The effects of infant feeding on short- and long-term growth, including the development of childhood obesity and the possible “programming” of long-term body size and health, have been of major interest throughout the world, including both developed and developing countries. Even if the question is narrowed, however, difficulties arise in making causal inference because of problems specific to this area of endeavor.
The specific causal question of interest is this: Does prolonged and exclusive breastfeeding cause infants to grow more slowly than other infants (those who are formula-fed, or those who are breastfed to a lesser degree or for a shorter duration)? This question is extremely difficult to answer. Human beings are not genetically identical laboratory rats who can be randomly assigned to various modes of infant feeding. Mothers and infants cannot be “forced” into prolonged and exclusive breastfeeding, not simply because of the obvious ethical and practical obstacles, but because infant feeding and growth are both dynamic processes, each able to influence the other.
Attempts to address the causal question can include both observational and experimental studies. One can use an observational (eg, cohort) design to compare infants who are breastfed with those who are formula-fed, with further refinement achieved by categorizing breastfeeding according to its duration and degree of exclusivity. Still, confounding and reverse causality plague causal inference in observational studies of infant feeding and health.1–3 Another design option is a randomized controlled trial (RCT) of a breastfeeding promotion intervention. In such attempts, noncompliance with the allocated intervention becomes a huge problem. Large overlaps in the distributions of duration and degree of breastfeeding are likely, even if eligibility is restricted to mothers who initiate breastfeeding, and the intervention is designed to promote greater exclusivity and duration. This problem of overlapping distributions requires very large sample sizes to detect the effect sizes that can be reasonably expected. Moreover, the causal question then changes, at least if the analysis is based on intention to treat (ITT); it is no longer the effect of prolonged and exclusive breastfeeding, but the effect of the breastfeeding promotion intervention. Other approaches, however, including instrumental variable analysis4 and Bayesian modeling of compliance-adjusted causal effects,5 maintain many of the advantages of randomized allocation without altering the causal question.
Figure 1 is plotted from data taken from a pooled analysis of seven small cohort studies carried out in high-income countries from the late 1970s to the early 1990s.6 The graph depicts the weight-for-age z-score of infants who were breastfed for at least 12 months in these seven studies, with the reference population (z = 0) consisting of the then CDC/WHO growth reference,7,8 based largely on formula-fed infants born between 1929 and 1975. Despite the small sample sizes, the weight trajectories are remarkably similar. These studies showed that the infants began life on the heavy side (not surprising for highly selected breastfeeding participants willing to participate in studies like these), with an increase in z-score in the first 3 months followed by a dramatic reduction relative to the reference, reaching a nadir at 12 months and then partial “recovery” (ie, increase toward the reference mean) in the second year of life.
As mentioned, however, the then CDC/WHO reference was based primarily on formula-fed infants. The use of that reference led many infants who received prolonged breastfeeding to be diagnosed with “failure to thrive” as they crossed weight percentiles, particularly in the second half of their first year. It was therefore believed that the reference was inappropriate for breastfed infants, ie, infants following WHO’s infant feeding guidelines. In addition, the sample size of infants on whom the reference was based was too small to classify the very low or very high weight extremes with adequate precision. This led WHO to carry out its Multicentre Growth Reference Study, based on healthy populations in Brazil, Ghana, India, Norway, Oman, and the United States.9 Participants were highly selected, advantaged populations in each of the study countries. Moreover, the published growth reference is based on the 50% who followed WHO’s breastfeeding recommendations, ie, those who were exclusively breastfed for at least 4 months, continued partial breastfeeding to at least 12 months, and began consuming complementary (solid) foods by 6 months of age. It is from this study that the current WHO infant growth reference was developed. It is a “prescriptive” reference (how infants should grow), based on a highly selected groups of infants who were healthy enough and growing well enough to continue following WHO’s feeding recommendations. It is not a descriptive reference that describes how any preselected sample of newborns actually grows during infancy and early childhood. This fact raises questions as to whether the reference reflects a true “effect” of the recommended feeding on infant growth, or whether the infant’s growth itself affects the mother’s and infant’s ability to comply with the feeding recommendations. In other words, how much of the observed association between feeding and growth is the result of reverse causality?
