Open access Protocol Delayed antibiotic prescribing for respiratory tract infections: protocol of an individual patient data meta-analysis Beth Stuart,1 Hilda Hounkpatin,1 Taeko Becque,1 Guiqing Yao,2 Shihua Zhu,1 Pablo Alonso-Coello,3 Attila Altiner,4 Bruce Arroll,5 Dankmar Böhning,6 Jennifer Bostock,7 Heiner C C Bucher,8 Mariam de la Poza,9 Nick A Francis,10 David Gillespie,11 Alastair D Hay,12 Timothy Kenealy,5 Christin Löffler,13 Gemma Mas-Dalmau,14 Laura Muñoz,15 Kirsty Samuel,16 Michael Moore,1 Paul Little1 To cite: Stuart B, Hounkpatin H, Abstract Becque T, et al. Delayed Strengths and limitations of this study Introduction Delayed prescribing can be a useful antibiotic prescribing for strategy to reduce antibiotic prescribing, but it is not respiratory tract infections: ►► This study uses individual patient data (IPD) from clear for whom delayed prescribing might be effective. protocol of an individual patient randomised controlled trials and observation- This protocol outlines an individual patient data (IPD) data meta-analysis. BMJ Open al studies to investigate the clinical effectiveness 2019;9:e026925. doi:10.1136/ meta-analysis of randomised controlled trials (RCTs) and potential cost-savings of a delayed antibiotic bmjopen-2018-026925 and observational cohort studies to explore the overall prescribing, both overall and for key subgroups of effect of delayed prescribing and identify key patient ►► Prepublication history and people. characteristics that are associated with efficacy of additional material for this ►► IPD provides sufficient statistical power to explore delayed prescribing. paper are available online. To interactions between treatment groups and patient Methods and analysis A systematic search of the view these files, please visit characteristics. the journal online (http://​dx.​doi.​ databases Cochrane Central Register of Controlled ►► The IPD will only include data from studies for which org/​10.​1136/​bmjopen-​2018-​ Trials, Ovid MEDLINE, Ovid Embase, EBSCO CINAHL the data are available. 026925). Plus and Web of Science was conducted to identify ►► Observational studies will use propensity score ap- relevant studies from inception to October 2017. BS and HH contributed equally. proach to adjust for confounding by indication on Outcomes of interest include duration of illness, measured covariates. severity of illness, complication, reconsultation and Received 3 October 2018 ►► Findings from this study may inform GPs decisions patient satisfaction. Study authors of eligible papers Revised 5 November 2018 on prescribing and increase use of appropriate de- will be contacted and invited to contribute raw IPD Accepted 22 November 2018 layed prescribing which in turn can help to reduce data. IPD data will be checked against published data, antibiotic consumption and antimicrobial resistance. harmonised and aggregated to create one large IPD database. Multilevel regression will be performed to explore interaction effects between treatment allocation and patient characteristics. The economic Introduction evaluation will be conducted based on IPD from the Antimicrobial resistance is a cause for great combined trial and observational studies to estimate concern prompting calls for action at the the differences in costs and effectiveness for delayed prescribing compared with normal practice. A decision local, national and international level to model will be developed to assess potential savings prevent ‘overuse, misuse and abuse’ of antibi- and cost-effectiveness in terms of reduced antibiotic otics,1 particularly in primary care where anti- usage of delayed prescribing and quality-adjusted life biotics are most prescribed.2 About 60% of years. antibiotics prescribed in primary care are for Ethics and dissemination Ethical approval was respiratory tract infections (RTIs).3 However, obtained from the University of Southampton Faculty most infections are self-limiting, symptom- of Medicine Research Ethics Committee (Reference atic benefit from antibiotics is modest4–6 © Author(s) (or their number: 30068). Findings of this study will be employer(s)) 2019. Re-use and patients prescribed antibiotics for RTIs published in peer-reviewed academic journals as permitted under CC BY. are more likely to carry antibiotic resistant well as General Practice trade journals and will be Published by BMJ. commensal bacteria and develop resistant presented at national and international conferences. For numbered affiliations see The results will have important public health infections.7 end of article. The Standing Medical Advisory Committee implications, shaping the way in which antibiotics Correspondence to are prescribed in the future and to whom delayed report recommends that the fewest number Dr Beth Stuart; prescriptions are issued. of antibiotic courses should be prescribed for ​bls1@​soton.​ac.