Methods

Intervention development and delivery

Theoretical framework

The research for intervention development drew on the framework that was used previously in the eCRT.16,39 McDermott et al.39 identified theoretical components that relate directly to effective implementation in health-care settings, identifying aspects of social cognitive theory41 and self-determination theory42 as possible influences on GP prescribing behaviour. The feedback received from the process evaluation for the eCRT is also incorporated.30

Social cognitive theory proposes that the environment plays a key role in influencing an individual’s behaviour.41 An individual’s belief in their ability to exercise control over their environment is one of the most important mechanisms involved in successful behaviour change. If an individual perceives their environment to be controllable and supportive, they will be more likely to succeed in performing the desired behaviour.41 In the present research, this suggests that interventions that are embedded into the consultation environment and become active during the flow of care are more likely to be successful. Social cognitive theory also proposes that the strength of an individual’s belief in his/her own ability to reach goals (i.e. their self-efficacy) functions as a key determinant of motivation for a specific behaviour. GPs’ self-efficacy has also been implicated as a predictor of intended adherence to recommendations for prescribing.43,44 Social cognitive theory also suggests that anticipated outcomes or ‘outcome expectancies’ of a behaviour influence the likelihood that it will be performed. Outcome expectancies relevant to prescribing decisions might include anticipated patient pressure45 or beliefs about risks and benefits associated with characteristics of a disease.46

Qualitative interviews in the previous study39 identified views that were consistent with self-determination theory.42 The theory proposes that behaviour change will occur and persist if it is autonomously motivated, in contrast to behaviour change that is brought about by perceived enforcement. GPs reported, for example, that they would be unlikely to engage with an intervention that they were forced to view or that they considered was attempting to control their behaviour, but in contrast they would be more inclined to engage with an intervention that they considered was there to support and aid them.

This study’s approach to developing the intervention aimed to create a controllable and supportive environment, increase self-efficacy, provide a meaningful rationale for engagement in the intervention, and promote expectations of positive outcomes, while reducing perceived negative risks, in order to support better GP adherence to prescribing recommendations.39

The description of the intervention was also guided by the TIDieR (Template for Intervention Description and Replication) checklist.47 This ensured that the intervention was sufficiently described to enable replication, future development and implementation into practice.

Design of intervention development study

A qualitative semistructured interview study was conducted to obtain empirical feedback on prototype interventions and to provide contextual information. Follow-up interviews were conducted with selected participants to further refine the study tools. Interviews were conducted by three experienced researchers, who met on a regular basis during data collection phase to compare and discuss their field notes.

Participant selection

Participants were GPs and nurse prescribers (NPs) employed by the NHS and working in non-trial general practices across England (London, Oxfordshire and Yorkshire). The minimum sample of 30 participants was set, aiming for saturation. GPs and NPs included a convenience sample and were recruited via clinical research networks, snowballing, mail-outs and social media [i.e. the young GP Facebook (Facebook, Inc., Menlo Park, CA, USA) community]. Those GPs and NPs interested in taking part were sent the participant information sheet and consent form. Participants were reimbursed for their time.

Procedure

Each interview followed an interview guide, which contained a list of questions and prompts. A semistructured interview was designed to identify factors likely to influence successful implementation of the intervention tools and discover likely responses to the proposed messages, in order to further inform development and aid refinement of the messages. During the interview participants were encouraged to interact with the study tools using a tablet as they would during their routine clinical practice. The initial section of the interview explored participants’ views on the webinar. Then the discussion moved onto the demonstration of the DSTs, followed by the discussion of practice-prescribing feedback. Field notes were made after each interview summarising the main points raised. All interviews were digitally recorded and then transcribed using professional services. Data collection and tools refinement were iterative processes and data collection continued until saturation was achieved. The study was progressively recruiting and interviewing GPs and introducing changes based on their comments until the last interviewed GPs considered no improvements could be made.

Analysis

The full analysis of transcripts was performed by two coders. A thematic analysis was conducted.48 An inductive approach to data analysis was used and themes were derived from the data. NVivo software (version 11; QSR International, Warrington, UK) was used to facilitate data analysis. Participants did not provide feedback on the findings; transcripts were not returned to participants for comments or corrections.

