Antibiotics for lower respiratory tract infection in children presenting in primary care: ARTIC-PC RCT

Summary of trial design

This was a randomised, placebo-controlled, parallel-group trial of amoxicillin for children presenting in primary care with uncomplicated LRTI, powered to detect benefits important enough to be worth prescribing antibiotics for, both overall and among key clinical subgroups in whom antibiotic prescribing is very common. We also included qualitative (see Chapter 3), microbiological (see Chapter 4) and economic (see Chapter 4) studies. Patients who were not randomised as a result of patient or clinician beliefs or preference were invited to participate in an observational study in which the same measures and outcomes were collected (see Chapter 2).

Ethics

The trial protocol was approved by the South West – Central Bristol Research Ethics Committee (reference 15/SW/0300), and the protocol can be found on the project web page of the NIHR Journals Library website (URL: www.journalslibrary.nihr.ac.uk/programmes/hta/133464/#/).

Intervention

The intervention was 50 mg/kg/day oral amoxicillin in three divided doses for 7 days, or placebo.

Amoxicillin was chosen as it is the first-choice antibiotic for LRTI and, at current levels of intermediate resistance, should cover most susceptible organisms. 25,37 The rationale for the dose is in line with guidance from the British National Formulary for children, and this was supported by a Monte Carlo simulation to achieve a minimal inhibitory concentration of around 1.5 mg/l to cover Haemophilus influenzae as well as intermediate-resistant pneumococci for 90% of the intended population. 25 We estimated that, to achieve bacterial eradication, the blood concentration of amoxicillin needs to be above the minimal inhibitory concentration for at least 5 days. A 7-day course was chosen to allow for poor adherence26 and on pragmatic grounds to match current practice at the time the study commenced to achieve greater clinician and parent acceptability; similar consensus had been required for the previous trial in adults. 25

Inclusion criteria

Included were children aged between 6 months and 12 years presenting to primary care with an acute LRTI, defined in several previous cohorts and trials as an acute cough as the predominant symptom and judged by the GP to be infective in origin, lasting < 21 days, and with other symptoms or signs localising to the lower respiratory tract (e.g. shortness of breath, sputum, pain). 27–30 These inclusion criteria reflect the clinical criteria used in daily practice to diagnose acute bronchitis31 and were used in the Cochrane review;17 they are also the key drivers of prescribing. 2,11,22

Exclusion criteria

Exclusion criteria were cough as judged by the clinician to have a non-infectious aetiology (e.g. hay fever or asthma) or to be almost certain viral aetiology (croup, for which antibiotics are not commonly prescribed); immune compromise; and antibiotic use in previous 30 days. Children with suspected pneumonia based on clinical examination or who were very severely ill as judged by the GP were also excluded from the trial, but they were eligible to enter a parallel observational study.

Consent

The parent or guardian of the child provided written consent. Children who were able to understand the study read an age-appropriate patient information leaflet and signed an age-appropriate assent form.

Randomisation

Parents and children who consented to the study and agreed to randomisation received oral suspension of either amoxicillin or placebo, randomised in a 1 : 1 ratio. Investigational Medicinal Product (IMP) packs were indistinguishable in appearance and packaging, and each pack was labelled with a unique identification number to maintain allocation concealment. A computer-generated random number list was provided by an independent statistician and kept only by the IMP manufacturer. Random block sizes of two to four packs were used, with practice sites receiving whole blocks. Investigators randomised and dispensed by selecting the next sequentially numbered IMP pack.

Data collection

Outcomes

Statistical analysis

Cox regression was used for the primary outcome and for total symptom duration, adjusting for age, baseline symptom severity, prior duration of illness and comorbidity. The proportional hazards assumptions was assessed visually using the Kaplan–Meier curve. Linear regression was used for symptom severity, and logistic regression was used for reconsultation, progression of illness and side effects, adjusting for the same baseline covariates as in the primary analysis. To aid interpretation, risk ratios were also calculated for the binary outcomes using a log-binomial model. Analysis was by intention to treat, as randomised regardless of non-adherence or protocol deviations. Multiple imputation was used as the primary analysis, agreed with the funder and documented in the statistical analysis plan, which superseded the protocol (it should be noted that the protocol incorrectly stated that multiple imputation was a secondary analysis). Multiple imputation comprised all variables from the analysis model and any predictors of missingness, and used 100 imputations. A complete-case analysis was performed as a sensitivity analysis. Prespecified subgroup analyses were carried out on chest signs (alpha = 0.05), sputum/rattly chest, history of fever, physician rating of unwell, shortness of breath, oxygen saturation below 95% and STARWAVe clinical prediction rule for hospitalisation39 (all alpha = 0.01). Analyses were carried out in Stata^® 16 (StataCorp LP, College Station, TX, USA) and IBM SPSS Statistics 26 (IBM Corporation, Armonk, NY, USA), with analysts blinded to the randomisation group.

