Methods

Study design and setting

The primary objective and first two secondary objectives were addressed by undertaking a controlled interrupted time series study.43 It comprised a cohort of children born in the UK between 2006 and 2018 and their mothers, whose health-care records were included on a primary care database [i.e. The Health Improvement Network (THIN) or the Clinical Practice Research Datalink (CPRD)] or a secondary care database [Hospital Episode Statistics (HES)].

The interrupted time series component of the study compared rates of diagnosis of asthma, eczema and other immune system-related diseases over the study period in children delivered by CS between time periods when the mother of each child was likely to receive antibiotics before skin incision and periods when it was likely that antibiotics would be administered after cord clamping.

Validity in this comparison of incidence rates is mostly threatened by temporal changes in the diagnosis and recording of health conditions and different exposures that can have an impact on the number of cases identified in routine data, rather than from case mix confounders, as indications for CS and CS incidence have changed little over the study period. As an example, patterns of asthma diagnosis have changed over time, which can, in part, be attributed to the revisions in the national asthma management guidance44 and the potentially conflicting compliance and prevalence issues faced in meeting asthma-specific Quality and Outcomes Framework (QOF) indicators introduced in 2004.45

Conducting an analysis dependent on adjustment for confounding factors was unlikely to succeed in controlling for these changes, because (1) it is uncertain what all the drivers of all these changes have been, (2) it is uncertain whether or not any covariates exist that accurately describe these changes without substantial missing data and (3) there are challenges in specification of a functional form for the relationship between these covariates and outcome. Instead, to account for such temporal changes, we included children born by VD as a control group, as these children would have been subject to all the same temporal changes as children delivered by CS, but would not have routinely received prophylactic antibiotics. Figure 1 shows that, although the national uptake of pre-incision antibiotic policy has increased over the study period, incidence rates of asthma decreased equally in children born by CS and those born by VD.

FIGURE 1 Illustration of the need to use VDs as the comparator group because of changing temporal patterns in outcome events unrelated to routine exposure to pre-incision antibiotics, using asthma as an example. (Rates after 2014 are further reduced because of < 5 years of follow-up and much lower incidence of asthma in very young children.) (Figure not reproduced here; see the original article.)

Our model first computed the ratio of the incidence of outcomes in those receiving CS compared with VD in each year and age group. The incidence of some outcomes, including asthma, is known to differ by delivery mode.46–48 It then compared these ratios between periods when mothers of children delivered by CS would have received antibiotics before skin incision and periods when mothers of children delivered by CS would have received antibiotics after cord clamping. This ‘ratio-of-ratios’ is an indicator of how the incidence of disease has increased (if > 1) with pre-incision antibiotics, stayed the same (if = 1) or decreased (if < 1), while adjusting for temporal changes through the matching of outcomes in children born vaginally. Subject to the assumption that the model of the difference between CS and VD rates is transferable across the time periods, differences between the observed and counterfactual CS event rates were interpreted as likely to be caused by the change in practice.

In the primary analysis, for primary care data, national policy uptake rates in the year of delivery were used as a basis for estimating a probability that each mother had received pre-incisional antibiotics and, therefore, the analysis actually used this probability as the explanatory variable, rather than a 0–1 variable. For secondary care data, we used each hospital’s response indicating the year of local policy implementation. This was obtained from our national survey of maternity care providers, which provided a 0–1 variable.

To address the final secondary objective, we compared the incidence of post-partum maternal infectious morbidity post intervention with pre-intervention rates to investigate if the effects of reduced morbidity that are shown in randomised controlled trials outside the UK9 that form the basis for the NICE guideline change7 can be replicated in the UK using routine health-care data. This analysis has been carried out for two reasons: (1) to measure the assumed impact using ‘real-world data’ to enable us to compare the benefits of pre-incisional antibiotics on maternal outcomes against any potential harms for the child and (2) to assure us that any lack of effect of the different prophylactic antibiotic policies was not due to the quality of the routine data. We calculated the crude estimates and, because of the observed overall changes in the incidence of different maternal post-partum infections recorded in routine data over time, we also undertook controlled analysis adjusting for delivery mode.

Population and eligibility criteria

To be eligible for inclusion in the study, records of all live-born children needed to meet the following criteria:

• The year of birth was between 2006 and 2018.
• It was possible to link the baby and their mother’s health-care record in primary care (THIN or CPRD) databases or a secondary care (HES) database.
• The delivery mode (CS or VD) could be determined based on recording in primary (THIN or CPRD) and/or secondary care (HES).

If children had missing delivery information, they were excluded. In the case of multiple births (such as twins), one of the children was included randomly to ensure independence of observations.49

The duration of follow-up of children born between 2006 and 2014 was until the age of 5 years. The follow-up of children born in 2015, 2016, 2017 and 2018 was limited to 4, 3, 2 and 1 years, respectively.

For primary care data, the exposure, outcomes and covariates were defined using the Read clinical code classification system. In HES, International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) was used for clinical diagnoses and the Office of Population Censuses and Surveys Classification of Surgical Operations and Procedures, 4th revision (OPCS-4) for procedures. The code lists are available online at URL: https://github.com/NIHR-Pre-incision-Antibiotics/ClinicalCodes (accessed 8 July 2021).

The process for Read code selection was as follows:

• Stage 1 – a list of key words and synonyms for each variable of interest was compiled.
• Stage 2 – the terms identified in stage 1 were used to search the description column of the Read code dictionary/database to identify the main chapter and appropriate level 3 stem code in the Read code dictionary.
• Stage 3 – stem codes were then used to identify other related Read codes, up to level 7, associated with the main stem codes. These Read codes were then selected and stored.
• Stage 4 – where available, for each outcome of interest, online Read code repositories and supplementary information from published articles that had used a similar set of Read codes were identified. These were then compared with our selected set of Read codes and further useful codes were added at this point.
• Stage 5 – with reference to the inclusion and exclusion criteria, the team identifying and selecting Read codes allocated weights to each Read code as follows: (1) definite (the team were sure that the Read code met the study inclusion criteria), (2) probable (the team felt that these codes could meet the inclusion criteria but needed further discussion), and (3) excluded (the study team felt that these codes definitely did not meet the inclusion criteria).
• Stage 6 – consensus was sought through consultation with study team researchers and public contributors, general practitioners (GPs) or specialty field experts on the final list of codes to be used, which were utilised for electronic data extraction.

