The study design involves several retrospective cohort studies using linked administrative health and social data from Statistics New Zealand’s Integrated Data Infrastructure (IDI) [28]. Antibiotic use and T1D, ADHD, and IBD are defined as described below. To date, data for all children born in New Zealand between October 2005 and December 2010 (N=334,204) and their mothers have been extracted from the Department of Internal Affairs birth data in the IDI. Children’s antibiotic use has been defined for 4 time periods (at pregnancy, at ≤1 year, at ≤2 years, and at ≤5 years). The development of T1D, ADHD, and IBD (which consists of Crohn disease and ulcerative colitis) has been measured from the end of the antibiotic use periods until death, emigration, or the end of the study in 2021, whichever came first, accumulating approximately 3,000,000 person-years. Children who emigrated overseas or died before the end of the antibiotic use period have been excluded from the analysis, as they cannot be followed up for the occurrence of these chronic childhood conditions. At the end of follow-up, children had reached the age of 11-16 years.

The IDI is a database of deidentified administrative and survey data about people and households in New Zealand [28]. It includes data about health, education, income, social support payments, migration, and other life events, which can be linked at the individual level. The IDI provides a longitudinal record of events and is a growing resource. As of September 2018, the IDI holds over 166 billion pieces of information from more than 14 organizations [28,29]. Table 1 lists the datasets that are being used for this study with a brief description of the data and the variables extracted from these datasets.

Antibiotic exposures in utero and for the first 5 years of life are identified for all cohort members born between October 2005 to December 2010 from pharmaceutical data. Dispensing dates and dose and number of purchases are identified, and each antibiotic prescription is categorized by (1) class, according to the Anatomical Therapeutic Chemical classification J01 “Antibiotics for systemic use” (eg, penicillins, cephalosporins, sulfonamides); (2) spectrum, that is, broad or narrow; and (3) whether antibiotics target Gram-positive or Gram-negative bacteria or both. These categorizations of individual antibiotics are provided in Table 2 [31].

The selected health outcomes of the study population are determined through linkage with the following data collections: (1) hospital discharges, (2) pharmaceutical data, (3) nonadmitted patient collection, and (4) laboratory claims, for the period starting from birth or end of antibiotic exposure period of each child to the end of 2021. The specific case definitions, including the International Classification of Diseases, Tenth Revision codes corresponding to the health outcomes under consideration, are provided in Table S1 of Multimedia Appendix 1 [32-35]. To ascertain the prevalence of T1D, we used 3 distinct algorithms to identify cases, facilitating a comprehensive comparison of results to ensure consistency. For IBD, the identification of cases, as outlined in Table S1 of Multimedia Appendix 1, will be subject to further validation against several cohorts of patients with IBD obtained from collaborating gastroenterologists. This validation will involve exploring various combinations of medications prescribed for IBD, including those listed in Table S1 of Multimedia Appendix 1. The analysis aims to provide insights into the diversity of medication regimens associated with IBD cases. Any refinements or enhancements to the algorithms as well as insights gained from the medication combination analysis will be documented and incorporated into the final analysis.

Fixed covariates/confounders that will be considered in the analyses include sex, ethnicity, deprivation index (based on mesh block) [36], birth weight, gestation, mode of delivery, rurality, and maternal age. Time-dependent covariates include hospitalization for infections and other chronic diseases and selected prescription medications (eg, paracetamol, antivirals, antifungals).

Linkage to border movements and mortality data are used to determine whether cohort members are still alive and are based in New Zealand. Those who have emigrated or died prior to their fifth birthday are excluded from the analyses; the follow-up time for those who died or emigrated after the fifth birthday is censored up to that point, which means that the event of interest or health outcome being investigated may not be observed for some individuals.

To date, the primary focus has been on data preparation that consisted of (1) constructing the cohort through data linkage; (2) identifying the antibiotic exposure variables for all cohort members; (3) identifying other variables, including confounders for all cohort members; and (4) identifying T1D, ADHD, and IBD cases within the cohort using the definitions described in Table S1 of Multimedia Appendix 1. The next stage involves analyses that focus on assessing associations between antibiotic use and the health outcomes described above. For this, we will use Cox proportional hazards regression, with attained age as the analysis time scale. As noted before, children have been followed up until the estimated date of diagnosis, emigration from New Zealand, death, or the end of the study period (December 31, 2021), whichever comes first.

