Evidence for causal associations between prenatal and postnatal antibiotic exposure and asthma in children, England

Higher risks of asthma have been observed in children with prenatal exposure to antibiotics and during early life compared with those who have not. However, the causality of such associations is unclear.


| INTRODUC TI ON
There is evidence that children whose mothers received antibiotics during pregnancy (prenatal exposure), and children who were exposed to antibiotics in early life (postnatal exposure), are at higher risk of allergic diseases and asthma. [1][2][3][4] However, it is not clear to what extent this association is due to bias or is a causal association.
Causal mechanisms are plausible: antibiotics taken by the mother in pregnancy and passed to her child, or taken by the child in early infancy, can lead to a Th-2 skewed immune response and asthma. 5,6 Given that most cases of asthma in the UK are allergic, it is plausible that changes in the immune system are a common ultimate proximal causal pathway shared by prenatal and postnatal antibiotic exposures. However, not all cases of asthma are allergic, and so, there can be different causal mechanisms in different asthma sub-types. 7 If it is a causal association, this is important from a public health perspective, given the large proportion of pregnant women and children who use antibiotics (33% and 70%, respectively, estimated for our target population), and the high prevalence of asthma in many settings. 7 However, there is also evidence that this association with antibiotics can be due to unmeasured confounders. For example, maternal exposure to antibiotics before pregnancy, [8][9][10] and paternal exposure 9 have been reported as associated with asthma. Furthermore, it has been described that this association between prenatal and postnatal antibiotics and asthma decreased when assessed with sibling analysis that controls for shared familial and indoor environmental factors. 11 In our observational study, we investigated whether prenatal and postnatal exposure to antibiotics had a causal association with an increased risk of asthma in children recruited to the Born in Bradford birth cohort study in England (BiB cohort). 12 We used several approaches to deal with different sources of bias and to assess causation (i.e., triangulation) 13 : maternal prescriptions 0-12 months before pregnancy were used as negative control exposure, the analysis was repeated excluding children with lower respiratory infection and/or asthma at 0-4 years to avoid reverse causation with postnatal antibiotic exposure, and sibling analysis was conducted to control for prenatal shared time-invariant maternal and environmental factors. We also assessed the heterogeneity of effects by mother's ethnicity and interaction between prenatal and postnatal exposures, and the disease burden.

| ME THODS
The target population was 25,534 children born at the Bradford Royal Infirmary between 2007 and 2011, of which 13,858 (54%) were recruited to the BiB cohort. 12,14 We included all children of the BiB cohort except 173 children who died. The study population was the 13,685 children alive at the time of this study, from 12 Evidence for causal associations between prenatal and postnatal antibiotics exposure and asthma in children, England.
• These associations were observed after procedures to minimize bias and not observed with exposure to antibiotics before pregnancy.
• We conclude that these findings support causal effects in our target population. was invited to take part in the BiB cohort. Consenting women signed a consent form.

| Sources of data
Data on ethnicity, socio-economic and lifestyle factors were obtained from a baseline questionnaire administered at recruitment during pregnancy. Primary care clinical event and prescription data were obtained from linked electronic health records (EHR). 15 Obstetric and birth data were extracted from the hospital electronic maternity records, and data on children's body mass at 4-5 years were taken from the National Child Measurement Programme.

| Outcome definition
We conducted case-insensitive text mining of the EHR for the term "asthma". Of the 162 Read codes found (diagnostic codes that are used in primary care in the UK), we selected 148 to define asthma (Table S1), including Read codes validated in previous studies. 16

| Exposure
We searched the EHR for generic and common brand names of oral and systemic antibiotics (case-insensitive text mining), and BNF codes. We classified maternal exposure during pregnancy (prenatal exposure) as four dichotomous variables: first trimester (<93 days of gestational age); second trimester (93-184 days); third trimester until 28 days before the child's birth; and the period 7-27 days before the child's birth (see Table S2). Maternal prescriptions at 0-12 months before pregnancy were used as the negative control exposure, under the assumption it is associated with the same confounders that are F I G U R E 1 Derivation of the study population associated with maternal exposure during pregnancy, but with an implausible causal effect with asthma in the child. 18 Postnatal exposure was defined as a child's prescriptions of systemic antibiotics between 0 and 24 months old.