Selection of ICD-10 and OPCS-4 codes to obtain admitted patient care data was generally straightforward. The validity of each coding algorithm was checked through triangulation of a number of sources. First, recently published studies based on routinely collected data were screened to check for consensus in the research community as to how to code specific reasons for admission. Second, checks of resources for clinical coders (the NHS coding clinic and the Terminology and Classifications Service in NHS Digital) were carried out. Third, technical appendices of outputs produced by the NHS and the Office for National Statistics were checked to see what coding definitions were used in reporting by public bodies. Last, to check for the specificity of coding (i.e. the extent to which fourth digits were commonly used or, if coding tended to default to ‘other’ or ‘not elsewhere classified’) individual analyses of English admitted patient care data were undertaken.

The delivery mode is recorded well in hospital records.2 We used OPCS-4 procedure codes, which are available for 98% of births in HES,50 to derive the mode of delivery and ICD-10 codes for the remaining deliveries. Our estimates showed that the delivery mode is also accurately recorded in primary care using Read codes (98% verified against hospital records). The recording, however, was incomplete (the delivery mode was known for 50% and 69% of children in THIN and CPRD data sets, respectively, with the higher proportion in CPRD due to the supplied data set being restricted to the records in the CPRD pregnancy register). For THIN and CPRD records not linked to HES, we used the Read codes to identify the mode of delivery, which increased the number of children in the sample.

Study outcomes

The primary outcomes in this study were the incidence of (1) asthma and (2) eczema. Asthma and eczema outcomes in the THIN–CPRD data set were defined using lists of relevant Read codes, which were created following the steps described above (see Population and eligibility criteria). Furthermore, Read code lists for both asthma and eczema have been previously validated in primary care.51,52 Therefore, existing validated lists were used as the basis for the outcome definition in this study and further codes were added following consultation with the wider research team, which included a GP and an allergy and immunology consultant. In line with the paper on the validation of eczema coding,51 in addition to a Read code for eczema diagnosis, at least two prescriptions on separate dates for any eczema-related treatment were also required to reflect the chronic nature of the condition. In the HES data set, asthma and eczema were defined if a ICD-10 code for asthma or eczema was used as the main primary diagnosis for the hospital admission.

The main analysis for these outcomes was carried out separately in the primary care data set and the secondary care (HES) data set (the latter analysis including only those hospitals for which the year of antibiotic prescribing policy change was known).

We also explored the relationship between antibiotic exposure and asthma and eczema severity in the primary care data set. Moderate/severe asthma was defined as having a Read code for asthma diagnosis and a prescription code for a corticosteroid inhaler and either a leukotriene receptor antagonist or a systemic steroid. As the severity of asthma is not reliably coded, a prescription for a leukotriene receptor antagonist or a systemic steroid is indicative of asthma that is uncontrolled on inhaled corticosteroids (i.e. moderate or severe asthma). Moderate/severe eczema was defined as having a Read code for eczema diagnosis and at least two prescriptions on separate dates, with at least one prescription for treating moderate or severe eczema (i.e. moderate, potent or very potent steroid, topical calcineurin, systemic therapy or phototherapy), as described in the NICE guideline for eczema.53

Secondary outcomes are listed in Table 1. We excluded immune deficiencies from the outcome list, as Read and ICD-10 codes did not allow us to differentiate between congenital and acquired immunodeficiencies. We also did not include pelvic abscess as a separate outcome in the analysis of secondary care data, as there was no specific ICD-10 code for this condition.

Covariates

To assess the validity of the model we used in controlling for temporal confounders, we explored changes in the case mix covariates to identify any differential changes over time in relation to the mode of birth. To identify relevant covariates, a literature review of reviews was conducted on the risk factors for the primary study outcomes (i.e. asthma and eczema) and postnatal maternal infections. MEDLINE and Google Scholar (Google Inc., Mountain View, CA, USA) were searched for articles written in English that reported the effect of any risk factor for the conditions mentioned. Identified covariates included maternal demographic characteristics,54 behaviour-related factors,54–58 pregnancy- and labour-related factors,54,57 antibiotic prescribing during pregnancy,56,59,60 maternal long-term allergy-related conditions,61–64 deprivation,54,55,58,65,66 child demographic and birth characteristics,54,55,58,59,62,64–67 breastfeeding,55,67 child health conditions55,61–65,67 and antibiotic prescribing in childhood55,60,67 (Table 2).

Data sources and mother–baby linkage

Electronic primary and secondary health-care records

THIN and CPRD

The Health Improvement Network and CPRD GOLD are two UK-wide primary care databases of anonymised patient records from general practices that use the same computer software (Vision, Cegedim Rx Ltd and In Practice Systems Ltd, London, UK), together covering over 10% of the UK population.68 In addition, a significant proportion of patients (about 30% in THIN and 60% in CPRD) have linked secondary care (HES) data. Both THIN and CPRD GOLD databases are broadly generalisable to the UK population in terms of patient demographics and health condition prevalence.69,70 The University of Birmingham (Birmingham, UK) holds a sublicence for the full THIN database. The CPRD GOLD data set was acquired from the CPRD centre of the Medicines and Healthcare products Regulatory Agency (MHRA) for use in this specific study.

Identification and removal of duplicate practices

The Health Improvement Network and CPRD GOLD databases have some overlap at practice level, with 37% and 46% unique practices, respectively, reported previously.68 Our sample size calculations showed that both primary care databases were required to provide sufficient power to answer the research questions. We adapted a previously published algorithm to identify practices common to both CPRD GOLD and THIN.71 Although the databases do not use the same identifiers for patients or practices, the overlapping practices can be identified and removed reliably using practice-related information, patient registration, demographic and medical record information.68,71

We extracted practice- and patient-level data to identify common/duplicate practices in THIN and CPRD. The following practice-level data were extracted: (1) the number of males and females born between 1905 and 1910, between 1910 and between 1915 and 1915 and 1920; (2) the number of males and females newly diagnosed with type 2 diabetes in the year 2010; and (3) the total number of patients who received at least one drug prescription in the year 2010. In total, this amounted to nine practice-level variables that we extracted from the THIN database and the same variables were provided to us on request from CPRD GOLD.

For patient-level data, we examined aggregated information from a sample of patients. To extract our patient-level variables, we calculated an eligible time period where both THIN and corresponding CPRD GOLD practices were contributing data to their respective data set. The eligible period was defined as the time between the ‘latest’ and ‘earliest’ set of dates for each database as follows: latest of (1) THIN Vision date, (2) THIN acceptable mortality reporting date, (3) THIN computerisation date or (4) CPRD GOLD ‘up to standard’ date (a quality indicator); and the earliest of (1) THIN collection date or (2) CPRD collection date. Defining the eligible time period was important, as the patient-level data we extracted for deduplication can vary outside this period. For example, there might be patients who were registered with THIN earlier, but were transferred out before the practice started contributing data to CPRD GOLD. If such patient-level data were extracted from THIN, then we would not be able to confirm their characteristics in CPRD GOLD, as their data were never present in it. Alternatively, we might have found similar patient-level characteristics by chance in candidate practices that are not duplicates.