Patient and public involvement

We engaged with a primary care patient participation group (in Lewisham, south London). The trial procedure and the proposed intervention were presented to the group, and feedback and views were obtained on all aspects of the intervention, including the way in which messages would appear on GP screens, and information that would be presented to patients (such as patient information sheets). A member of this patient participation group continued in an advisory role throughout the intervention development, commenting on the intervention materials and giving feedback on the clarity, style and wording.

Cluster randomised trial

The trial was conducted in general practices contributing to the UK’s CPRD. The CPRD is one of the world’s largest databases of primary care EHRs; it includes data from general practices throughout the UK. CPRD data have been extensively evaluated and employed for epidemiological research.49 Recently, the CPRD has begun to be used as a resource for interventional research.16,50–52

Research ethics

The protocol for the study was approved by the NHS Health Research Authority National Research Ethics Service (NRES) Committee London–Dulwich (reference number 14/LO/1730). The protocol was also reviewed and approved by the CPRD Independent Scientific Advisory Committee (ISAC; reference number 14_130). An Independent Trial Steering Committee, and a Data Monitoring Committee, had oversight of the trial. The responsible partner at each practice gave written informed consent for the participation of the practice in the trial. The intervention was implemented at the general practice level and there was no requirement for individual patient consent because all individual treatment decisions were at the discretion of GPs and NPs. The study protocol has been reported previously,53 and the updated protocol can be accessed from the NIHR Health Technology Assessment (HTA) programme website (https://njl-admin.nihr.ac.uk/document/download/2010718; accessed on 1 November 2018).

Target population

The target population for this trial was the general population registered with general practices in the UK, including England, Scotland, Wales and Northern Ireland. The immediate participants in the research were health professionals who may issue prescriptions for antibiotics at UK general practices. Outcomes were evaluated using the anonymised EHRs for individual patients registered with UK general practices who presented with RTIs and were eligible to be prescribed antibiotics.

Inclusion/exclusion criteria

General practices were included in the trial if they were contributing up-to-standard data to the CPRD and consented to participation in the trial, and research governance approval for the study was in place. Data for individual participants were included if they were currently registered with the CPRD general practices. There were no exclusion criteria.

Recruitment and allocation

All practices contributing to the CPRD were invited to participate in the study from September 2015. Cluster randomisation was employed because intervention was at cluster level. Allocation was performed at King’s College London by Martin Gulliford and Toby Prevost, and CPRD staff responsible for recruiting practices to the trial, and communicating allocations (JS and KS), had no access to the allocation procedure. Practices were allocated to intervention and control trial arms using minimisation according to the MINIM program (www-users.york.ac.uk/∼mb55/guide/minim.htm; accessed October 2018),54 stratifying for CPRD region/country and pre-trial antibiotic-prescribing quartile. Practices were allocated in six waves: November 2015 (19 practices), January 2016 (12 practices), February 2016 (19 practices), June 2016 (18 practices), July 2016 (8 practices), and August 2016 (4 practices). For analysis, the six waves were combined into three periods: period 1, practices randomised in November 2015; period 2, January and February 2016; and period 3, June to August 2016. One practice allocated to the intervention trial arm withdrew from the CPRD before the intervention started and was excluded from further analysis because no data were available. The trial was stopped when the last general practices completed 12 months of follow-up.

Sample size calculations

Key measures included the consultation rate for RTIs, the antibiotic-prescribing rate for RTIs (both per 1000 registered patient-years) and the proportion of RTI consultations with antibiotics prescribed (%). As the primary aim of the study was to reduce antibiotic prescribing, antibiotic prescriptions for RTIs per 1000 registered patients was the primary outcome for the trial. Design parameters for the eCRT, which included participants aged 18–59 years, are shown in Table 2.

