Breastfeeding and the risk of childhood asthma: A two‐stage instrumental variable analysis to address endogeneity

    loading  Checking for direct PDF access through Ovid


Decades of research on the effect of duration of breastfeeding on childhood asthma, a chronic illness inflicting millions in developed countries, remains inconclusive.1 Although the World Health Organization (WHO), the United Nations Children's Fund (UNICEF) and the American Academy of Pediatrics (AAP) recommend 6 months of exclusive breastfeeding, many studies show that breastfeeding may actually increase the risk of asthma, some show breastfeeding has no effect, and others show a modest protective effect.1
Several meta‐analysis studies suggest the inconsistent and inconclusive results on breastfeeding and childhood asthma could be due to a number of research issues such as the inability to conduct controlled experiments, the sampling criteria used, and the analysis model.1 One limitation due to lack of control over the experiment and sampling criteria is that researchers often are unable to randomly select and randomly assign participants into treatment groups. This inability can lead to serious unintended consequences such as unmeasured confounders, selection bias, and reverse causality.1 If selection bias, reverse causality, and unmeasured confounders are not appropriately addressed in the statistical analyses, then the estimate of the coefficient, odds ratio, or relative risk ratio related to duration of breastfeeding will be biased.1
In the literature, the most common approach to control for the consequences of selection bias is to identify, measure, and adjust for potential confounders.1 These confounders must be related to both the treatment and health outcome. However, because of sampling criteria and data limitations, researchers are not always able to statistically adjust for a comprehensive set of confounders.1 For example, prior research has not statistically controlled for often unquantifiable behavioral and psychological factors related to breastfeeding although these can affect a mother's decision to breastfeed, which in turn can affect childhood asthma.1 If mothers of children with a family history of asthma are likely to breastfeed more believing that breastfeeding will protect their child against asthma or because highly educated mothers value breastfeeding, then a positive or no relation between breastfeeding and subsequent childhood asthma could be perceived as breastfeeding “causing” a diagnosis of childhood asthma or breastfeeding is not protective of childhood asthma.1 In contrast, mothers of children with a family history of asthma may decide to not breastfeed for the fear she may be “passing” asthma to the child,5 which can lead to the belief that not breastfeeding prevents asthma. As a further example, mothers suffering from post‐partum depression may not breastfeed the child and engage in other unhealthy behaviors such as smoking that increases the risk of the child developing asthma.7 Non‐breastfeeding mothers who engage in other unhealthy behaviors could heighten the risk of the child developing asthma.8
Thus, a second limitation of observational studies on breastfeeding and childhood asthma is that prior research uses statistical techniques that assume breastfeeding is an exogenous variable when in fact breastfeeding could be correlated with unadjusted confounders related to childhood asthma. If this occurs, then the statistical relationship between duration of breastfeeding and childhood asthma is affected by the presence of endogeneity between breastfeeding and childhood asthma.9 Endogeneity is a relatively serious statistical problem because it violates the assumption that breastfeeding is exogenous and uncorrelated with the error term, and this violation can bias the statistical effect of breastfeeding on asthma.1 If the exposure variable breastfeeding is correlated with the error term of the asthma model, then it is an endogenous variable. As the error term represents the effect of unadjusted confounders, the presence of endogeneity in this context suggests the presence of vital correlated omitted variables.

Related Topics

    loading  Loading Related Articles