For each THIN and CPRD GOLD practice pair, we used CPRD GOLD data and randomly selected 10 patients who were registered in the practice within the eligible period. We extracted their registration date, year and month of birth, sex, family identification number, first three consultation types, consultation dates and duration of consultations for matching. We created two new scores, (1) ‘patient point’ and (2) ‘consultation point’, and set both to zero. We looked in the corresponding THIN practice to find patients with the same information. If a patient was found in THIN with the same registration data, sex and date of birth information, then the patient point was incremented by 1 and, for the same patient, we incremented consultation points by 1 for each consultation information that was matched as equal in both databases.

After considering patient- and practice-level scores, we found 450 practices contributing to both THIN and CPRD. Among the duplicate practices, we retained the practice (from THIN or CPRD GOLD) that contributed the highest number of babies to the study.

Linkage of mothers and babies

Information on mothers and their babies can be linked in CPRD GOLD and THIN using the family identification number, delivery codes, mother’s registered or estimated date of delivery, child’s birth month, gestational age at delivery and, where available, the sex of the baby, with reported identification of the vast majority of mother–child pairs.72,73

Clinical Practice Research Datalink GOLD data held by the CPRD centre of the MHRA include a mother–baby link and, because of CPRD data minimisation requirements, we were given access to already linked mother–baby CPRD data. CPRD has linked mothers and children using several methods, including delivery-related information, information indicating pregnancy and pre- and postnatal information, which is explained in detail elsewhere.74

To link mothers and babies in the THIN–HES and THIN primary care database, for each mother and for each delivery record she had within the study period, we looked in the associated database record to see if there was a baby with the same family number as that of the mother and linked them to the mother if the following conditions were true:

• The baby’s year and month of birth were the same as the mother’s year and month of delivery.
• The baby was registered with the practice within a year of birth.

In addition, after linkage, if the sex of the newborn was provided in the mother’s HES delivery record, then we cross-checked it with that in the primary care data to ensure accuracy. In addition, where a mother had more than one baby matching (e.g. twins), we randomly chose one baby.

We excluded delivery records and baby records that did not meet our matching criteria.

Defining index and exit date for patients in THIN and CPRD

For each of the records containing delivery information in HES, we checked if patient data existed in the associated THIN/CPRD practice records. If found, for each mother–baby pair, we set the index date as the date of delivery, defined as the OPCS-4 date or the hospital episode end date if the OPCS-4 date was missing. For mother and baby pairs for whom HES data did not exist, we assigned the date of birth of the baby as the index date for the baby and the date of delivery for the mother. These two dates can differ because only the month of birth is given in the baby’s records. At the same time, delivery date may reflect the date the information was entered in primary care records, as opposed to the true delivery date. We also defined the patient start date separately for the mother and baby as the latest of the date the quality of data was verified, study start date (1 January 2006) or registration date with the practice. For the mother, the day when they reached age 13 years was also included in determining the patient start date. We also defined the patient end date separately for the mother and baby, to determine the end of maximum follow-up for each patient, as the earliest of the death date, the date the patient left the general practice, 31 December 2018 or the last practice data collection date. If the patient was a mother, we added 6 weeks to the date of delivery in the list when calculating the patient end date, as maternal outcomes were limited to the post-partum period. If the patient was a baby, we included their date of birth plus 5 years in the list when calculating the patient end date. Once we defined these three key dates (i.e. the index date, patient start date and patient end date), we removed any patients whose start date was after their end date (e.g. error in recording of dates), whose index date was not between the start and end dates (e.g. if the HES data had information after the patient left the practice) or whose age at index date was not between 13 and 50 years (mother only).

HES data covering the whole of England

A mother–baby linked database was created using anonymised HES data collected for all NHS hospital admissions in England75 to allow the investigation of more severe outcomes of interest resulting in hospital admissions.

We requested HES data from NHS Digital on all delivery events occurring in English hospitals from financial years 2005/6 to 2017/18. We also asked for any admissions data relating to patients under the age of 5 years with the same encrypted HES identifier as a baby delivered in any of these delivery events and, for the mothers who underwent a delivery episode, any admissions either 1 month before delivery or 6 weeks after the start of the delivery episode. The application process was protracted. We completed the application in August 2018, but received data for the financial years 2005/6 to 2016/17 in April 2019. A data update for 2017/18 and 2018/19 was received in March 2020, but because of the project time restrictions, we were able to add only the latest admission data for the mother–baby pairs that we had already linked in the database using records for 2005/6 to 2016/17.

There is no shared identifier in the UK to link maternal and baby records in HES. We adapted a strategy that had already been validated by another recent study76 using deterministic and probabilistic linkage algorithms, allowing linkage of up to 98% of baby and mother secondary care records in England and still remaining nationally representative for the main birth characteristics (e.g. gestational age, birthweight, sex and maternal age). The processing workflow is outlined below.

We identified all of the delivery episodes and the spells to which they belonged. Episodes in which maternity care is delivered have a specific set of fields that contain critical data about the delivery pathway, as well as the usual data set of diagnostic and procedure fields that occur in any other kind of hospital admission episode. These are called the ‘maternity tail’ and include variables such as the weight of the baby, the maternal age, the pregnancy term in weeks and other clinical details. There is not an automatic linking field between babies and mothers in the anonymised data available in HES and so a simple linkage allowing details of a birth to be linked to future admissions involving the child is not possible. However, a baby born in hospital is technically admitted to the hospital and has an admission episode, even if they require no clinical treatment. This episode should contain all of the same data as the ‘maternity tail’ in the mother’s record. In practice, there are often a number of missing data items in the mother’s record and, more often, in the baby’s record. It is possible, however, to create an algorithm that evaluates the likelihood of a particular baby’s episode being linked to a particular mother’s birth event. For example, if all of the maternity fields match, if the baby’s admission date is the same day as the delivery and if the registered GP, the hospital and the postal district prefix of the postcode of residence are the same, then the chance that this delivery did not relate to this baby would be very low. If most of the data matched, but there were some missing items and no data items that contradicted each other, then the chances of mismatch would also be very small. If there was only one likely baby match, we could assume a deterministic linkage. The chances of an erroneous match would increase a little with each missing pair of data items, and it is possible that a situation could occur in which more than one baby could be matched to the birth episode could occur. In these cases, we had to apply probabilistic matching based on a matching threshold.