The estimation of sample size requirements was informed by the analysis of the CPRD data for the eCRT.16 The mean general practice-specific antibiotic-prescribing rate in the CPRD was 111.9 per 1000 patient-years [standard deviation (SD) 39.8 per 1000 patient-years]. The study employed analysis of covariance, with measures over 12 months before and after the intervention, giving a correlation coefficient of 0.82. Initially, it was aimed to recruit 120 CPRD general practices, which would have enabled detection of a 12 per 1000 reduction in antibiotic-prescribing rate. It was considered that a recruitment target of 120 practices was achievable because of the two previous cluster trials in the CPRD each recruiting more than 100 CPRD general practices.16,50

The REDUCE study depended on data from Vision (In Practice Systems Ltd, London, UK) practices contributing to the CPRD. Of the 385 Vision practices invited to take part, 21% agreed, 26% declined, 17% were undecided and 36% did not respond. During the recruitment phase, 26% (99 practices) changed their GP software system from Vision, precluding their involvement in the current study. The original target of 120 practices for this study was not therefore achievable, so the sample size calculation was revised. It was estimated that if there were 40 practices in each of two trial arms then, with an alpha of 0.05, there would be 80% power to detect an absolute reduction in antibiotic-prescribing rate of 15 per 1000 registered patient-years, and > 90% power to detect an absolute reduction of 3.5% in the proportion of RTI consultations at which antibiotics are prescribed. This revised sample size calculation was discussed and agreed with the HTA programme representatives at a HTA monitoring visit on 11 July 2016. The revised sample size calculation was also submitted to the CPRD ISAC and the Research Ethics Committee as an amendment.

Main measures

Outcomes were evaluated from the EHRs of registered patients at participating CPRD general practices. For each participant, the person-time at risk during the 12-month intervention period of the trial was evaluated. In addition, whether or not each patient had consultations for RTIs was evaluated. Self-limiting RTIs, including cough and bronchitis, otitis media, rhinosinusitis, sore throat and common cold, were evaluated using Read medical codes. Repeat consultations during the same episode were excluded using a 14-day time window. It was determined whether or not antibiotics were prescribed on the same date as the RTI consultation. Antibiotic prescribing was evaluated using product codes for antibiotics included in the British National Formulary (BNF),55 section 5.1, excluding antituberculous and antilepromatous drugs. Comorbidity was classified as present or absent using the ‘seasonal flu at-risk Read codes’ as employed in the NHS.56 Seasonal flu at-risk Read codes include diagnoses of significant heart, lung, renal, liver or neuromuscular disease, as well as cystic fibrosis, diabetes and immunosuppression or immunosuppressive treatment. Criteria based on age and pregnancy were not included.

Outcome measures

The proposed outcome measures for the trial are outlined in Table 3. The primary outcome measure was the rate of antibiotic prescribing for RTI per 1000 participant-years over the 12-month intervention period. Secondary outcome measures were the proportion of RTI consultations with antibiotics prescribed; the consultation rate for RTI per 1000 participant-years, and estimates for each of cough and bronchitis, colds, otitis media, rhinosinusitis and sore throat. In addition, total antibiotic prescribing for all indications was evaluated. Deferred prescriptions using the Read code for ‘deferred antibiotic prescription’ (8BP0) was evaluated. Health-care utilisation and costs, using methods reported previously, were also evaluated,57,58 obtaining utilisation estimates from the CPRD and costs of care from reference sources.59

Analysis plan

The trial protocol envisaged that a general practice-level analysis would be performed, with data aggregated to practice level.53 Subsequently, two considerations favoured an individual level for analysis of the primary, secondary and safety outcomes. First, there was attrition of trial general practices, with six practices withdrawing from the intervention trial arm and five from the control trial arm during the 12-month intervention period. This attrition was accounted for by practices leaving the Vision practice system that is employed by CPRD general practices. Data from practices that withdraw from the study were included in the intention-to-treat analysis but, because the conditions of interest have a pronounced seasonal distribution, bias might be introduced if comparable periods of time are not included for each practice. Second, a preliminary publication from the study group,31 as well as analyses for the trial DMC, drew greater attention to safety outcomes of the study. Analysis of safety outcomes requires consideration of individual-level covariates (including age and comorbidity), and these are also relevant for decisions to prescribe antibiotics. Thus, inferences are intended both at general practice level and at individual patient level. Consequently, an analysis was conducted of individual participant-level data as the primary analysis. The statistical analysis plan included prespecified subgroup analyses by age group, gender, comorbidity, region, type of infection and baseline antibiotic-prescribing quartile. Age group was categorised from 0 to 14 years, then in 10-year bands until ≥ 85 years. The subgroup effect was assessed statistically on this basis. The effect was summarised, more simply, in those aged 0–14 years being classed as children, and those aged 15–84 and ≥ 85 years being classed as adults.