​uk PROSPERO registration number CRD42018079400. the shortest period possible.8 However, in a Stuart B, et al. BMJ Open 2019;9:e026925. doi:10.1136/bmjopen-2018-026925 1 Open access primary care context, it can be difficult to tell whether or higher risk of complications/reconsultation. These antibiotics are appropriate for an individual patient. It is findings may inform GPs decisions on prescribing and not always clear which patients are at risk of prolonged increase use of appropriate delayed prescribing which in illness or developing complications nor whether the RTI turn can help to reduce antibiotic consumption and anti- is bacterial or viral in nature. Point of care tests such as C microbial resistance. reactive protein and clinical scores both show promise in This protocol describes an IPD meta-analysis of both helping to guide GP management decisions.9–12 Delayed randomised controlled trials (RCTs) and observational prescribing is a useful strategy that may be used on its own cohort studies of delayed antibiotic prescribing for acute or in conjunction with clinical scores and point of care respiratory infections. Although RCTs represent the ‘gold testing. It allows the patient to collect a prescription to standard’ design for assessing the impact of an interven- be taken if their symptoms do not start to improve within tion, the participants in these studies may differ systemat- a specific duration of their initial consultation, providing ically from the patients encountered in everyday clinical a potential aid to negotiating treatment acceptable to practice due to the inclusion criteria. By including obser- the patient and provider while reducing inappropriate vational studies in a meta-analysis, with suitable tech- prescribing.13 14 Delayed prescribing is recognised as niques to control for potential confounding by indication, part of the toolkit available to GPs to help reduce antibi- it may be possible to obtain better estimates of the effect otic use, especially in the context of consultations where of delayed prescribing in routine practice. patients expect to receive an antibiotic prescription and is part of the National Institute for Health and Care Excel- lence guidelines.3 A recent large prospective primary care Methods cohort (MRC DESCARTE) demonstrated that delayed Aims and objectives prescription was likely to be as effective as immediate This IPD study will investigate the clinical effective- antibiotic in reducing complications and more effective ness and potential cost-savings of a delayed antibiotic at reducing reconsultations.4 prescribing approach, both overall and for key subgroups While in some situations delayed antibiotic prescribing of patients. The study started in October 2017 and is due is appropriate, for other patients it may be unsuitable to be completed in September 2020. and controversial.15 It is important to understand which The objectives of this study are as follows: subgroups of patients may require immediate antibi- ►► To achieve more accurate estimates of the clinical otics to avoid complications and which patients might effectiveness of a delayed prescribing strategy on the benefit from a delayed or no prescribing strategy.4 A primary outcome, symptom severity, by harmonising 2013 Cochrane review16 of 10 trials found that a delayed the outcomes from all trials and observational studies prescribing strategy was not significantly different to which have included this strategy. immediate prescribing in terms of clinical outcomes for ►► To estimate the clinical effectiveness of delayed cough and cold. In patients with acute otitis media and prescribing for secondary outcomes: duration of sore throat immediate antibiotics were more effective illness, development of complications, reconsultation than delayed for reducing fever, pain and malaise in some and patient satisfaction. studies. However, the review noted a high level of hetero- ►► To explore whether there are key subgroups geneity between studies made combining them in a tradi- (informed by the literature and described below) for tional meta-analysis difficult and did not allow sufficient whom delayed antibiotic prescribing may or may not power for the examination of subgroups or complications be beneficial. Subgroups will be: prior duration of either in meta-analysis or in meta-regression. illness (above/below median for the condition), age Where aggregation is possible, a traditional meta-anal- (under 16 years, 16–64 years, age over 65 years), fever ysis or even meta-regression can be used to determine at baseline consultation (>37.5°C), comorbid condi- the overall main effect of delayed antibiotic prescription. tions including lung comorbidity such as asthma or Although such techniques can also be used to explore how chronic obstructive pulmonary disease (COPD) and effect varies across study characteristics, these techniques severity of symptoms at baseline consultation. The lack statistical power and are prone to confounding as effectiveness of delayed antibiotic prescribing for patient-level characteristics are not taken into account.17 subgroups of symptom complexes such as cough An individual patient data (IPD) meta-analysis is consid- (acute bronchitis), pharyngitis or otitis media will also ered the gold standard for meta-analyses and can over- be explored. come the limitations of meta-analysis of aggregate data. In ►► To investigate whether symptom trajectories are influ- an IPD meta-analysis, original data for all participants in enced by antibiotic prescribing/consumption. each study are obtained and synthesised. This allows suffi- ►► To estimate the costs of treatment