The first major data manipulation task was identifying all of the delivery episodes that we were going to match. This is not a straightforward task, as episodes that contain data relating to deliveries are sometimes not specifically flagged as a delivery episode. In addition, some episodes that are clearly flagged as delivery episodes do not contain all of the data needed to obtain a match. This is because there can be a number of episodes in a spell in hospital that include a delivery and, despite coding and clerking standards, some data that belong in one episode can be spread across more than one. We identified any episode that met the criteria used by Harron et al.,76 namely that the episode should contain a code indicating the delivery of a live infant, a procedure code for a delivery procedure or have two or more ‘maternity tail’ fields that are completed with a valid (i.e. not null or system missing) value. When these episodes had been identified, we used the unique spell identifier to extract all episodes in each spell containing a delivery episode.

Cleaning was carried out to remove duplicate records, episode durations with implausible date ranges and births where there was an implausible interval between two or more birth events. This happens rarely and is usually because a mother is misidentified in the data set and is given the same NHS identifier (ID) as someone else. If both these women had given birth within, for example, 6 weeks of each other, then it appears that one woman has given birth twice, with an implausible duration between the two deliveries. In these cases, records for both women were censored.

Mother–baby linkage

The Harron et al.76 methodology was based on a complex synthetic data set where iterations of data removal were undertaken in key mother and baby variables and probabilities of mismatches were calculated given that different variable or combinations of variable were missing on one side or the other of the match. Through this process, a series of probabilities for different matching scenarios were calculated, as well as a series of weights to determine the likelihood of a valid match, given that a specific variable was matched or not matched. The process was based on 1 year of data for 2012/13. Data recording in HES was improving year on year in the period of our study. It is likely that the data years before 2012/13 would have had more missingness than the Harron et al.76 data and the years afterwards. As Harron et al.’s76 weights were calculated roughly in the middle of our period of study, we felt that they could be applied reliably to our data. We decided that rather than replicating the original synthetic data set creation, which would have been outside the resource capabilities of this study as originally designed, we would repeat Harron et al.’s76 matching methods, but use the matching weights published in their paper.76

During key stages in the work, we had clarification from Harron et al.76 on some of the aspects of their method that were not covered in detail in the publication. The linkage was carried out in two stages. Where data could be linked using a simple rules-based approach following the original study, we followed the original authors in referring to this as deterministic linkage. Any delivery events that could not be linked deterministically were linked probabilistically.

Deterministic (rules-based) linkage

Mother and baby records were treated as matching if records matched on general practice, maternal age, birthweight, gestation, birth order and sex. The matching could allow for missing data, but not for any contradictory data on either side of the match. If three or more of the variables matched exactly, and no variables were mismatched, the match was assumed to be valid.

Probabilistic linkage

Any delivery records that were unmatched deterministically were compared with baby records on 21 separate variables. Each one had a positive match weight where there was a match and a negative weight where there was a mismatch. These were applied to a matching matrix and the weights summed to derive a total weight for each candidate pair. The weights were determined from the synthetic data analysis produced by Harron et al.76 and were available in their published paper. Their team also used the same threshold (a match score of 20) to decide whether or not the match was probable.

Linkage to survey data

The final stage involved matching national survey data on hospital prophylactic antibiotic administration policies (see National survey data for more details on the survey) to births. This involved identifying the NHS organisation code for each site about which the respondent was answering. This was carried out by cross-checking all of the hospital names with the NHS Coding and Classification Service records. In addition, each site had to be checked to see whether or not its site code had changed during the period of study; this happens when the trust responsible for the site changes. This also involved manual searches on a database of successor and predecessor organisations held by the NHS Coding and Classification Service. Not all deliveries were matched because not all sites replied to the questionnaire. There were some cases where individual units did not collect their site codes with sufficient specificity. This can happen when one trust manages a large number of smaller sites. There is a highly specific site code capable of attributing activity to an individual site; however, some organisations do not use the code appropriately, choosing to apply a more generic code with a trust prefix but no suffix (i.e. the part of the codes that identifies the site specifically). We linked only those questionnaire data where a valid and site-specific code was captured in HES and was the same as a site code attributed to a hospital that had responded to the survey.

National survey data

National survey of hospital prophylactic antibiotic administration policies

Information about which prophylactic antibiotics are given and the timing of their administration is not available in HES. Therefore, we obtained the year when the policy to administer pre-incision antibiotics was introduced in each hospital from a survey of maternity care providers.

A national survey of the clinical directors for UK maternity units was completed to investigate (1) to what extent the recommendation in the NICE Clinical Guideline 1327 regarding the timing of prophylactic antibiotic administration for CS has been adopted across the country, (2) when the updated national guidance was implemented and (3) the types of pre-incision antibiotics used in practice.

The survey was developed using a secure, web-based software platform REDCap (Research Electronic Data Capture) designed to support data capture for research studies (version 7, Vanderbilt University, Nashville, TN, USA)77 and a link to the survey was sent to all clinical directors of maternity units in 2017, with follow-up of non-responders undertaken by e-mail, telephone and at national professional conferences. During the follow-up, an option to complete the survey in Microsoft Word (Microsoft Corporation, Redmond, WA, USA) was also offered. A copy of the survey questionnaire is included in Report Supplementary Material 1. If the respondents did not specify the exact year of the policy change, but had entered a free-text comment saying, for example, a year before which the policy was changed (e.g. ‘before 2009’) then this year was used in the analysis as the year of the policy change.

National maternity care surveys

Information on breastfeeding in health-care data is incomplete and we, therefore, investigated the trends in this case mix variable by the mode of delivery using data from the national maternity surveys, with the survey methodology described in detail elsewhere.78,79 Briefly, the Care Quality Commission undertook nationwide surveys of women who had a live birth and received care from NHS trusts in 2007, 2010, 2013, 2015 and 2017 in England, asking about their experiences during labour and birth, including the type of delivery and newborn feeding in the first few days after birth.79 Care Quality Commission survey data were available from the UK Data Service.80 National surveys were also carried out in 2006, 2010 and 2014 by the National Perinatal Epidemiology Unit (Oxford, UK) using a random sample from live birth registrations in England. The questionnaires were posted to women at 3 months after delivery and asked questions on antenatal and postnatal care and health, including questions on the mode of delivery and infant feeding at the time of the survey (e.g. breast milk only, both breast milk and formula milk, and formula milk only).78

Estimated sample size and statistical power

THIN–CPRD database sample size estimates

To estimate the number of CSs and VDs in primary care over our study period, we first calculated the completeness of recording in THIN database using THIN data linked to secondary care records in 2010. This was extrapolated over all years in our study (2006–18), taking into consideration the changes in the numbers of births nationally over time. We assumed that the overall size of the THIN and CPRD databases, the proportions with linked secondary care data and the reported overlap between THIN and CPRD databases remained constant over the study period. We also assumed that 86% of mothers and babies could be linked. The number of births we expected to have data on over the study period was 777,389 (206,615 children born by CS and 570,774 children born by VD).