| Covariates
We first drew a directed acyclic graph with the main risk factors for asthma and then selected the variables related to the following: (1) the mother (history of asthma, eczema or hayfever, smoking, selfdefined ethnicity, country of birth, age at child's birth, education, employment, diabetes during pregnancy, parity); (2) family size and (defined as EHR records with the terms "lower respiratory tract infection", "pneumonia", "syncytial infection" or "lower respiratory infection"). Maternal history of asthma and child's gender were always maintained in the regression models. Model 5 (minimally adjusted model) was derived from model 3 after F I G U R E 2 Strategy of analysis to assess associations between exposure to antibiotics and asthma. After the directed acyclic graph, we selected the available variables representing the factors with a strong association with the outcome and their proxies based on a hypothetical causal framework. The analysis started with an exploratory data analysis to describe missing data, extreme values, sparse data, conflicting data from different sources and multicollinearity backward elimination (package "abe" in R), 19 based on change-inestimate of odds ratio (OR) ⪆10% and without using statistical significance. We also ran a matched analysis with conditional logistic regression with groups of siblings from the same mother, to control for shared time-invariant maternal and environmental factors. 20 We estimated the E-value: "the minimum strength of association, on the risk ratio scale, that an unmeasured confounder would need to have with both the [exposure] and the outcome to decrease the association to the null (RR = 1)." 21 The interaction between prenatal and postnatal exposures, and heterogeneity of effects according to the mother's ethnicity (the distribution of the risk factors for asthma vary in relation to ethnicity 22 ) and mode of delivery (caesarean vs.

| Statistical analysis
vaginal) were assessed with log-binomial regression due to the high prevalence of asthma. The risk of disease in the population due to the exposure (population attributable risk, PAR), 23 the population attributable fraction and the excess number of cases of asthma in the target population attributable to antibiotics were estimated for different hypothetical scenarios if the antibiotics had no effect on the asthma cases (Stata commands punaf and regpar). 24 We conducted sensitivity analyses to assess the following: ascertainment bias (stratification by the total number of days with GP visits or to the clinics a child had between 0 and 24 months old, excluding days with records on asthma, wheeze and LRTI); clustering effect (robust variance estimator, mothers as clusters); effect of each variable excluded from the MLR (missing data or bivariable screening); and effect of breastfeeding, positive skin prick test, child care and pets, with data available in subgroups of the BiB cohort 22,25 (unmeasured confounders). 26,27

| RE SULTS
The CCAs were conducted on 12,476 children (Table 1). Differences were observed between CCA and 151 children with missing data, especially in relation to lower proportion of children whose mothers were born in the UK, mothers not employed and higher maternal education. In the CCA, 10.6% (n = 1,322) of children had asthma, 46.3% (n = 5,774) had mothers of Pakistani heritage, and 46.1% (n = 5,746) had mothers born outside the UK.

| Initial data analysis
In the 12,476 children, several factors were associated with an increased risk of asthma, in addition to maternal and postnatal exposure to antibiotics (Table S3): they were related to children (male, born by caesarean section, low birthweight, prematurity, obesity at 4 years old and LRTI in infancy) and to the mother and family (Pakistani origin, history of allergic diseases, not born in the UK, not being employed at baseline and as number of people in the household increased).

| Association between exposure to antibiotics and childhood asthma
In the unadjusted association, the risk of asthma was higher among children whose mothers received antibiotics at 0-12 months before pregnancy (negative control exposure) and in each trimester of pregnancy, and among children exposed to antibiotics at 0-24 months old (  Table S4. These associations were observed for all classes of antibiotics, but the small numbers did not allow further comparisons between these different classes ( Table 3).
The risk of asthma increased as the number of prescriptions of any antibiotics at 0-24 months increased, however with overlapping confidence intervals ( Table 4). Analysis of dose-response with exposure during pregnancy at 7-27 days before birth was not conducted due to small numbers.
Adjustment for clustering effect and the variables excluded from the CCA changed the OR between −0.7% and +15% in comparison with the fully adjusted model (model 3 in Table 2 of delivery, child obesity at 4 years, low birthweight and maternal age at birth). There was a higher risk of asthma among those children exposed at 0-24 months with OR = 1.99 (1.00, 3.93) (Table S5).
Results for maternal exposure had small numbers of children and therefore were difficult to interpret.

| Sensitivity bias analysis
The adjusted risk ratio (

| E-value
For the association with exposure at 7-27 days before birth, the ob-

| Heterogeneity of effect/interaction
The children who were exposed at 7-27 days before birth (prena-

For all classes of antibiotics b
To the mother 7-27 days before birth c a Each class of antibiotics each time and also adjusted for maternal history of asthma and child's gender. b With the variables for all classes of antibiotics together in the regression, and adjusted for maternal history of asthma and child's gender.
c The variable on macrolides was excluded because there were no asthma cases among the exposed.