of patients with cient statistical power18 to explore interactions between RTIs (from the UK National Health Service (NHS) treatment groups and patient characteristics. This and Personal Social Services (PSS) perspective data) would allow us to identify key groups in which delayed and investigate any potential cost-savings and cost-ef- prescribing or no prescribing might be inadvisable due fectiveness in terms of antibiotic usage of delayed anti- to longer durations of illness, greater severity of illness biotic prescriptions compared with antibiotic use. 2 Stuart B, et al. BMJ Open 2019;9:e026925. doi:10.1136/bmjopen-2018-026925 Open access ►► To identify priorities for future delayed prescribing intended interventions (participant and study personnel research. blinding, deviations from intended interventions and analysis in groups to which they were randomised), Study approach attrition bias and appropriate methods to account for A full systematic review will be conducted to identify and missing data, detection bias (blinding of outcome asses- select eligible papers. Study authors of eligible papers will sors) and selective outcome reporting.20 21 The “Risk be contacted and invited to contribute raw data. Of Bias In Non-randomised Studies - of Interventions” (ROBINS-I) tool will be used to assess quality of each Systematic review to identify eligible papers observational study. The ROBINS-I tool contains items on Eligibility criteria bias due to confounding, selection bias, bias due to devi- Design: RCTs or eligible observational cohort studies. ations from intended intervention, bias due to missing Population: all patients attending primary, ambulatory data and selective reporting.22 We will use GRADE to rate or acute care settings with a RTI. the overall certainty (quality) of evidence that includes Intervention: delayed antibiotic prescription. the evaluation of risk of bias, inconsistency, indirectness, Comparator: immediate antibiotic prescription or no imprecision and publication factors.23 antibiotic prescription. Outcomes: severity of illness (symptom severity), dura- Data extraction and database creation tion of illness, complication, reconsultation, patient satis- Authors of eligible trials and observational studies will be faction, costs and quality of life. contacted via an email, or where this is not possible by Studies on antibiotic prescribing that are not an RCT or a letter, outlining the study aims. They will be invited to observational cohort (eg, survey or cross-sectional studies collaborate and share their data in a format of their choice. and case–control studies) will be excluded. Studies of A data sharing agreement may be provided on request. hospital inpatients are also excluded. A complete database (containing data on all available Search strategy for identification of studies measures) rather than key variables used in the publica- A systematic search of the databases Cochrane Central tion will be requested. If resource usage and quality of Register of Controlled Trials, Ovid MEDLINE, Ovid life data were collected in separate forms, such informa- Embase, EBSCO CINAHL Plus and Web of Science will tion should be provided too. This will allow us to recalcu- be conducted to identify relevant papers published from late the primary outcomes where necessary to ensure a inception to October 2017. The search strategy will be consistent approach has been used across studies with the based on criteria set out by the Cochrane Collaboration aim of reducing the level of between-study heterogeneity. in a recently published systematic review and is available Collaborators will be able to provide the data in a format in the online supplementary appendix.19 The Interna- of their choice. Study data will be considered unavailable tional Standard Randomised Controlled Trial Number in the event that none of the authors have responded to Registry, a primary clinical trial registry recognised by multiple contact attempts or if all contacted study authors WHO and ICMJE that accepts and records all clinical indicate they no longer have access to the data. research studies in order to improve the publicly avail- Data will be checked by comparing key variables (eg, able information about clinical studies, will also be size of sample, descriptive statistics of demographic searched for any relevant trials. Literature search results and outcome measures) with published data. We will will be exported to EndNote and then uploaded to COVI- aim to reproduce individual study results to ensure the DENCE, a web-based software that facilitates collabora- current analysis is consistent with what has been done in tion among reviewers during the study selection process. each study. Collaborators will be contacted if important Full-text articles will be uploaded to COVIDENCE. Two discrepancies are identified and asked for clarification. reviewers (HH and TB) will independently assess articles The level of missing data within each included study will to determine eligibility for inclusion and any discrepan- be assessed and discussed with study collaborators. If less cies will be resolved by discussion or referred to a third than 5% of data is found to be missing on key baseline reviewer (BS). Reference