Our estimates of asthma incidence in 0- to 5-year-old children in the UK using the THIN database matched the previously published figures.38 In 2014, the incidence ranged from 0.8 and 6.5 per 1000 person-years in those aged 0–1 and 1–2 years, respectively, to 12.1 and 11.4 in those aged 3–4 and 4–5 years, respectively. Based on our estimates of the incidence of asthma in each age group in each year from 2006 onwards, and assuming a 5% annual loss of follow-up due to patients changing their general practice, Table 3 shows the estimated number of children available for follow-up depending on their year of birth, and the number of new asthma cases in each of the study years by the mode of delivery.

HES database sample size estimates

Using the HES secondary care database available for England and assuming 95% linkage between mothers and babies (the linkage estimate varied between 91% in the earlier years and 98% in 2012/13),76 we estimated that we would obtain data from 8,043,600 children with a known mode of delivery across 2006–18 (2,070,500 children born by CS and 5,973,100 born by VD).

Annually, there are around 10,000 hospital admissions with asthma as the primary diagnosis in children aged 0–4 years.81 Assuming a re-admission rate during the follow-up of 50% (our estimates using HES indicate that, in the later years of the cohort, the proportion of children re-admitted was as low as 29–39%), we calculated the numbers of children who were admitted to hospital for the first time with asthma as the primary diagnosis in each year of follow-up (Table 4).

Estimates of statistical power

To estimate statistical power, and to assess the impact of misclassification of antibiotic exposure on estimates of increase in risk with pre-incision prophylactic antibiotics for CS, we simulated the study (1) based on our estimates using the THIN database regarding the number of children for whom maternal data could be linked and the asthma diagnosis rates in each year group between 0 and 5 years, and (2) using HES data based on our estimates of children with linked maternal data and rates of children newly hospitalised for asthma based on HES statistics.

In each simulation, we created a data set for the whole study, with 13 birth cohorts from 2006 to 2018 and follow-up included across the first 5 years of life (curtailed at the end of 2018). Each birth was classified by the delivery mode and, for those children born by CS, whether or not prophylactic antibiotics were administered prior to skin incision (generated randomly using a binomial random number generator using an underlying probability identified in the national survey of hospital prophylactic antibiotic administration policies during that year of birth).49

We randomly simulated outcome events according to the year-age event rates using a binomial random number generator, with increased rates in all those born by CS and further increases in those who had pre-incision antibiotics (Table 5). A ratio of incidence ratios of 1.20 was used for increased risk of asthma with CS.47 Further increases (ratio of incidence ratios from 1.10 to 1.20, increasing in steps of 0.02) were used for the increase with pre-incision antibiotics rather than antibiotics after cord clamping.

Simulations were repeated 1000 times and statistical power estimated by noting the proportion of simulations for which the lower limit of the 95% CI for the variable indicating if antibiotics were administered prior to skin incision was greater than a ratio of incidence ratios of 1. We also assessed attenuation bias created by misclassification.

The model that was fitted to analyse the simulation data included a trend term for the probability of receiving antibiotics before skin incision, where the value for 2006–9 = 0, 2010 = 0.2, 2011 = 0.4, 2012 = 0.6, 2013 = 0.8 and 2014–18 = 1 (with a zero value for those who were born by VD). Although in the analysis we used values obtained from the national survey on hospital prophylactic antibiotic administration policies, as the simulations used the national survey to allocate women to a treatment, we would have overestimated power if we had used exactly the same function in the analysis model.

For outcomes in primary care, we estimated that we would have 80% power of detecting a ratio of incidence ratios of 1.16 and > 90% power of detecting ratio of incidence ratios of 1.18, being able to estimate them with a maximum of 15% underestimation from misclassification.

For outcomes recorded in secondary care, in the full HES data set, we estimated that we would have > 80% power to detect a ratio of incidence ratios of 1.10 and 90% power to detect a ratio of incidence ratios of 1.12, with similar underestimation rates due to misclassification.

This relates to detecting an increase of seven additional new cases of asthma per 1000 children aged 0–4 years or an increase of 4% (1400/35,000) of all new asthma cases in 0- to 4-year-olds annually in the UK. This also relates to detecting an increase of one additional new case of asthma requiring hospital admission per 1000 children aged 0–4 years or an increase of 3% (200/7000) of all newly hospitalised asthma cases in 0- to 4-year-olds annually.

We also estimated that this study would be adequately powered to detect differences in some of the other outcomes of interest, as their incidence is higher than asthma incidence in children, including:

• GP-diagnosed eczema (with incidence rates of 2.5 and 6.6 per 1000 person-years reported in children aged 0–1 and 1–2 years, respectively36,37)
• maternal post partum infectious morbidity, including caesarean wound infection (with the incidence ranging from 1.4 to 8.2 per 100 births82).

Statistical methods and analyses

The statistical analysis methods for this project can be broken down into the methods used to analyse the primary care data set (THIN–CPRD) and the methods used to analyse the secondary care data set (HES). We describe these methods separately for each data set.

Primary care data (THIN–CPRD)

Summarising data in primary care

The primary outcomes (of asthma and eczema) are summarised in the THIN–CPRD data set using the number of children diagnosed with asthma or eczema and the length of time at risk. Each child is classified as either diagnosed with the outcome or not, using the methods of diagnosing asthma and eczema described earlier (see Study outcomes). For each outcome, the length of time at risk is calculated as the time from birth to the earliest of the child’s fifth birthday, the date the child is lost to follow-up (e.g. the child moves practice or the practice is not included in THIN–CPRD), the study end date (31 December 2018) or the date at which the child is diagnosed with the outcome. We present the rate of the outcome per 1000 person-years, which is calculated as the ratio of the number of events and the exposure time multiplied by 1000. We do not present results stratified by prophylactic antibiotic status, as this was not possible (see Analysis methods used).