PROMOTION OF BREASTFEEDING INTERVENTION TRIAL
During the mid-1990s, we initiated the Promotion of Breastfeeding Intervention Trial (PROBIT), a cluster-randomized trial of a breastfeeding promotion intervention in the Republic of Belarus.10 The intervention was based on the ten steps of the WHO/UNICEF Baby-Friendly Hospital Initiative.11 The clusters randomized were 31 maternity hospitals and one affiliated polyclinic (outpatient clinic where children are followed for routine health surveillance and illness care) per hospital. A total of 17,046 healthy breastfed infants born at ≥37 weeks’ gestation and weighing ≥2500 grams at birth were enrolled during their postpartum stay. This sample size (number of clusters and number of infants) was based on the primary trial outcome, which was the occurrence of one or more gastrointestinal infections in the first year of life. The births occurred between June 1996 and December 1997.
Follow-up research visits at the polyclinics were scheduled at 1, 2, 3, 6, 9, and 12 months of age, at which time data were collected on current infant feeding, infections (particularly respiratory and gastrointestinal) since the previous visit, and measurements of weight, length, and head circumference. Of the 17,046 mothers and infants who were enrolled, 16,491 (97%) were followed through 12 months of age.10
Figure 2 shows the effect of intervention on the duration of any breastfeeding. All of the infants were initially breastfed (an inclusion criterion), but the experimental and control groups began to deviate by the first month of life, and the difference persisted throughout the first year of life. The cluster-adjusted hazard ratio for weaning (discontinuation of breastfeeding) was 0.70 (95% confidence interval [CI] = 0.59–0.83) in the experimental group vs. the control group. An even greater effect of the intervention was observed for degree of breastfeeding, with 43% of infants in the experimental group exclusively breastfeeding at 3 months vs. 6% in the control group (cluster-adjusted hazard ratio for discontinuation of exclusive breastfeeding = 0.29 [95% CI = 0.19–0.46]). Very few infants in either group were exclusively breastfed at 6 months (7.9% vs. 0.6%), likely reflecting the fact that WHO’s feeding recommendations at the time were for exclusive breastfeeding for 4 to 6 months, rather than the current recommendation of 6 months.12
GROWTH ANALYSES IN PROBIT
To illustrate the biases inherent in observational studies of infant feeding and growth, we have analyzed infant growth in PROBIT using several approaches.13 The main approach, of course, is by the ITT, ie, comparing growth in infants randomly assigned to the experimental vs. control groups, irrespective of “compliance” with the randomized allocation. We also used observational analyses, however, based on actual feeding received. The 1378 infants who were weaned in the first month were used as a proxy for formula feeding. In addition, we categorized another group comprising 1271 infants following, as close as possible, the WHO’s feeding recommendations at the time: exclusively breastfeeding for at least 3 months with continued partial breastfeeding to at least 12 months (the fact that no PROBIT visit was scheduled at 4 months prevented us from identifying infants exclusively breastfed for 4–6 months). Finally, a small group of infants (n = 269) was identified following the current feeding recommendations of WHO, ie, exclusive breastfeeding for at least 6 months with partial breastfeeding to at least 12 months. The latter group is contained within the previous group. Figure 3 shows the results of these analyses in graphical terms. The ITT analysis showed a reduction in weight for age in the first month, with continued increases thereafter of weight for age relative to the then WHO/CDC reference. Weight for age was higher in the experimental group than in the control group starting at 1 month of age, but by 12 months of age, the two groups were virtually indistinguishable. The early weaning group had the sharpest decline in weight for age in the first month (probably contributing to the decision to wean), with catch-up weight gain relative to the two experimental groups by the end of the first year of life. The two groups following the WHO’s feeding recommendations (pre- and post-2001) showed larger weight-for-age z-scores in the first 3 months, followed by a decline that continued until 12 months; the decline, however, never dipped as low as the reference mean of 0. For the observational analyses, this pattern of results raises the question as to whether the trajectories seen in the two WHO’s feeding recommendation groups represent a true effect of feeding on growth or, as mentioned earlier, an effect of growth on the ability to follow the breastfeeding recommendations—reverse causality.
DOES GROWTH CAUSE CHANGES IN INFANT FEEDING?