lists of included articles will be characteristics and outcome measures, the data will be reviewed to identify other potential papers not retrieved analysed on a complete cases basis. However, if there is in the initial search. Contributing authors and content substantial missing data, the pattern and nature of the experts will be asked if they have, or are aware of, any missingness will be explored and multiple imputation by additional studies (published or unpublished). chained equations will be used if appropriate. A copy of the data to be included in the IPD meta-anal- Risk of bias assessment and certainty of evidence ysis will be converted for analysis in STATA V.14 or above. assessment A database will be created containing study ID, patient Two reviewers will independently assess the risk of bias ID, outcomes measures, subgroups of interest, antibiotic of each included study. RCTs will be assessed using the prescribing approach and propensity scores, where neces- Cochrane risk of bias tool that includes items on allo- sary. Outcome measures from the different studies will be cation bias (random sequence generation, allocation harmonised once data from all study authors have been concealment, baseline imbalance), departures from received. In the case that coding for an outcome in a Stuart B, et al. BMJ Open 2019;9:e026925. doi:10.1136/bmjopen-2018-026925 3 Open access study differs greatly from other studies, it may be possible reconsultation outcome, side effects and patient satisfac- to simplify measures to allow the study to contribute data tion. All models will control for baseline severity of illness or exclude the study from certain analyses. Any difficul- and all individuals will be included on a modified inten- ties with harmonisation will be discussed with the full tion to treat (ITT) basis (ie, as randomised).27 collaborators group to decide on the most meaningful To explore whether there are differences in treat- approach. Harmonised data from the different studies ment effects for certain sub-groups, we will consider five will be aggregated (using STATA) into a single large IPD prespecified subgroups which have been identified as database, with an indicator variable to identify patients potential effect modifiers by previous research: prior dura- from the same trial. The IPD database will be rechecked tion of illness (above/below median for the condition), for accuracy. age (under 16 years, 16–64 years, age over 65 years), fever The key resource information on medication, primary at baseline consultation (>37.5°C), comorbid condition care consultation, hospital inpatients, hospital day cases, including lung comorbidity such as asthma or COPD, and outpatient visits, accident and emergency (A&E) atten- severity of symptoms at baseline consultation. We will also dance from individual studies will be recorded if such explore differences according to the diagnostic group information were collected and length of usage and (acute sore throat, cough/chest infection, otalgia/otitis frequency of attendance will be adjusted to the same media, upper RTI). The analyses described above will be time periods. We will use the UK national published tariff repeated, testing for interactions between treatment allo- (British National Formulary (BNF), national reference cation and sub-group characteristic. cost and PSS Research Unit (PSSRU)) to cost-associated Symptom trajectories in patients who do not take anti- resource usage. biotics will also be explored. For symptoms recorded in symptom diaries, we will calculate the median and IQR Analysis of the reported time until all symptoms have settled Study and patient level characteristics will be described completely, in those patients who did not receive an anti- for all studies that contribute IPD. Characteristics of any biotic prescription or who received a delayed prescrip- studies that declined or were unable to provide data will tion but did not fill it (where these data are available). be considered in order to determine the extent to which As sensitivity analyses, we will also explore duration of the included studies are a representative sample. A tradi- moderately severe of symptoms. It is possible that the use tional meta-analysis may be performed, using data from of multiple outcomes (symptom severity, development of published papers, to test for differences between studies complications, patient satisfaction) to determine clinical included in the IPD meta-analysis and those that could effectiveness may result in varying conclusions. We will not provide data.24 report the results for each outcome, though the discus- Heterogeneity of eligible studies will be summarised sion and recommendations will be based primarily on the using an I2 statistic (tested by Higgins I2 test). A substan- primary outcome– symptom severity. tial statistical heterogeneity will be considered if I² statistic Work on cross design synthesis by the Program and is >50%.21 Sources of heterogeneity will be explored and Methodology Division of the US General Accounting reported and a random effects model will be used for all Office (1992) has provided a theoretical framework for analyses. combining database data and RCT data. Drawing on this work, the current review