All dichotomous secondary outcomes are summarised in the same way as the primary outcomes. Secondary outcomes that are of a count form, such as the number of primary care consultations in the first 12 months, are summarised using the total number of events and the length of time at risk. The length of time at risk is calculated as the time from birth to the earliest of the date the child is lost to follow-up or the date at which the outcome-specific length of time has passed (e.g. 12 months). The rate of the outcome is presented per 1000 person-years.

Maternal outcomes are summarised using the total number of events and the percentage of total births to which this corresponds. All maternal outcomes in primary care were dichotomous. For the composite maternal infectious morbidity, we categorised mothers as having either ‘none’ or ‘one or more’ of the individual composite outcomes (i.e. wound infection, endometritis/endomyometritis, sepsis and pelvic abscess) recorded in primary care data. Results are presented stratified by delivery mode (VD or CS).

Analysis methods used

To analyse the primary outcomes (asthma and eczema), we calculated for each child’s year of life whether or not a diagnosis of the outcome was made and the exposure time for that particular year of life. For example, a child diagnosed with asthma aged 18 months would have zero events and 12 months’ exposure during their first year of life, and one event and 6 months’ exposure during their second year of life. This allowed us to explore cases by year of birth and year of life. To estimate the impact of antibiotic timing (pre-incision or post-cord clamping prophylactic antibiotics), we fitted a Poisson regression model to the data, with an offset for exposure time and robust standard errors to allow for misspecification. The model included terms for year of birth and age (year of life), and their interaction. Two further terms were also included in the model: (1) mode of birth (CS or VD’) to allow for differential risk of the outcome in those receiving CS compared with VD’ and (2) antibiotic timing’ to allow for differential risk in those receiving pre-incision rather than post-cord clamping prophylactic antibiotics. For all CS births, as the exact timing of antibiotic provision was not known, we used the probability of antibiotic use before skin incision derived from the national survey. This coefficient provides an estimate of the effect of the policy change in antibiotic timing, adjusting for misclassification. As the policy change was applicable for CS births only, all VD births were classified as pre-policy change. We present the incidence rate ratio (IRR) and the 95% CI for the antibiotic timing variable, comparing the rate of events in those given prophylactic antibiotics pre incision with the rate of events in those given prophylactic antibiotics post cord clamping.

All dichotomous secondary outcomes are analysed in the same manner as the primary outcomes. For outcomes with a small number of events (< 200), we present the summary statistics only. For count outcomes, a Poisson model without robust standard errors was fitted to the data. All other terms in the model were the same as those used for the primary outcomes.

Sensitivity analyses

To assess the robustness of the results, we conducted several sensitivity analyses. Sensitivity analyses were performed for the two co-primary outcomes (asthma and eczema). Some secondary outcomes were included in the sensitivity analyses to check the robustness of the primary analysis, particularly where results suggested a significant association with the timing of antibiotic provision and the association was unexpected.

In the primary care data set, we conducted the following sensitivity analyses:

• outcome definition
• definition of timing of the prophylactic antibiotic policy change
• data recording quality
• random effect.

The rationale for these sensitivity analyses and the statistical methods used for each of them are described in more detail below.

Outcome definition

The main outcome definition for asthma in primary care was based on relevant Read codes and the main definition of eczema was based on relevant Read codes in combination with two eczema-related prescriptions on separate dates, as described earlier. We considered the impact of prophylactic antibiotic timing on severity of asthma and eczema based on prescribing of medications indicative of moderate or severe asthma and eczema. In addition, we also considered several alternative definitions of asthma and eczema, including:

• asthma diagnosed in children aged 3–5 years (as it is more difficult to distinguish asthma from other more common diagnoses, such as bronchiolitis, in very young children)
• asthma definition based on Read codes used in the QOF
• eczema diagnosis based on Read codes in children aged 1–5 years and at least two prescriptions within 90 days before or 365 days after the eczema Read code record
• eczema diagnosis based on Read codes only.

Definition of timing of the prophylactic antibiotic policy change

For the primary analysis, we used a probability that the prophylactic antibiotics were given pre incision using data from the national survey. To test the sensitivity of this method, we restricted the data set to the births from 2006 to 2010 and from 2013 to 2018. This was carried out to compare births in the years before the change in the NICE guidelines (i.e. 2006–10) with the years in which > 50% of hospitals had introduced the policy (i.e. 2013–18). Children born by CS were dichotomised as having prophylactic antibiotics post cord clamping (children born 2006–10) or prophylactic antibiotics pre incision (children born 2013–18). A Poisson model was fitted to the data, with an offset for exposure time and robust standard errors. The model included terms for year of birth, year of life, their interaction, delivery mode (CS or VD) and antibiotic timing (pre incision or post cord clamping).

Data recording quality

To evaluate the impact of data recording quality, we restricted the data set to THIN–CPRD records linked to secondary care records. It was expected that these records would be the most accurate source for mode of delivery. A Poisson model was fitted to the data in the same manner as the primary analysis.

Random effect

Our primary analysis assumes that all observations are independent. As the outcomes are based on Read codes that have been entered onto an electronic medical health system at the general practice, it is possible that observations from children within the same general practice are correlated in some way (both because of access to the same general practice and also because they are likely to share socioeconomic similarities). To explore the possibility of non-independence of observations, we explored whether or not the results were robust to the inclusion of a random effect for general practice. This allows observations within a general practice to be correlated. To this end, a mixed-effects Poisson model was fitted to the data, with an offset for exposure time and robust standard errors and a random effect for general practice. Fixed effects were included for year of birth, year of life, their interaction, mode of birth (CS or VD) and antibiotic timing (i.e. probability of pre-incision prophylactic antibiotics).

Subgroup analyses

Children born by elective CS are likely to be exposed to antibiotics in utero for longer than children born by emergency CS. To assess whether or not the impact of the antibiotic policy change is different in elective CS and emergency CS, we categorised each birth as being elective CS, emergency CS or VD. A Poisson model was fitted to the data, with an offset for exposure time and robust standard errors. Fixed effects were included for year of birth and year of life, and their interaction. There were also terms for mode of delivery (i.e. emergency CS, elective CS or VD) and antibiotic timing (i.e. probability of pre-incision prophylactic antibiotics), and an interaction between them. We report the IRR for antibiotic timing in the emergency CS and elective CS groups, and the p-value for the interaction between them. This p-value indicates the strength of evidence of a differential effect of antibiotic timing based on CS type.