For each health outcome, analyses will be conducted to measure associations with antibiotic use during specific early life periods (pregnancy, ≤1 year, ≤2 years, and ≤5 years, as well as combinations of these periods). Antibiotic use will be based on the number of prescriptions, which will enable the assessment of dose-response associations. In addition to considering all antibiotic classes combined, we will also conduct analyses where antibiotics will be grouped into different classes/categories (Table 2); this will provide insights into which specific groups of antibiotics may be most strongly associated with the 3 outcomes of interest. Analyses will be stratified by mode of delivery to assess whether associations may be different in different subgroups (effect modification) as has been shown for cesarean section births, with larger effect sizes shown for associations between antibiotic use and T1D for caesarean section births [20,37]. In addition to stratified analyses, we will also assess the role of potential confounders such as sex, prioritized ethnicity, deprivation index, and rurality by using multivariable analyses.

Nested case-control analyses will be conducted as an additional way to address potential bias and confounding. Controls will be matched to cases on year and month of birth, sex, ethnicity, and other potential confounding factors such as residence and deprivation. In addition, to address potential medical surveillance bias, matched controls that occur in the same data collections as the cases will be selected. Nested case-control analyses will also enable the evaluation of possible reverse causation (ie, the health outcome of interest resulting in the prescription of antibiotics rather than the other way around) by disregarding antibiotic use in the 6 months before diagnosis of the cases and the equivalent time point of the matched controls. Moreover, control for confounding by maternal factors will be achieved through within-mother analysis of disease-discordant pairs of siblings. Factors remaining constant between pregnancies could, for example, be the mother’s attitude toward antibiotic prescriptions as well as the general practitioner’s antibiotic prescription practices, which will influence the child’s exposure to antibiotics; other types of analyses can typically not adjust for this. Further, confounding by indication, where the reason for prescribing antibiotics may be linked to the development of chronic conditions, will only be addressed once primary care data become available, enhancing the rigor of the study.

Based on national and international data, we have assumed that at least half of the children will have been prescribed antibiotics within the first year of life (52% in a Finnish study, 15% for specific antibiotic classes [38]). Based on age-specific statistics of the Virtual Diabetes Register, which is an annually updated national register of all patients with diabetes mellitus from 2010 to 2015, we estimate that 400 cases of T1D can be identified within the cohort [39]. Furthermore, based on hospitalization data, as per the age-specific data, we estimate that at least 200 IBD cases (170 cases of Crohn disease and 30 cases of ulcerative colitis) can be included in the study based on hospitalization data. Finally, for ADHD, and as noted earlier, medication dispensing has doubled from 516 per 100,000 in 2007-2008 to 996 per 100,000 in 2016-2017 in New Zealand [9]. Although a breakdown by age group is not available, we estimate that at least 1000 cases can be identified in the cohort. This is a conservative estimate based on our experience with other IDI projects [34], and it is likely that this number is substantially higher (up to 3% of the study population). Thus, we assume that case sets will have a minimum size of 1000 for ADHD, 400 for T1D, and 200 for IBD. Hazard ratio (HR) estimates that are detectable with 80% power (P<.05, 2-sided) under different population size and exposure prevalence scenarios are summarized in Table 3. Assuming an exposure prevalence of 33%, this study has 80% power to detect an HR of 1.2 for ADHD, 1.4 for T1D, and 1.5 for IBD. Analyses of specific strata of the study population (eg, based on sex or ethnicity) will have sufficient study power to detect similar effect sizes (Table 3). Considering a lower exposure frequency of 10% (eg, for specific antibiotic classes), the study has 80% power to detect an HR of 1.3 for ADHD, 1.5 for T1D, and 1.7 for IBD. Assuming a further reduction in exposure frequency of 5% (eg, for specific antibiotics), the study has 80% power to detect an HR of 1.5 for ADHD, 1.7 for T1D, and 2.0 for IBD (Table 3).

This study was approved by Human Research Ethics Committee of the University of Otago (reference: HD21/053). Microdata access approval for the project was provided by Statistics New Zealand. To ensure confidentiality, all data are deidentified before access. Access to linked datasets is limited to authorized personnel, and data handling follows ethical guidelines. Sensitive data will be managed securely. Although the data within IDI are fully deidentified, there are several requirements that govern the use of IDI data that this study will adhere to. These are (1) statistical outputs can only be disseminated after outputs have been checked and approved by Statistics New Zealand, (2) the IDI confidentiality rules require the suppression of counts and associated results of analyses on samples smaller than 6, and (3) the random rounding of counts can be up or down to the next multiple of 3.