Breastfeeding, a subject where trials are nearly impossible

Randomly assigning a feeding method is not conceivable. Available studies therefore compare families that made different choices, and those families differ in income, education, smoking and working conditions.

Those characteristics are themselves associated with the outcomes measured, which produces associations that tell us nothing about the effect of feeding. The direction of those gaps is not the same everywhere, moreover: in some countries prolonged breastfeeding is more common in well-off groups, in others the reverse, which alters the structure of the biases.

The Belarusian trial run by Michael Kramer and colleagues from 1996 randomly allocated maternity units, not families, between a breastfeeding support programme and usual practice. Known as PROBIT, it rests on cluster randomisation: it is institutions that are drawn by lot, which avoids assigning a practice to mothers.

The programme clearly increased the duration and exclusivity of breastfeeding, which allows two populations differing on that point to be compared without families having been sorted.

Publications report a reduction in gastrointestinal episodes and in certain skin conditions during the first year, results regarded as solid.

On long-term indicators, results are more mixed: published follow-ups find no clear effect on several later measures, notably indicators of body size and allergy. A follow-up published in 2008 reported a difference in some cognitive scores, a result that has been questioned because the people conducting the assessments knew which group the children were in.

Another approach compares, within the same family, children fed differently. It controls for everything the family shares. It has its own limitations: when a mother breastfeeds one child and not the other, it is often for a reason, the infant's health, the circumstances of the birth, a return to work, which may itself be linked to the outcomes.

That work, notably by Cynthia Colen and David Ramey published in 2014, finds markedly smaller differences than between-family comparisons, on most medium-term indicators.

Public discussion frequently mixes claims of different natures: public health recommendations, short-term established effects, disputed long-term effects, and considerations about the concrete conditions of implementation. Each belongs to a different type of evidence, and the solidity of one does not carry over to the others.

Health organisations maintain explicit recommendations while noting, in their technical documents, that the level of evidence varies by indicator, a mention that almost always disappears in coverage.