Secondary care data (HES)

Summarising data in secondary care

The primary outcomes (asthma and eczema) are summarised in the HES database using the number of children admitted to hospital at least once and the length of time at risk. For each outcome, an admission is indicated by ICD-10 codes, as described earlier (see Study outcomes). Children are categorised as having ‘no admission’ or as having ‘one or more admissions’. The length of time at risk is calculated as the time from birth to the earliest of the child’s fifth birthday or the study end date of 31 March 2019. We present the rate of the outcome per 1000 person-years, which is calculated as the ratio of the number of events and the exposure time multiplied by 1000.

Most dichotomous secondary outcomes are summarised in the same way, where we present the number of children admitted at least once, the total follow-up time and the outcome rate per 1000 person-years. For secondary outcomes with short exposure time (such as neonatal sepsis), we present the outcome rate per 1000 births.

Dichotomous maternal outcomes are summarised in the HES data set using the number of events and the percentage of total events to which this corresponds. One continuous maternal outcome (i.e. hospital length of stay) was not normally distributed and so the median and interquartile range are presented. All maternal outcomes are summarised stratified by delivery mode (VD or CS).

Analysis methods used

To analyse the primary outcomes (asthma and eczema), for each child we calculated whether or not they had been admitted to hospital admission for the primary reason of asthma or eczema and their exposure time. Owing to data limitations, we did not have the date of admission and so could not look at cases by year of life. We did, however, examine cases by year of birth. To estimate the impact of antibiotic timing (pre-incision or post-cord clamping prophylactic antibiotics), a Poisson model was fitted to the data, with an offset for exposure time and robust standard errors to allow for misspecification. The model included terms for year of birth, mode of birth (CS or VD) and antibiotic timing. Data were not available on the antibiotic timing (pre incision or post cord clamping) for individual births, but data on the hospital policy for each year were known. It was assumed that prophylactic antibiotics for each CS were administered in accordance with the hospital’s policy for that particular year. Antibiotic timing was categorised as before policy change, year of policy change and after policy change. As the policy change was applicable to CS births only, all VD births were classified as ‘before policy change’. The comparison between children born after the policy change and those born before the policy provides an estimate of the effect of the policy change in antibiotic timing. This comparison did not include those children born in the year of policy change, as their exposure to prophylactic antibiotics is not known. We present the IRR and 95% CI for the antibiotic timing variable, comparing the rate of events in those given prophylactic antibiotics pre incision with the rate of events in those given prophylactic antibiotics post cord clamping.

All dichotomous secondary outcomes are analysed in the same manner as the primary outcomes. For outcomes with a small number of events (< 200), we present the summary statistics only. For continuous outcomes (i.e. maternal length of stay), we fitted a linear regression model, with fixed effects for year of birth, mode of birth (CS or VD) and antibiotic timing. We present the difference in mean length of stay between those children born by CS with pre-incision prophylactic antibiotics and those children born by CS with post-cord clamping prophylactic antibiotics.

Sensitivity analyses

To assess the robustness of the results, we conducted several sensitivity analyses. Sensitivity analyses were performed for the two co-primary outcomes (asthma and eczema). Some secondary outcomes were included in the sensitivity analyses to check for the robustness of the primary analysis.

In the secondary care data set (HES), we conducted the following sensitivity analyses:

• definition of timing of the prophylactic antibiotic policy change
• random effect
• exploratory discordant sibling analysis.

Definition of timing of the prophylactic antibiotic policy change

The primary analysis of the secondary care data set is limited to hospitals that responded to the national survey, as we required information on whether or not the antibiotic policy was changed and the year in which the change was made. As not all hospitals responded to the survey, we wanted to explore the sensitivity of the results to the timing of antibiotic policy change. To this end, we estimated the probability of pre-incision prophylactic antibiotics based on calendar year. These probabilities were calculated using responses to the national survey from England hospitals, as HES includes data from hospitals in England only. A Poisson model was fitted to the data, with an offset for exposure time and robust standard errors. Fixed effects were included for year of birth, mode of birth (CS or VD) and antibiotic timing (probability of pre-incision prophylactic antibiotics).

Random effect

Our primary analysis assumes that observations are independent. However, it is possible that observations from children born from the same hospital are correlated in some way, perhaps because of sociodemographic and health-care access similarities. To explore the possibility of non-independence of observations, we explored whether or not the results were robust to the inclusion of a random effect for hospital. This allows observations from children born within the same hospital to be correlated. To this end, a mixed-effects Poisson model was fitted to the data, with an offset for exposure time and robust standard errors, and a random effect for hospital. Fixed effects were included for year of birth, mode of birth (CS or VD) and antibiotic timing (pre incision or post cord clamping).

Discordant sibling analysis

In an attempt to control further for family-related genetic and environmental factors, we employed a discordant sibling analysis. For this, we created a subset of data in which women gave birth at least twice using the same mode of birth (VD or CS) during the study, including before and after the change in antibiotic policy. A Poisson model was fitted to the data in the same manner as the primary analysis of the secondary care data.

Subgroup analyses

In the secondary care data set, we performed two subgroup analyses: (1) CS type analysis (i.e. elective or emergency) and (2) antibiotic regiment analysis (i.e. cefuroxime only, cefuroxime and metronidazole, and co-amoxiclav only).

Caesarean section type

To assess whether or not the impact of the antibiotic policy change differs between elective CS and emergency CS, we categorised each birth as being elective CS, emergency CS or VD. A Poisson model was fitted to the data, with an offset for exposure time and robust standard errors. Fixed effects were included for mode of delivery (emergency CS, elective CS or VD) and antibiotic timing (using probability of antibiotic policy change), and an interaction between them. Adjustments were included for year of birth. We report the IRR for antibiotic timing in the emergency and elective CS groups, and the p-value for the interaction between them.

Antibiotic regimen

Different antibiotic regimens may have a differential effect on longer-term child health. Based on the national survey, the most common antibiotics used were co-amoxiclav alone, cefuroxime alone, and cefuroxime and metronidazole. We restricted the analysis to hospitals that used the same antibiotics before and after the policy change. A Poisson model was fitted to the data, with an offset for exposure time and robust standard errors. Fixed effects were included for year of birth, mode of delivery (CS of VD) and antibiotic timing (pre incision or post cord clamping). We fitted a model separately for each antibiotic type (co-amoxiclav alone, cefuroxime alone, cefuroxime and metronidazole). We report the IRR for antibiotic timing separately for each antibiotic type. This exploratory subgroup analysis was undertaken for primary outcomes (asthma and eczema) and necrotising enterocolitis because of the previously reported increased risk of this condition in babies born to mothers who were exposed to co-amoxiclav.83

Analysis was carried out using Stata^® MP version 16 (StataCorp LP, College Station, TX, USA).