The trial data set comprised full EHRs data for all participants who consulted with an RTI on one or more occasion during the trial baseline and intervention periods, together with denominator data for all patients registered at trial practices. In the primary intention-to-treat analysis, data for antibiotic prescriptions issued at consultations for RTIs and person-years at each practice were aggregated by age group, gender, comorbidity status and study quarter (as aggregation by month led to non-convergence) following the intervention start date. A random-effects Poisson model was fitted using the ‘hglm’ package in the R program (The R Foundation for Statistical Computing, Vienna, Austria).60,61 The dependent variable was a count of antibiotic prescriptions. Explanatory variables were trial arm, gender, age group, comorbidity status, region, study quarter and baseline antibiotic-prescribing rate. The baseline antibiotic-prescribing rate was included as an age-standardised rate for each practice, using the European standard population62 for reference. For practices that withdrew during the intervention period, baseline time was included pro rata. The period of randomisation was included, as well as the interaction of period with the baseline antibiotic-prescribing rate. A random effect was included for general practice. The offset was the log of person-years. The intervention effect was tested by considering the statistical significance of the effect of trial arm. Analyses were conducted using the same analytical framework for secondary outcomes. The log of respiratory consultations was employed as offset in the analysis of the antibiotic-prescribing proportion. Interaction terms were tested and prespecified subgroup analyses were conducted in the same framework. A forest plot was constructed. Results for the primary outcome are expressed as the number of antibiotic prescriptions for RTIs per 1000 registered patient-years, referred to as the antibiotic-prescribing rate; and, as secondary outcomes, the number of consultations for RTIs per 1000 registered patient-years (the RTI consultation rate) and the number of consultations for RTIs with antibiotics prescribed/total RTI consultations (%) (the antibiotic-prescribing proportion). Data for all antibiotic prescriptions for any indication were also analysed as a secondary outcome.

Cluster-level analyses were conducted, using general practice-specific aged-standardised rates as observations, as a secondary analysis. Scatterplots were presented in order to visualise the data, and rates were compared using analysis of covariance weighted for the number of person-years at the practice.

Analyses for safety outcomes

The incidence of safety outcomes was estimated as rates per 100,000 person-years. Safety outcomes were analysed by fitting a Poisson model and adjusting for age group, gender and comorbidity. A random effect for general practice was included for the most common outcome of pneumonia. The random effect for general practice was omitted for the remaining, more sparsely distributed outcomes, because of non-convergence.

Sensitivity analyses

Several sensitivity analyses were conducted for the primary outcome. The impact on estimates of a better-fitting model was evaluated, which included respiratory consultation rates as covariates as well as interactions of age group with covariates; the effect of fitting baseline age-specific rather than age-standardised rates as covariates was explored. Extended quasi-likelihood (EQL) and EQL1 estimation methods for ‘hglm’ were compared. The goodness of fit of the ‘hglm’ mode was evaluated and, as there was evidence of overdispersion, double hierarchical generalised linear models (DHGLMs) were evaluated using the ‘dhglm’ package in R.61 DHGLMs allow for additional random effects in the dispersion part of the model, as well as in the mean model for fixed effects, which can be used to model heavy-tailed distributions and may provide more robust analysis when outliers are present.61 An overdispersed Poisson model was fitted, incorporating an additional random effect in the dispersion part of the model.