will apply a similar approach IPD meta-analysis to the synthesis of observational cohort and RCT data.28 IPD meta-analysis will be conducted using a one-stage Synthesis of observational studies will additionally use approach. The one-stage approach combines all the data inverse probability weighting by propensity score anal- in a single meta-analysis based on a regression model ysis to adjust for confounding by indication on measured stratified by study.25 One-stage analysis is often more covariates. This serves to balance observational studies on appropriate for exploring treatment-covariate interac- key covariates) so that the distribution of baseline covari- tions as it has increased power and is less likely to suffer ates is similar between treated and untreated subjects, from aggregation bias.26 The aggregated IPD will be making observational studies more like those in a RCT.29 30 examined using multilevel regressions clustering on the Potential covariates to be included in the propensity score individual study level to take into account any heteroge- include: demographic factors (eg, age, sex), comorbid neity between studies. Using the aggregated IPD dataset, health conditions and signs and symptoms at baseline we will calculate the key outcome measures on a consis- consultation. Observational studies after this adjustment tent basis. The model selected will be one that is appro- for potential confounding variables will be included in priate to the outcome measure of interest. For example, the main analysis if appropriate. However, sensitivity assuming the underlying assumptions of the model are analyses will explore whether the estimates from observa- met, a linear regression will be used to model the severity tional cohort studies differ from those found in the RCTs. of symptoms, a proportional-hazards model will be used If there is evidence of substantial differences, data from to assess the duration of illness, or a suitable count model observational studies and RCTs will not be pooled in a where time to event data is not available, and logistic single meta-analysis, and reasons for these differences will regression models will be used for the complications/ be explored narratively. 4 Stuart B, et al. BMJ Open 2019;9:e026925. doi:10.1136/bmjopen-2018-026925 Open access Economic analysis The economic analyses will follow the same approach as The economic evaluation will be taken from the NHS and the statistical analyses with calculations undertaken using PSS perspective, as most of the studies in this area have a one-stage approach. collected such data. The NHS and PSS perspective is the most commonly used approach to economic evaluations Sensitivity analyses conducted in the UK. It covers medication, primary care All analyses will be repeated using a two-step approach. consultation, walking in centre, NHS telephone service, The two-stage approach involves calculating the effect outpatient attendance, A&E visit and hospital admis- for each study separately and then combining the results sion. Outcomes will be differences in cost, costs per anti- using traditional meta-analysis techniques. Results using biotic prescription avoided and costs per quality of life the one-stage and two-stage approach will be compared gained. Individual patient resource usage data such as and any discrepancies explored. This will also allow us to medication use, GP consultations, outpatient visits, A&E include and explore the impact of the aggregate results attendance and use of secondary care service from each from trials for which we are unable to obtain data, as aggre- individual trial or observational study will be costed using gate results can only be included in a two-step approach. UK published data sources (BNF, national reference costs Aggregated data from observational studies will not be and PSSRU). Total costs for individual patients during included because in the absence of full data, it will not be study periods will be calculated. We will also take a soci- possible to control for confounding by indication. etal perspective by including the societal cost of antimi- Sensitivity analysis will explore whether there are differ- crobial resistance and time of work due to illness in our ences in inferences based on differences in the approach modelling and sensitivity analyses. taken to delayed prescribing and whether there are any A decision analytical model will be developed to project differences if studies at high risk of bias are excluded. We the potential saving (due to reduced antibiotic usage, and will also consider a higher cut-off for presence of fever reduced frequency of attending primary care consulta- informed by the literature. tion) and cost-effectiveness in terms of reduced antibiotic Given the potential long-term problem of antibi- usage, and quality-adjusted life years (QALY) gained of otic resistance, a sensitivity analysis will performed to delayed antibiotics. The model will consider reinfection incorporate the costs of antibiotic resistance in the and present the results at different time lengths, up to economic analyses. However, given the difficulty previous a maximum of 5 years. The model will be an individu- researchers31 32 have had in estimating such costs we will