TA B L E 4
Percentage of asthma cases (based only on Read codes between 5 and 8 yo) at 5-8 years old and the 95% confidence intervals, separately for number of prescriptions of any antibiotic to the children at 0-24 months old, adjusted for maternal history of asthma and child's gender. N = 12,476

| Impact
Among the 12,476 children, 3.1% of the mothers received antibiotics during pregnancy at 7-27 days before birth, and the excess number of asthma cases due to this exposure was estimated as 16 cases (33 cases in the target population, 25,534, PAR = 0.1%, PAF = 1.2%) (scenario 1 in Table S7). 70% of children received antibiotics at 0-24 months old, and the excess number of cases was 569 cases (1,164 cases in 25,534, PAR = 4.6%, PAF = 43.0%) (scenario 2).

| DISCUSS ION
The two exposure variables, maternal antibiotics prescriptions at 7-27 days before the child's birth and prescriptions for the child at 0-24 months, were associated with an increased risk of asthma, even after assessment for different biases: adjustment for relevant Evidence of dose-response has also been reported. 10 In previous studies, similar associations were observed for antibiotic exposure during, before and after pregnancy, 8,10 and the association decreased with sibling analysis. 11 These findings are in conflict with the causal hypothesis. However, in our study population, neither antibiotics given to the mother before pregnancy was associated with asthma nor did the siblings' analysis decrease or change the direction of the association. Therefore, in our study population the results were not conflicting.
The fact that the association was not observed before pregnancy gives support to specificity of the exposure during pregnancy. 33 Causal mechanisms have been suggested as biologically plausible: prenatal exposure to antibiotics by the mother and in early infancy could change the gut microbiome of the child and ultimately can lead to asthma. 5 Alternatively, but not mutually exclusive, antibiotics can change the maternal microbiota, which are transmitted to the child and could lead to changes in the child's immune response. 34 The diversity of the gut and lung microbiota of children would be disrupted, and reduced diversity and increased proportions of some specific pathogenic species have been associated with the development and severity of asthma. 5,6,35,36 A similar process would be involved in the higher risk of asthma among children born by caesarean section, 37 and antibiotics and obesity in children. 38 Therefore, there is evidence of this causal mechanism from other events. Heterogeneity of ef- This study had a well-defined target population, which is important to make causal effects meaningful. 40 Instead of trying to address our causal question based simply on a single regression model, we used triangulation. 13 The variables representing maternal antibiotic exposure during pregnancy were defined after looking at the data, which could be considered a reason for the cautious interpretation of their associations with asthma.
Unmeasured confounding remains a possibility, and it is a potential limitation in an observational study. 41 Relevant data such as on indoor environmental factors and breastfeeding were only available for sub-group of the study population, and thus, we conducted a bias analysis, but it does not replace an adjustment with the variables. If the magnitude of association is high, it is less likely to be explained by unmeasured confounders. 42 We took the E-value as an indicator for the magnitude of association for this hypothetical confounding bias, by assuming that there would be one confounder that was not adjusted for, a strong assumption. 43 Unmeasured confounders with magnitude of association with asthma could be higher than 2.0 44 ; therefore, confounding by an unmeasured confounder is plausible.
Evidence supporting a causal association is needed before proposing public health interventions and changes in clinical practice.
The decision-making process in public health ultimately is a subjective judgement considering our expectations of the benefits and harms based on evidence from different sources and including our current knowledge. 45,46 In our study and target population, the attributable excess number of cases would be small for prenatal exposure at 7-27 days before the birth, but this would be much higher for postnatal exposure in early childhood, supporting the potential benefits if exposure to antibiotics is reduced. As a caveat, the interpretation of the excess number attributable to the exposure depends on whether the association is causal and whether elimination of the exposure has no effect on the distribution of other risk factors. 47 Furthermore, the excess number attributable to the exposure is not equal to the number of cases caused by the exposure. 48 Evidence of any exposure-disease association depends also on the frequency of the different causes, and asthma results from the combination of several factors, which may not have the same distributions in different populations. Therefore, evidence of causation from a specific population is not necessarily generalizable to another population, consistency does not necessarily support causation, 40 and inconsistency in different studies is not necessarily due to methodological issues. If there is good evidence of a causal association, whether the magnitude of the effect would justify a change in recommendations and practice should be contextualized for different populations, considering the trade-offs between the potential harms and benefits, and feasibility of reducing the consumption of antibiotics in infancy. For example, the decision to administer antibiotic prophylaxis to mothers during caesarean before the surgical cut to prevent surgical site infection should be balanced with the potential of longterm harm to their children's health. 49 This more pragmatic approach should be considered, instead of trying to generalize the findings or wait for ultimate proof of a causal association or consistency of results from different settings. If, in a specific population, there is fair evidence of a significant effect from a public health perspective, and it is feasible to reduce the use of antibiotics, an intervention may be justified. Furthermore, these findings support the policy that seeks to minimize the usage of antibiotics in children, 50 given the potential considerable numbers if our results were extrapolated to the UK population. In our target population, it would also be prudent to be judicious in the use of antibiotics in the last month of pregnancy.

CO N FLI C T O F I NTE R E S T
None.

AUTH O R CO NTR I B UTI O N S
SSC and LP conceived this investigation and wrote the manuscript.
SSC executed analyses. All authors critically revised the manuscript for intellectual content, interpreted study findings, and read and approved the final manuscript.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available on request from the corresponding author (https://borni nbrad ford.nhs. uk/resea rch/how-to-acces s-data/). The data are not publicly available due to privacy or ethical restrictions.