Follow-up

Data were available on symptom duration for 317 participants (73.3%), on symptom severity for 298 participants (69.0%), on reconsultation with new or worsening symptoms for 401 participants (92.8%), on children who required hospital assessment for 415 participants (96.1%) and on side effects for 310 participants (71.8%) (see Table 3). Most of the symptom data came from the diaries, and the estimates of median duration of illness were a day longer for both diary (placebo, 6 days; antibiotic, 5 days) and telephone (placebo, 7 days; antibiotic, 6 days). The missingness of diary data (including adherence data) was associated particularly with parental qualification (those with higher qualifications were less likely to have missing data), and hence this variable was included in the imputation model.

Primary outcome

The duration of moderately bad or worse symptoms was similar in both groups [antibiotic, median 5 (IQR 4–11) days; placebo, median 6 (IQR 4–15) days; HR 1.13, 95% confidence interval (CI) 0.90 to 1.42)] (see Table 4 and Figure 2). None of the prespecified clinical subgroups or the additional post hoc exploratory subgroups (low oxygen saturation; STARWAVe categories) modified the effect of treatment on the duration of moderately bad or worse symptoms (see Table 5).

FIGURE 2.

Kaplan–Meier curve of duration of moderately bad or worse symptoms in days.

Secondary outcomes

There was a small but statistically significant difference between the groups in symptom severity on days 2–4 after seeing the doctor (antibiotics 1.8, placebo 2.1; mean difference–0.28, 95% CI –0.51 to –0.04) (see Table 4), equivalent to less than 1 in 3 children rating symptoms as slight rather than very little problem.

The duration of symptoms until very little problem was also similar between the groups (antibiotics median 7, IQR 4–17; placebo median 8, IQR 5–20), with no significant difference between the groups (HR 1.09, 95% CI 0.86 to 1.38) (see Table 4).

The number of participants reconsulting with new or worsening symptoms was 60 (29.7%) in the antibiotics group compared with 76 (38.2%) in the placebo group (risk ratio 0.80, 95% CI 0.58 to 1.05) (see Table 4). Few children needed hospital assessment (antibiotics, n = 5, 2.4%, placebo, n = 4, 2.0%; risk ratio 1.23, 95% CI 0.32 to 4.44) (see Table 4). The number of participants experiencing side effects was similar (antibiotics, n = 60, 38.2%, placebo, n = 52, 34.0%; risk ratio 1.20, 95% CI 0.87 to 1.55) (see Table 4).

Adherence

A limitation of the study is that only 232 (53.7%) participants provided data on adherence to medication. Among those who reported adherence, over the first 5 days it was most common to report 14 or 15 doses being taken (out of a maximum of 15; Figure 3), and the majority (95%) started taking their medication on day 1 (see Figure 3). Adherence was maintained in the antibiotics group over days 1–5 but decreased gradually to 84% by day 5 in the placebo group. A total of 98 out of 119 (82.4%) participants in the antibiotics group and 87 out of 113 (77.0%) participants in the placebo group took at least 11 doses of medication over days 1–5. A per-protocol analysis suggested no statistically significant difference in the duration of moderately bad or worse symptoms (HR 1.06, 95% CI 0.77 to 1.46). Reported adherence was higher for longer prior duration of illness (odds ratio 1.08, 95% CI 1.01 to 1.16), adjusting for group and other prespecified covariates. Among children whose parents thought they were receiving antibiotics, 68 (85%) adhered to their medication, and among children whose parents thought they were receiving placebo, 82 (82%) adhered to their medication. There was little evidence for clustering by GP practice (intracluster correlation 0.01, 95% CI 0.00 to 0.99).

FIGURE 3.

Number of doses of medication taken in total over days 1–5 by treatment group: (a) placebo; and (b) antibiotics.

Strengths and limitations

The study is one of the very few to report on the effectiveness of prescribing antibiotics among younger children presenting with chest infections in primary care, and the parents of two-thirds of eligible children agreed to be randomised in the trial. The study was designed to detect a 3-day improvement in the duration of more severe symptoms in key clinical subgroups. Our patient and public involvement and engagement (PPIE) group regarded the duration of illness as the most important outcome, and they judged that, given the problem of antibiotic resistance, a duration of 3 days was important enough to warrant prescribing an antibiotic. 2,3 Our prior data suggested an illness duration of 21 days, and so 3 days would represent a 15% improvement,15 but from the current study 3 days represents a 25% improvement for an illness lasting 12 days. Although the IQR was slightly wider in the placebo group, the Kaplan–Meier curves did not separate significantly. We used the most patient-relevant outcome (parent-reported symptoms) and documented illness progression. The study confirmed that illness progression is uncommon, but the study was not specifically powered to assess the progression of illness or reconsultations. In the final sample, imputed and complete cases analyses were adequately powered overall and for the subgroups, but slightly underpowered for the complete-case analysis in the chest signs subgroup, owing to, in part, slightly fewer children than expected having chest signs and the COVID-19 pandemic prematurely ending recruitment. However, the estimates for the primary outcome for complete and imputed cases in the chest signs subgroup were very similar (6 days in antibiotics and placebo groups in both analyses) and the HRs were near unity (0.91 and 0.97, respectively), and the upper 95% CIs of the HRs (1.41 and 1.43) suggest that the ‘true’ benefit for children with chest signs is unlikely to be more than 2 days, that is, not likely to be very important clinically. Although the study was placebo controlled, it was at the pragmatic end of the spectrum in that there was no close monitoring of parents and children; parents behaved as they might in practice as to whether or not they gave their child medication, and, although we had limited data on medication adherence, a per-protocol analysis provided similar estimates to those for the total trial population. The antibiotic (amoxicillin) was chosen as it is the first-choice antibiotic in UK national guidance for use in LRTIs among children. 37 Compared with representative observational cohorts, our trial population had a more severe clinical presentation,39 so we may have overestimated the impact of antibiotics.