al-based micro-simulation taking into account individual vary the assumptions about the likely cost to see whether risk profiles with associated risk equations. The model the inferences are modified. will simulate individual treatment pathways with defined baseline characteristics and will capture the time spent at Patient and public involvement different levels of disease severity, time to recovery and The study has two patient and public involvement (PPI) recurrence rates. The probabilities of staying or transi- team members who have been actively involved in shaping tioning to another state will be derived from the statis- the research questions and will continue to contribute to tical analyses. The clinical analyses based on all data will all stages of the research. Specifically, they will contribute provide more robust estimates to populate the input to data harmonisation, providing advice to ensure the values for the data. The costs and cost-effectiveness of result will yield useful outcomes for patients and members subgroups will be explored in the decision model. of the public, interpretation of study findings and what Quality of life questionnaire will be translated into they might mean for patients and the public, dissemina- utility scores based on the UK tariff appropriate to the tion at relevant conferences/meetings and how best to instruments used in each study. The different instruments disseminate findings to key patient and public groups. (such as EQ5D, SF6D or Health Utilities Index (HUI)) In return for their time, PPI representatives will be reim- will be recorded and weighted in the aggregated anal- bursed in line with INVOLVE recommendations33 and yses. Utility scores at each measurement point will be esti- will be recognised as co-authors on research outputs. mated, and a mean utility score for symptom severity will be calculated. For studies where no quality of life data Ethics and dissemination were collected, we will apply the utility scores estimated There will be no identifiable patient data in any of the from the other studies according to symptom severity. datasets, and data will be stored in a password-restricted QALYs will be calculated based on an area under the folder on the university server. This protocol has been curve approach. Due to the potentially skewed distribu- prepared in accordance with the Preferred Reporting tion in costs, a generalised linear mixed model will be Items for Systematic Reviews and Meta-Analyses Protocol employed to analyse costs and QALYs, controlling for guidelines.34 baseline patient characteristics, length of follow-up and The results of this work will provide the best possible disease severity. The data will be hierarchical, with indi- evidence regarding when and for whom delayed vidual patients nested within studies, with study modelled prescribing is appropriate. The results, disseminated as a random effect. Incremental costs and cost-effective- through academic and trade journals, as well as confer- ness and associated CIs will be estimated by bootstrapping. ences, should help GPs in their decision-making about Stuart B, et al. BMJ Open 2019;9:e026925. doi:10.1136/bmjopen-2018-026925 5 Open access 3 prescribing during consultations. The results will also be Iberoamerican Cochrane Centre, Department of Clinical Epidemiology and Public shared with patients and members of the public through Health, Biomedical Research Institute Sant Pau (IIB Sant Pau), Barcelona, Spain 4 Office for Educational Affairs, Department of General Medicine, University of our PPI collaborations. There is potential for PPI to help Rostock, Rostock, Germany address the issue of patients seeking antibiotics inappro- 5 Department of General Practice and Primary Health Care, University of Auckland, priately which may help reduce unnecessary prescribing Auckland, New Zealand 6 and in turn, antimicrobial resistance. Southampton Statistical Sciences Research Institute, University of Southampton, Southampton, UK 7 Divison of Health and Social Care Research, King's College London, London, UK 8 Basel Institute for Clinical Epidemiology and Biostatistics (CEB), University Hospital Basel and University of Basel, Basel, Switzerland Discussion 9 Institut Català de la Salut, CAP Doctor Carles Ribas, Barcelona, Spain IPD meta-analysis is more appropriate and has higher 10 Division of Population Medicine, School of Medicine, Cardiff University, Cardiff, UK 11 statistical power than traditional meta-analysis to identify Centre for Trials Research, College of Biomedical & Life Sciences, Cardiff which key subgroups or patient characteristics may or may University, Cardiff, UK 12 Centre for Academic Primary Care, Population Health Sciences, Bristol Medical not benefit from delayed prescribing. Another advan- School, University of Bristol, Bristol, UK tage of IPD is that some of the biases in studies can be 13 Institute of General Practice, Rostock University Medical Center, Rostock, Germany explored more fully than in aggregate data meta-analysis. 