Research Ethics Committee approval

The Ethics Review Committee of the University of Birmingham provided the ethics approval required for this study (reference number ERN_17-1675).

The NHS South-East Multi-Centre Research Ethics Committee approved the use of the THIN database, subject to independent scientific review to help ensure appropriate analysis and interpretation of the data.84 The Independent Scientific Ethical Advisory Committee – Scientific Review Committee Panel of the data provider IQVIA (Durham, NC, USA) approved the use of the THIN and HES-linked data in this study (reference number 18THIN047).

The CPRD data set has ethics approval from a National Research Ethics Service Committee for observational research using anonymised data.85 For this study, the use of CPRD has been approved by the Independent Scientific Advisory Committee for MHRA Database Research (reference number 18_181AR2).

The use of the HES database is exempt from NHS Research Ethics Committee approval because it involves the analysis of an existing data set of non-identifiable data. The standard NHS Digital data approval process was followed to obtain approval for the use of HES data.86 The Health Research Authority confirmed that Health Research Authority approval was not required for this research, as it involved linking anonymised patient data from established databases for this study only.

Changes to the protocol

There were no changes to the protocol.

Patient and public involvement

Aims

There were two main aims of patient and public involvement (PPI) in this project. First, we involved women to ensure that we included outcomes that are important for parents and children. We explored with parents and parents-to-be which health conditions, particularly allergy- and immune system-related conditions and maternal infections after birth, are most important to them and, therefore, which health conditions the study should ideally investigate (assuming that information for these conditions is well recorded in routine medical practice). Second, our public involvement activities helped ensure that key messages from the research about the timing of prophylactic antibiotic administration in CS were communicated to relevant audiences in ways that were clear and accessible. We report our PPI here using the format recommended by the Guidance for Reporting Involvement of Patients and the Public Short Form (GRIPP2-SF).87

Methods

The chief investigator (DS) provided leadership for all public involvement in the project and was the first point of contact for public contributors and researchers on all matters relating to public involvement. Two women (RH and MS) were recruited to be involved in the study. Ruth Hewston is a volunteer with the National Childbirth Trust (London, UK) and has experience of involvement in a range of maternity projects. Magdalena Skrybant had a CS in 2013 and has an interest in maternity research. Both Magdalena Skrybant and Ruth Hewston were invited to attend all Project Management Group meetings. Materials were sent in advance of meetings and Ruth Hewston and Magdalena Skrybant could contact Dana Šumilo at any point to discuss the project. Both Magdalena Skrybant and Ruth Hewston were invited to contribute to all reports and they are acknowledged as co-authors to publications. Public contributors were also offered an honorarium for their involvement.

We also recruited an independent public contributor to our Project Steering Group (MC). MC was recruited through the Royal College of Obstetricians and Gynaecologists (RCOG) (London, UK) Women’s Voices Involvement Panel. Maria Clark was invited to attend all meetings, received paperwork in advance and could contact Dana Šumilo to discuss the study.

To gain wider public perspectives, public involvement sessions took place as part of the ‘Yoga for Bump’ sessions, which were facilitated by National Institute for Health and Care Research (NIHR) Collaborations for Leadership in Applied Health Research and Care and funded through a Wellcome Trust (London, UK) Engagement Fund. These sessions were free for women to attend and included an informal chat with a midwife, a research engagement session and then a pregnancy yoga session. The sessions were held in June 2018 in two different community locations in the West Midlands (Hall Green and Perry Barr) and at two different times of the day, which helped maximise the diversity of the contributors. Eight mothers or mothers-to-be took part in the sessions. They came from a variety of backgrounds, and included women from ethnic minority communities, an often under-represented group in research. The researcher (DS) and a PPI advisor (MS) of the project explained the background of this research and the research questions. This was followed by the discussion, in relation to this study, of which health conditions are most important to women and how the findings of the study can be shared.

We also obtained perspectives from women from the RCOG Women’s Voices Involvement Panel who responded to the wider clinician and public consultation survey in July 2018. A link to the online survey was also sent to the British Intrapartum Care Society (London, UK), the British Maternal & Fetal Medicine Society (London, UK) and the British Association of Perinatal Medicine (Bristol, UK). In the survey, respondents were asked to select up to three conditions from each group of health conditions that we were interested in for inclusion as study outcomes (the survey questionnaire is included in Report Supplementary Material 2). Twenty-nine respondents (22 clinicians and seven parents) completed the survey.

Both Magdalena Skrybant and Ruth Hewston contributed to the writing of this report, including Patient and public involvement and the Plain English summary.

Results

Patient and public involvement, including the public involvement sessions and the wider survey, helped us to confirm that we should look at a wide range of outcomes, including neurodevelopmental conditions. From the proposed outcome list in the online survey, there were no outcomes that were not selected by the respondents, apart from septic thrombophlebitis, which we excluded from the final outcome list. Our public involvement sessions also confirmed that we should consider the severity of outcomes.

Patient and public involvement also helped us to identify other information mothers-to-be wanted explained in relation to the study, including how antibiotics work, what antibiotics are recommended, the difference between intrapartum antibiotics and antibiotics during pregnancy, and what impact antibiotics may have depending on when they are provided. PPI also confirmed the importance of balancing the risks and benefits to the mother and baby when communicating the study findings.

Discussion and conclusions

Patient and public involvement helped us to confirm the finding from our initial PPI while developing this research proposal that a broad range of outcomes should be investigated. PPI engagement sessions informed the development of the wider consultation through the online survey and the final outcome list. PPI also confirmed the importance of capturing and interpreting the data, taking into account the severity of health conditions, which was addressed in this study by assessing the health conditions in both primary care and hospital data. Furthermore, PPI strengthened the importance of comparing the impact on children with the impact on mothers and helped us to better understand what relevant information women need to know when communicating the results of the study.

Reflections

The study benefited from having two very experienced public contributors as full members of the Project Management Group since the inception of this research and also a very experienced independent public contributor as a member of the Project Steering Group.

Wider PPI activities were useful in confirming that the outcomes we were planning to investigate were relevant to parents and parents-to-be and in identifying additional outcomes of interest. Public involvement sessions were also useful in identifying what contextual information needs to be provided when communicating the findings to the wider public. The 30-minute duration of the public involvement sessions was sufficient for addressing the main objectives of this PPI. Given the complexity of the research topic and uncertainty around the potential impact the antibiotics may have, it was initially difficult for women to comment on the outcomes that were important to them. It was very helpful to have an experienced PPI advisor facilitating public involvement sessions, maximising the benefit gained from the sessions.