We recently used our observational data from PROBIT to empirically demonstrate reverse causality, ie, the effect of prior growth on subsequent feeding decisions.14 We restricted our analysis to infants who were breastfed (or exclusively breastfed) at the prior PROBIT visit. We assessed crude associations between weaning (or discontinuation of exclusive breastfeeding) during an interval and infant size at the prior visit, ie, the beginning of the interval. We also examined mixed random-effects logistic models fit using penalized quasi-likelihood (SAS PROC GLIMMIX) with random effects for polyclinics, also adjusted for important potentially confounding covariates including maternal education, maternal smoking at the onset of each interval, geographic region (east vs. west Belarus), and urban vs. rural residence.
The Table presents the results of the multivariate models, showing adjusted odds ratios for weaning and discontinuation of exclusive breastfeeding according to the weight-for-age z-score at the prior PROBIT visit. As the Table clearly shows, a low weight-for-age z-score (<−1) at 1 month increased the risk of both weaning and discontinuation of exclusive breastfeeding by 2 months relative to z-scores from −1 to +1 (reference). Conversely, high weight-for-age z-score at 1 month reduced the odds of weaning or discontinuation of breastfeeding. Similar results were seen for low weight-for-age z-score at 2 months and weaning and discontinuation of exclusive breastfeeding at 3 months, and for weight-for-age z-score at 3 months and feeding at 6 months. These results clearly show a strong association between prior weight for age and subsequent infant feeding, conditioned on feeding at the time of the previous visit (when the weight was obtained). The findings strongly suggest that infant size affects breastfeeding, perhaps via the mother’s or the physician’s confidence in the mother’s milk supply and her ability to continue breastfeeding.
Although our results show that reverse causality can lead to associations between larger prior infant size and subsequent feeding decisions, we suspect another source of reverse causality to be responsible for the observational association (seen in Figs. 1 and 3) between prolonged and exclusive breastfeeding and smaller subsequent size. We suspect that slow-growing infants are “satisfied” with breastfeeding, whereas fast-growing infants are hungry in response to their faster growth, leading to crying and other signals of need for increased energy intake. We further hypothesize that the crying undermines the mother’s confidence in her milk supply and hence leads to formula supplementation, which subsequently leads to reduced infant demand for breast milk and to consequent reduction in the mother’s milk supply and to eventual weaning.
Figure 4 shows our postulated directed acyclic graph for infant feeding and growth based on the results of PROBIT and our analysis of the current evidence base.15 The inferential problem here is that the mother’s decision whether to continue breastfeeding or to supplement or wean her infant is made between visits and often based on infant signals (crying, fussing, etc.) that are extremely difficult, if not impossible, to measure (and therefore to control for) on a dynamic, day-to-day basis. In PROBIT (and all other studies of which we are aware), both infant size and infant feeding are measured at follow-up visits or telephone interviews, whereas infant signals and maternal feeding responses to those signals occur between those visits or interviews. The causal inference made based on the feeding at time j and the size at time j + 1 is likely to be interpreted as a change in feeding affecting the growth, rather than a change in growth (or hunger) affecting the feeding. Thus, feeding at time j is used as a proxy for feeding within the interval j to j + 1. It may be a poor proxy, however, because of change in feeding (supplementation or weaning) induced by hunger signals reflecting the infant’s faster growth.
This interpretation leads to problematic inferences in the direction of cause and effect. In practical terms, sorting out the cart from the horse in this dynamic relationship between infant growth and feeding is a formidable obstacle to causal inference. We believe that the “cleanest” causal interpretation is the one emanating from the RCT design and ITT analysis (Fig. 3), which shows very different results from those obtained in observational studies. Mothers and infants randomized to the breastfeeding promotion intervention in fact grew more rapidly in the first 3 to 6 months than those randomized to the control intervention. Although similar differences in the first 3 months of life have also been seen in observational studies (see Fig. 1), the two randomized groups in PROBIT converge rather than deviate from one another after 3 months, unlike the results reported from observational studies of prolonged breastfeeding relative to a growth reference based primarily on formula feeding (Fig. 1).