14 Knowledge and Research Management Unit, Hospital de la Santa Creu i Sant Pau, For example, it is possible to check the randomisation Barcelona, Spain 15 integrity and to perform a full ITT analysis to minimise Agència de Qualitat i Avaluació Sanitàries de Catalunya (AQuAS), Barcelona, Spain 16 ASPIRE PPI Panel, Leeds, Institute for Health Sciences, University of Leeds, Leeds, attrition bias, even if this was not performed for the orig- UK inal publication. Furthermore, with the collaboration and communication required for a successful IPD, it should Acknowledgements We would like to thank the following collaborators which be possible to obtain information from the original study assisted us by allowing us to use the data from their studies: Dr Jennifer Chao, authors about potential sources of bias. A further advan- Dr David McCormick, Ms Jacqueline Nuttall. tage of IPD meta-analysis is that it allows study outcomes Contributors BS is the guarantor. BS conceived the original study concept to be calculated in the same way and the models adjusted and design and obtained funding from the NIHR Research for Patient Benefit (RfPB) Programme. BS and HH wrote the first draft of the manuscript. GY and SZ for similar confounders which can reduce heterogeneity contributed to the section on health economics. BS, HH, TB, GY, SZ, PA-C, AA, BA, across studies. DB, JB, HCCB, MdlP, NAF, DG, ADH, TK, CL, GM-D, LM, KS, MM and PL contributed However, there are some challenges associated with to the concept and design of the study and manuscript editing, read, provided IPD meta-analysis.35 The key risk is not obtaining all the feedback and approved the final manuscript. relevant data and failure to do so can result in biased Funding This paper presents independent research funded by theNational Institute results. It may be particularly difficult to obtain data for for Health Research (NIHR) under its Research for PatientBenefit (RfPB) Programme (Grant Reference Number PB-PG-0416-20005). The viewsexpressed are those of older studies, as the data may have been lost or destroyed the author(s) and not necessarily those of the NHS, theNIHR or the Department of or the lead researcher may have retired or changed field. Health & Social Care. We therefore plan to undertake a sensitivity analysis using Disclaimer The NIHR RfPB is not involved in any other aspect of the project, a two-step approach which will allow us to incorporate such as the design of the project’s protocol and analysis plan, the collection and the published estimates from any trials for which we are analyses. The funder will have no input on the interpretation or publication of the study results. unable to obtain IPD into a meta-analysis. Although IPD offers the ability to conduct additional and more accu- Competing interests None declared. rate and appropriate analyses, in many cases similar Patient consent for publication Not required. results and conclusions can be drawn from IPD and stan- Ethics approval University of Southampton Faculty of Medicine Research Ethics dard meta-analysis.35 It is also possible that the outcome Committee (Reference number: 30068). measures may not be sufficiently similar to harmonise. Provenance and peer review Not commissioned; externally peer reviewed. An initial scoping review suggests that almost all studies Open access This is an open access article distributed in accordance with the have collected the key outcome measures in a similar Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits way. However, if it is deemed impossible to perform others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, a meta-analysis with the data; a narrative review will be and indication of whether changes were made. See: https://​creativecommons.​org/​ undertaken instead. We hope the results of the study will licenses/​by/​4.​0/. help GPs to communicate more effectively with patients about the normal course of illness, which patients are likely to benefit from a delayed prescribing approach, the costs of antibiotic use for RTIs and help the general public References to feel more confident about waiting to see whether their 1. Ridge KW, Hand K, Sharland M, et al. Infections and the rise of antimicrobial resistance. London: Department of Health, 2011. illness settles before visiting their GP. 2. Public Health England. English surveillance programme for antimicrobial utilisation and resistance report, 2016. Author affiliations 3. National Institute for Health and Care Excellence. Prescribing of 1 antibiotics for self-limiting respiratory tract infections in adults and Academic Unit of Primary Care and Population Sciences, Faculty of Medicine, children in primary care (Clinical Guideline 69). 2008. University of Southampton, Southampton, UK 4. Little P, Stuart B, Hobbs FD, et al. Antibiotic prescription strategies 2 Biostatistics Research Group, Department of Health Sciences, College of Life for acute sore throat: a prospective observational cohort study. Sciences, University of Leicester, Leiceister, UK Lancet Infect Dis 2014;14:213–9. 6 Stuart B, et al. BMJ Open 2019;9:e026925. doi:10.1136/bmjopen-2018-026925 Open access 5. Spinks A, Glasziou PP, Del Mar CB. Cochrane Acute Respiratory international clinical trials methodology conference, ICTMC and the Infections Group. Antibiotics for sore throat. Cochrane Database Syst 38th annual meeting of the society for clinical trials United kingdom. Rev 2013;139:CD000023. 2017;18(Supplement 1). (no pagination). 