Process evaluation

A process evaluation of the trial was conducted, including a questionnaire to participating GPs and analysis of data on intervention utilisation collected into DXS Point-of-Care™ (DXS International plc, Farnham, UK). The design of the process evaluation questionnaire was informed by criteria suggested by Linnan and Steckler63 for the process evaluation of public health interventions and research. These authors suggest that rigorous process evaluation should encompass five dimensions: fidelity (the extent to which the intervention was delivered as planned), reach (how many intended recipients took part in the intervention), dose (how much of the intended intervention was delivered), recruitment and context. The questionnaire included items concerning respondents’ experience with each component of the intervention. In addition, data on the utilisation of DSTs were also collected. Data were analysed for each practice on the number of times DSTs were viewed and for the number of information leaflets printed. As these two measures were highly correlated, the number of DST views was analysed. For each general practice the proportion (%) of RTI consultations at which DSTs were viewed was estimated. The intervention trial arm practices were divided into quartiles and a linear trend was evaluated across levels of DST utilisation for the primary outcome, adjusting for the same covariates as in the primary analysis.

Analysis of costs of health-care utilisation

There were some reasons why a full cost-effectiveness analysis was not appropriate for this study: reducing antibiotic prescribing may result in benefits to non-trial participants and the intervention may contribute to a reduction or delayed development of AMR. The latter has the properties of a public good whose benefits may be widespread in geography and time. A cost-effectiveness analysis would require a model that would relate trial outcomes to the wider impact of reducing AMR. This trial has the potential to generate information that could be used in a full cost-effectiveness analysis, but such an analysis will require wider research.

For these reasons, a cost analysis that focused on analysis of health-care utilisation was conducted. Previously developed and reported methods were employed to evaluate the costs of health-care utilisation,57,58,64 obtaining utilisation estimates from the CPRD and unit costs of care from reference sources.59 Health-care utilisation was evaluated from participants’ EHRs in the CPRD. Hospital Episode Statistics data were not used because linked data for the relevant time period were not yet available at the time of analysis. Analyses evaluated primary care utilisation, including consultations at the general practice, emergency consultations, home visits, out-of-hours visits and telephone consultations, and hospital utilisation (as noted in primary care records) including inpatient admissions, outpatient episodes, day cases and emergency episodes, as reported previously.57,58,64 The number of events was calculated for each patient. The costs of health-care utilisation were estimated using unit costs for 2017 from standard reference sources.59 The cost of a general practice consultation was £38, a telephone consultation was £27, an inpatient stay was £634, a home visit was £96.40, an outpatient visit was £137, an out-of-hours and emergency consultation was £38, a day-case procedure was £727, and an emergency referral was £137.59 Drug prescriptions were enumerated and prescription costs calculated by linking the product code for each prescription to the drug price from a dictionary obtained from RESIP Drug Database UK Ltd (Chertsey, UK).

The cost analysis focused on the question of whether or not the total costs of health-care utilisation for patients who consulted at least once for RTIs during the trial intervention period might be increased through intervention. This may be a concern if patients are more likely to re-consult if they are not given a prescription. The distribution of total costs was compared for participants in the control and intervention trial arms from the time of their first RTI consultation to the end of their general practice’s participation in the trial. As the analyses included participants who had consulted at least once, there were no cost zero values. As expected, the distribution of total costs was skewed. A hierarchical general linear model (HGLM) was fitted with a log-link and gamma error distribution, and a random effect was included to allow for correlation by general practice. Analyses were also adjusted for the fixed effects of gender, age group, comorbidity, region and trial period. The coefficient for trial arm was used to evaluate whether or not there was evidence that costs of health-care utilisation might be increased through interventions that aim to reduce antibiotic prescribing.

Changes to original trial protocol

This section outlines changes to the original trial protocol. Each modification was submitted to the CPRD ISAC as an amendment for approval; the Research Ethics Committee was also notified, when appropriate.

