Reply

We thank Sethi and Hanif^{Reference Sethi and Hanif1} for their thoughtful comments on our recent publication in the British Journal of Psychiatry.^{Reference Green, Wrobel, Todd, Marx, Berk and Lotfaliany2} We agree with many of the points they raise.

Sethi and Hanif make recommendations pertaining to extant field limitations, including the high statistical heterogeneity (I ² > 75%) impacting all five of the non-sibling-controlled meta-analyses we reviewed. We agree that subgroup analyses would have helped assess the impact of such heterogeneity, but unfortunately the data constrained this. For example, only 13 of the 30 included studies stratified results by antibiotic type and, even in these, the reporting was inconsistent and no studies used sibling control groups. Moreover, no studies presented data on dosage effects, and only five included data relating to the number of antibiotic exposures. Our efforts to perform antibiotic subgroup analyses and the limitations of these data are described in the paper (see the ‘Results’ section).

As noted by Sethi and Hanif, data were sufficient to perform sibling subgroup analyses for childhood antibiotic exposure and both attention-deficit hyperactivity disorder and autism spectrum disorder (ASD), both of which presented non-significant statistical heterogeneity (I ² < 50%). Subgroup analyses were also performed by study type (i.e. cohort versus case control), which did not appear to significantly impact results. We suspect that the sibling subgroup analyses may explain some of the heterogeneity, while also noting that these analyses were conducted on different data-sets.

We agree with Sethi and Hanif that future studies should make efforts to include sibling-controlled antibiotic subtype data, stratified by specific antibiotic type rather than broad groupings. We also recommend sibling-controlled stratification of data by number of antibiotic exposures, type of childhood or maternal infection and severity data (e.g. presence, duration and extent of febrile episodes and presence of hospitalisation).

Sethi and Hanaf also point to potential bias arising from reliance on administrative data and retrospective study designs, limitations that we noted in our ‘Discussion’ section. We agree that further large cohort studies will help address some of these biases, particularly through prospective cohort studies, which combine data derived from regular participant questionnaires or interviews, paired with administrative data (e.g. hospital records, prescription data and other available national data). This is likely to reduce gaps in data, as well as allowing for sufficient follow-up periods.

Sethi and Hanif underscore the very low certainty of evidence across six of the seven primary meta-analyses. We suggest that this relates to significant statistical and methodological heterogeneity, lack of adjustment for key confounders (particularly sibling controls), inadequate lengths of follow-up and gaps in data. One sibling-controlled meta-analysis was rated as low certainty, but this must be viewed in the context of the Grading of Recommendations Assessment, Development and Evaluation (GRADE) certainty assessment system, where ‘low certainty’ is usually the highest level of certainty possible for observational data,^{Reference Yousefifard and Shafiee3} with a few uncommon exceptions (e.g. significant dose response, large magnitude of effect and substantial adjustment). Moderate or high certainty ratings are mostly reserved for randomised controlled trial data only. As such, a rating of low certainty for a meta-analysis is the highest certainty usually achievable for the observational data that were included in the analysis. A randomised controlled trial assessing whether early antibiotic exposure is associated with increased neurocognitive outcomes would be neither feasible nor ethical, because it would involve either withholding clinically indicated antibiotics in pregnancy or the first year or two of life, or prescribing these without indication. Hence, our recommendation to elevate the data to the highest level of certainty usually available for observational data (i.e. low certainty) is to ensure the use of sibling-controlled data, as well as that of well-designed prospective studies to address the sources of bias, and the data gaps outlined above. Long follow-up periods are also recommended and, in the case of neuropsychiatric outcomes, this follow-up period would ideally comprise at least 25 years, by which time the onset of most neuropsychiatric disorders would be expected.^{Reference Solmi, Radua, Olivola, Croce, Soardo and Salazar de Pablo4}

Sethi and Hanif also recommend further exploration of confounding by indication. Although we attempted subgroup analyses by childhood or maternal infection, this was limited by a lack of data. With respect to infection type, only five included studies used childhood or maternal infection as covariates in their analyses, and detailed data were not provided. A number of studies presented detailed data regarding childhood or maternal infection, but they reported these factors separately from antibiotic exposure; in other words, it was not reported whether those with infections or fevers were also exposed to antibiotics. These data are therefore unsuitable for determining whether infection factors or antibiotics were related to the likelihood of neuropsychiatric outcomes. In saying this, Atladóttir et al reported on a large, prospectively designed Danish cohort study (N = 96 736), which combined the use of interview data with health records reporting on maternal infection factors, including type of infection, presence of fever, duration of fever, highest fever (> or <38.5 °C), timing of infection or febrile episode by trimester and subsequent likelihood of ASD in offspring.^{Reference Atladóttir, Henriksen, Schendel and Parner5} That study reported no evident association between any of these factors and subsequent likelihood of ASD in offspring. Considering the findings of this single study, it would appear that in utero exposure to maternal infection and fever may not represent a significant risk factor for ASD (while noting that other neuropsychiatric outcomes were not reported in this study), weakening the argument for confounding by indication.

Finally, Sethi and Hanif^{Reference Sethi and Hanif1} recommend a discussion of strategies that may protect against unnecessary antibiotic exposure. This is an interesting and important idea worthy of further exploration, and indeed the field of antibiotic stewardship and the reduction of inappropriate use is an active area of public health focus. The focus of our review was to explore data relating to risks associated with early life antibiotic exposure and later neurocognitive outcomes; to this end we concluded that there is no strong evidence of such an association.

This discussion hopefully will guide the design of further studies in this field that may address some of these unanswered questions.