6. Kenealy T, Arroll B. Antibiotics for the common cold and acute 21. In: Higgins JPT, Green S, eds. Assessing risk of bias in included purulent rhinitis. Cochrane Database Syst Rev 2013;4:CD000247. studies: The Cohrane Collaboration, 2011. 7. Costelloe C, Metcalfe C, Lovering A, et al. Effect of antibiotic 22. Sterne JA, Hernán MA, Reeves BC, et al. ROBINS-I: a tool for prescribing in primary care on antimicrobial resistance in individual assessing risk of bias in non-randomised studies of interventions. patients: systematic review and meta-analysis. BMJ 2010;340:c2096. BMJ 2016;355:i4919. 8. Standing Medical Advisory Committee S-GoAR. The path of least 23. Guyatt G, Oxman AD, Akl EA, et al. GRADE guidelines: 1. resistance 1998. Introduction-GRADE evidence profiles and summary of findings 9. Cooke J, Butler C, Hopstaken R, et al. Narrative review of primary tables. J Clin Epidemiol 2011;64:383–94. care point-of-care testing (POCT) and antibacterial use in respiratory 24. Riley RD, Simmonds MC, Look MP. Evidence synthesis combining tract infection (RTI). BMJ Open Respir Res 2015;2:e000086. individual patient data and aggregate data: a systematic review 10. Little P, Hobbs FDR, Moore M, et al. Clinical score and rapid antigen identified current practice and possible methods. J Clin Epidemiol detection test to guide antibiotic use for sore throats: randomised 2007;60:431.e1–12. controlled trial of PRISM (primary care streptococcal management). 25. Riley RD, Lambert PC, Abo-Zaid G. Meta-analysis of individual BMJ 2013;347:f5806. participant data: rationale, conduct, and reporting. BMJ 11. Hay AD, Redmond NM, Turnbull S, et al. Development and internal 2010;340:c221. validation of a clinical rule to improve antibiotic use in children 26. Stewart GB, Altman DG, Askie LM, et al. Statistical analysis of presenting to primary care with acute respiratory tract infection and individual participant data meta-analyses: a comparison of methods cough: a prognostic cohort study. Lancet Respir Med 2016;4:902–10. and recommendations for practice. PLoS One 2012;7:e46042. 12. Hay AD, Sterne JAC, Hood K, et al. Improving the Diagnosis and 27. Akl EA, Kahale LA, Agoritsas T, et al. Handling trial participants Treatment of Urinary Tract Infection in Young Children in Primary with missing outcome data when conducting a meta-analysis: a Care: Results from the DUTY Prospective Diagnostic Cohort Study. systematic survey of proposed approaches. Syst Rev 2015;4:98. The Annals of Family Medicine 2016;14:325–36. 28. General Accounting Office. Cross-design synthesis: a new strategy 13. Little P. Delayed prescribing of antibiotics for upper respiratory tract for medical effectiveness research. Washington DC, 1992. infection. BMJ 2005;331:301–2. 29. Rosenbaum PR, Rubin DB. The central role of the propensity score in 14. Peters S, Rowbotham S, Chisholm A, et al. Managing self-limiting observational studies for causal effects. Biometrika 1983;70:41–55. respiratory tract infections: a qualitative study of the usefulness of 30. D'Agostino RB. Propensity score methods for bias reduction in the the delayed prescribing strategy. Br J Gen Pract 2011;61:e579–89. comparison of a treatment to a non-randomized control group. Stat 15. Ryves R, Eyles C, Moore M, et al. Understanding the delayed Med 1998;17:2265–81. prescribing of antibiotics for respiratory tract infection in primary 31. Smith R, Coast J. The true cost of antimicrobial resistance. BMJ care: a qualitative analysis. BMJ Open 2016;6:e011882. 2013;346:f1493. 16. Spurling GK, Del Mar CB, Dooley L, et al. Delayed antibiotics for 32. Oppong R, Smith RD, Little P, et al. Cost effectiveness of amoxicillin respiratory infections. Cochrane Database Syst Rev 2013:CD004417. for lower respiratory tract infections in primary care: an economic 17. Berlin JA, Santanna J, Schmid CH, et al. Individual patient- versus evaluation accounting for the cost of antimicrobial resistance. Br J group-level data meta-regressions for the investigation of treatment Gen Pract 2016;66:e633–e639. effect modifiers: ecological bias rears its ugly head. Stat Med 33. Payment- What rates should be offered for involveent in research? 2002;21:371–87. National Health Research Involve website. http://www.​invo.​org.​uk/​ 18. Lambert PC, Sutton AJ, Abrams KR, et al. A comparison of summary posttypefaq/​payment-​what-​rates-​should-​be-​offered-​for-​involvement-​ patient-level covariates in meta-regression with individual patient in-​research. data meta-analysis. J Clin Epidemiol 2002;55:86–94. 34. Shamseer L, Moher D, Clarke M, et al. Preferred reporting items for 19. Spurling GKP, Del Mar CB, Dooley L, et al. Delayed antibiotic systematic review and meta-analysis protocols (PRISMA-P) 2015: prescriptions for respiratory infections. Cochrane Database Syst Rev elaboration and explanation. BMJ 2015;349:g7647. 2017;51:CD004417. 35. Tudur Smith C, Marcucci M, Nolan SJ, et al. Individual participant 20. Savovic J, Page M, Elbers R, et al. A revised tool for assessing data meta-analyses compared with meta-analyses based on risk of bias in randomized trials (RoB 2.0). Trials Conference: 4th aggregate data. Cochrane Database Syst Rev 2016;9:MR000007. Stuart B, et al. BMJ Open 2019;9:e026925. doi:10.1136/bmjopen-2018-026925 7