages 10⁽Reference Fewtrell, Prentice and Jones²³⁾, 13–15 years⁽Reference Singhal, Cole and Lucas^12, Reference Singhal, Cole and Fewtrell¹⁹⁾ and 20 years of age⁽Reference Fewtrell, Williams and Singhal²⁴⁾ found no significant impact of neonatal diet, including human milk, on body composition. The latter was measured using the in vivo gold standard ‘four component’ model at the 10- and 20-year follow-ups. The 6-year follow-up of the PROBIT cohort also reported no significant differences in weight, BMI, waist or hip circumference and skinfold thicknesses between intervention and control groups⁽Reference Kramer, Matush and Vanilovich¹⁴⁾. However, at 13–15-year follow-up of the preterm trial, subjects randomised to standard diets (donor breast milk or term infant formula) had a significantly lower leptin:fat mass ratio (a measure of leptin resistance) compared with those randomised to the ‘enriched’ diet (preterm infant formula; P<0·05), with an inverse dose–response relationship between neonatal human milk intake and later leptin resistance⁽Reference Singhal, Farooqi and O'Rahilly²⁵⁾. This finding raises the hypothesis that a lower nutrient diet during the neonatal period, of which human milk is an example, may result in lower leptin resistance that could potentially influence later fatness and obesity risk. However, this requires further investigation with longer follow-up.