Because PROBIT was restricted to mothers and infants who initiated breastfeeding (the vast majority of mothers and infants in Belarus and many other countries), these results cannot necessarily be extrapolated to differences between prolonged and exclusive breastfeeding and formula feeding. Nonetheless, they suggest the need for caution when inferring causal effects of infant feeding on growth during this critical period. The complete convergence in growth (weight, length, and head circumference) achieved at 12 months by infants in the PROBIT experimental and control groups, as well as the absence of differences at age 6.5 years in height, weight, BMI, or skinfold thicknesses,16 also argue against causal effects of prolonged and exclusive breastfeeding on long-term risks of obesity. Although breastfeeding has many other well-established positive (and likely causal) effects on child health, survival, and neurocognitive development, it should not be expected to provide protection against obesity (and its metabolic and cardiovascular health consequences) in later childhood and adulthood.
ABOUT THE AUTHORS
All three authors are members of the Department of Epidemiology, Biostatistics, and Occupational Health at the McGill University Faculty of Medicine. MICHAEL KRAMER and ROBERT PLATT are also members of McGill’s Department of Pediatrics. ERICA MOODIE’s main research is in statistical methodology for causal inference, with a particular focus on adaptive treatment strategies, and Robert Platt’s is in statistical methods in perinatal epidemiology and pharmacoepidemiology. Michael Kramer is a perinatal epidemiologist with primary interests in the causes of adverse pregnancy outcomes and the child health effects of breastfeeding.
1. Bradford HA A Short Textbook of Medical Statistics. 1977 London Hodder & Stoughton:27
2. Sauls HS. Potential effect of demographic and other variables in studies comparing morbidity of breast-fed and bottle-fed infants. Pediatrics. 1979;64:523–527
3. Kramer MS, Kakuma R. Optimal duration of exclusive breastfeeding. Cochrane Database Syst Rev. 2002:CD003517
4. Greenland S. An introduction to instrumental variables for epidemiologists. Int J Epidemiol. 2000;29:722–729
5. Frangakis CE, Rubin DB, Zhou XH. Clustered encouragement designs with individual noncompliance: Bayesian inference with randomization, and application to advance directive forms. Biostatistics. 2002;3:147–164
6. Dewey KG, Peerson JM, Brown KH, et al. Growth of breast-fed infants deviates from current reference data: a pooled analysis of US, Canadian, and European data sets. World Health Organization Working Group on Infant Growth. Pediatrics. 1995;96(3 Pt 1):495–503
7. World Health Organization. The Growth Chart: A Tool for Use in Infant and Child Health Care. 1986 Geneva; Switzerland World Health Organization
8. US Department of Health, Education and Welfare. . NCHS Growth Curves for Children, birth-18 Years. DHEW Publication (PHS). 1977:78–1650
9. WHO Multicentre Growth Reference Study Group. . WHO Child Growth Standards based on length/height, weight and age. Acta Paediatr Suppl. 2006;450:76–85
10. Kramer MS, Chalmers B, Hodnett ED, et al.PROBIT (Promotion of Breastfeeding Intervention Trial) Study Group. Promotion of Breastfeeding Intervention Trial (PROBIT): a randomized trial in the Republic of Belarus. JAMA. 2001;285:413–420
11. WHO/UNICEF. Protecting, Promoting and Supporting Breastfeeding: The Special Role of Maternity Services. 1989 Geneva, Switzerland World Health Organization
12. World Health Organization. . Infant and Young Child Nutrition. 54.2. 2001 Geneva, Switzerland Fifty-Fourth World Health Assembly
13. Kramer MS, Guo T, Platt RW, et al.PROBIT Study Group. Breastfeeding and infant growth: biology or bias? Pediatrics. 2002;110(2 Pt 1):343–347
14. Kramer MS, Moodie EEM, Dahhou M, Platt RW. Breastfeeding and infant size: evidence of reverse causality. Am J Epidemiol. 2011;173:978–983
15. Kramer MS, Moodie EEM, Dahhou M, Platt RW. Response to “Causation or ‘noitasuaC’?”. Am J Epidemiol. 2011;173:988–989
16. Kramer MS, Matush L, Vanilovich I, et al.PROBIT Study Group. Effects of prolonged and exclusive breastfeeding on child height, weight, adiposity, and blood pressure at age 6.5 y: evidence from a large randomized trial. Am J Clin Nutr. 2007;86:1717–1721