Definition of safety outcomes

Initially, safety outcomes were listed as pneumonia and lower RTIs, peritonsillar abscess, mastoiditis, and skin and bacterial infections. Following a review of the recent National Institute for Health and Care Excellence (NICE) guidance,65 intracranial abscess and empyema were included as safety outcomes, whereas bacterial infections and skin infections were omitted as lacking specificity. Following a meeting of the trial Data Monitoring Committee (DMC), held on 20 October 2015, it was recommended to include the following conditions as safety outcomes to be monitored during the trial and reported to the DMC: scarlet fever, pyelonephritis, septic arthritis, osteomyelitis, meningitis, toxic shock syndrome and septicaemia. Amendments to the protocol to evaluate these safety outcomes were submitted to the ISAC and these were approved.

Sample size calculation

As noted in Sample size calculations, an amendment was submitted for approval of the reduced recruitment target of 80 general practices, rather than the initially envisaged 120.

Statistical analysis plan

As noted in Analysis, the trial protocol envisaged that a general practice-level analysis would be performed, with data aggregated to practice level. However, two considerations favoured an individual level for the analysis of primary, secondary and safety outcomes. (1) There was significant attrition of CPRD trial practices, with six practices withdrawing from the intervention trial arm and five from the control trial arm. Although data from practices that withdraw from the study can be included in the analysis, bias may be introduced if comparable periods of time are not included for each practice, when the condition of interest has a seasonal distribution. (2) The cohort study,7 as well as analyses for the trial DMC, drew greater attention to safety outcomes of the study. Analysis of safety outcomes requires consideration of individual-level covariates (e.g. age and comorbidity), and these are also relevant for decisions to prescribe antibiotics. Consequently, it was decided to conduct an individual-level analysis as the primary analysis with a cluster-level analysis being considered as secondary.

The statistical analysis plan (SAP) was written by the trial statistician (AP) and the principal investigator (MG). The SAP was reviewed and approved by the independent DMC (meeting on 20 April 2017) and the Trial Steering Committee (meeting on 8 June 2017).

Population-based cohort study of safety outcomes

Material in this section is reproduced from Gulliford et al.31 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 3.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/3.0/.

The data source for the cohort study was the UK’s CPRD.49 For the present study, data were included for the 10-year period, from 2005 to 2014, before the start of the cluster trial. The CPRD included data for an open cohort of about 4.5 million registered patients during this period.

Definition of infective complications of respiratory infections

The number of first episodes of infective complications were evaluated in the entire registered population of patients in the CPRD from 2005 to 2014. Infective complications of RTIs were defined using Read medical codes recorded in participants’ EHRs. EHR data include diagnoses recorded at primary care consultations and home visits. In addition, the CPRD referral file includes coded data for hospital referrals and hospital discharges. Analyses evaluated pneumonia (57 codes), empyema (14 codes), peritonsillar abscess (five codes), mastoiditis (13 codes), bacterial meningitis (19 codes) and intracranial abscess (14 codes). Codes for pneumonia were drawn from section H2 of the Read code classification, which includes codes for pneumonia and influenza. Codes were included if they indicated the presence of pneumonia without a viral aetiology. Bacterial meningitis included codes for meningococcal meningitis, meningococcal septicaemia, pneumococcal meningitis and Haemophilus meningitis, as well as unspecified bacterial meningitis. Data were extracted for all participants with records of infective complications from 2005 to 2014. Incident events were defined as the first record of an event in a participant that was recorded > 12 months after the start of the participant’s CPRD record. Gender, year and age group were included as individual-level covariates. Nine 10-year age groups were employed, with categories of 0–14 years and ≥ 85 years. Incident events were aggregated by year, age group, gender and general practice. Person-time for the registered CPRD population was estimated by year, age group, gender and general practice in order to estimate rates of infective complications. Cluster-level covariates included CPRD region, with 10 regions in England, as well as regions in Wales, Scotland and Northern Ireland. Deprivation quintile was included, based on general practice-level data for indices of multiple deprivation score66 for England and equivalent scores in Wales, Scotland and Northern Ireland. Linked data were provided by the CPRD.67 The distribution of deprivation quintiles is not expected to be equivalent between countries.

Definition of respiratory consultation and antibiotic-prescribing rates

Age-standardised measures were estimated for RTI consultations and antibiotic prescribing, as reported previously.7 For each CPRD general practice, the rate of RTI consultations per 1000 registered patients was estimated; the antibiotic-prescribing rate for RTI per 1000 registered patients, and the proportion (%) of RTI consultations with antibiotics prescribed. These prescribing measures were estimated on a sample of CPRD data because it was not feasible, and the licence did not allow us, to perform the analysis on the entire CPRD database. Participants were sampled from all acceptable patients included in the CPRD. A random sample of 75 currently registered patients was drawn without replacement from each year from 2005 to 2014. This gave a maximum sample of 750 participants, with up to 7500 person-years of observation per practice. The study aimed to achieve a total sample of < 0.5 million and the total sample for analysis was 411,226 participants from 643 general practices. This allowed the estimation of practice-specific proportions with a margin of error of 1%. For participants in the sample, person-years were estimated as the denominator from the start of CPRD registration, or 1 January 2005, to the end of the participant’s CPRD record, or 31 December 2014. Self-limiting RTIs were identified using medical codes recorded during general practice consultations. These were classified into five groups following the recommendations of NICE:6 colds and upper respiratory infections; sore throat, including pharyngitis and laryngitis; cough and acute bronchitis; otitis media; and rhinosinusitis. Acute bronchitis was included because current recommendations are to avoid antibiotic treatment.6 Consultations for RTIs were identified and first consultations within a 14-day time window were selected. Data for participants aged ≥ 100 years were excluded. Antibiotic prescriptions issued on the same day as respiratory consultations were identified. For each general practice rates of consultations for respiratory infections per 1000 person-years (i.e. the RTI consultation rate), rates of antibiotic prescribing for RTIs per 1000 person-years (i.e. the antibiotic-prescribing rate) and proportions (per cent) of RTI consultations with antibiotics prescribed (i.e. the antibiotic-prescribing proportion) were estimated. Rates and proportions were standardised for age and gender using the 2013 European standard population.62 After excluding practices with insufficient data, because of short periods of CPRD contribution rates were estimated for 610 CPRD general practices.

Analysis

Age-standardised incidence rates (per 100,000) were estimated by year for each safety outcome, and trends over time were estimated using a linear model. In the final stage of the analysis, the number of infective complications, with person-years at risk, were evaluated in relation to general practice-specific rates of RTI consultations and antibiotic prescribing. Mixed-effects Poisson models were fitted using the ‘hglm’ package60 in the R program.68 General practice was fitted as a random effect. The log of person-years was included as offset. Fixed effects included gender, year, age group, region and deprivation quintile. The association of age-standardised RTI consultation rate with rates of infective complications was evaluated. The association of the antibiotic-prescribing rate, and the antibiotic-prescribing proportion, with infective complication rates were evaluated after adjusting for the RTI consultation rate. Incident rate ratios (IRRs) [along with 95% confidence intervals (CIs)] were estimated for each quartile of RTI consultation rate, antibiotic-prescribing rate or antibiotic-prescribing proportion using the lowest quartile for reference. The RTI consultation rate, antibiotic-prescribing rate or antibiotic-prescribing proportion were also fitted as continuous predictors and IRRs were estimated per 10-unit change in the predictor. Whether or not the addition of quadratic terms improved goodness of fit was evaluated. As there were small numbers of events for intracranial abscess and mixed-effects models did not converge, the random effect for general practice was omitted for this outcome. Regression models were not fitted for Lemierre syndrome because this was extremely rare. The ‘ggplot2’69 and ‘forestplot’70 packages in R were used to present the results.

In order to present the clinical implications of these findings, the number of events expected in a general practice with 7000 patients, the general practice mean list size for England during 10 years of follow-up was calculated. The median (95% range) was used for the RTI consultation rate to estimate the expected number of RTI consultations. The disease incidence and distribution of antibiotic-prescribing proportion for the highest-prescribing quartile was used to estimate expected numbers of complications and antibiotic prescriptions. The relative risk increase for a 10% change in antibiotic prescribing was used, from the Poisson model, to estimate the expected change in number of infective complications.

Research ethics

The protocol for the study was approved by the CPRD ISAC (reference 14_130A2). The CPRD has broad NRES ethics approval for observational research studies.