Interpretation and comparison with previous literature

Only one trial in the Cochrane review of antibiotics for acute bronchitis included young children aged ≥ 3 years presenting with uncomplicated acute chest infections. 15,17 In that trial there were only 100 children aged ≤ 12 years, and the estimate of immediate antibiotics compared with no offer of antibiotics on symptom duration (HR 1.00) and symptom severity (a mean reduction of –0.3 points on a scale of 0–6 points) was similar to the non-significant result of the whole trial cohort. 15 These results are consistent with the results of the current study. As in the data from adults,15 the duration of illness in the current study is dominated by the duration of cough. A Cochrane review40 found inconclusive evidence of the effect of antibiotics in preventing RTIs, but a more recent trial of azithromycin used in early infections was effective in preventing severe illness among preschool children with recurrent infection41 (from 8% to 5%), although concern remains about the longer-term effects on antibiotic resistance from the use of long-acting macrolides. 6 A placebo-controlled trial42 of pneumonia in young children in a developing country found low failure rates in both the placebo (4.9%) and antibiotics (2.6%) groups.

Our results suggest that antibiotics do not provide a clinically important benefit, on average, for symptom reduction or symptom severity. The question of there potentially being children who receive a meaningful benefit that is imperceptible among the large numbers of children who receive no benefit remains. We explored this hypothesis by conducting subgroup analyses in five prespecified subgroups. Our subgroup analyses results suggest that none of the groups we specified is likely to achieve substantial benefits in terms of symptomatic improvement from antibiotics, although we did not have the power to exclude more moderate benefits. On the other hand, the average benefit from antibiotics in the general population may be even lower than our findings suggest. We had significantly fewer children with a very low risk STARWAVe score in this study than did the STARWAVe observational study, which recruited a representative sample of children with RTIs from the population (children with a very low risk score 67%, current trial 54%; see Table 1). 39 This suggests that the present trial successfully recruited more unwell children, in whom antibiotics might be expected to be more effective, and that the average impact of antibiotics in a more generalisable low-risk population is likely to be even lower than reported here. There was no evidence of selective benefit among children in whom pathogenic bacteria were isolated, which could be due to high carriage rates among children rather than true infection. The estimates of resource use suggest that not only are consultations, referral and hospitalisation costs significant,43 but societal costs are high. Antibiotic prescribing was not associated with health or societal resource savings and, if anything, resulted in slightly higher costs. If the costs of antibiotic resistance were included, then the adverse impact on health and societal resource use would be even greater. 7

Sample size: developing a new prognostic model

The standard rules used by statisticians suggest that the number of variables that can be assessed robustly in a prognostic model is 1 variable per 10 cases, so the three-variable model should be adequately powered. 53 However, the traditional rule of thumb has been questioned, and making the recommended assumptions53 (a margin of error of ≤ 0.05, mean absolute prediction error of ≤ 0.05, shrinkage factor of < 10% and expected optimism factor of 0.05) with an expected outcome event rate of 5% suggests that 566 participants were required.

Statistical analysis

We planned to control for confounding by indication by using inverse probability of treatment weighting using propensity scores in each of the regression models. However, the inverse probability of treatment weighting approach did not achieve good balance on the key covariates, whereas stratification by propensity score improved balance and, therefore, was used in analysing both the observational data and the combined data set that included both observational and trial data. Given the numbers of missing data, we imputed missing data using a chained equations model with 100 imputations.

We used a backwards-fitting approach to model the predictors of progression of illness. For the first, most inclusive, model, we retained in the multivariable model those variables with a p-value of < 0.20. Given the danger of overfitting, we also assessed another model for the variables that were significant at a p-value of < 0.05 in multivariate analysis in the first model, and finally for the three most significant variables. We explored whether or not progression of illness could be predicted from the following baseline characteristics: age, baseline severity, heart rate, respiratory rate, temperature, duration of illness prior to consultation, gender, comorbid conditions, history of asthma, chest signs, feeling generally unwell, oxygen saturation, sputum/rattly chest, vomiting, dry cough, chills, diarrhoea, disturbed sleep and passing urine less often. We present the discrimination of the model in the area under the receiver operator curve (AUROC), bootstrapping the estimates to limit overly optimistic estimates due to overfitting, and as a more efficient approach to internal validation rather than using split samples. 55 Model calibration was assessed with a Hosmer–Lemeshow test. We also tested the STARWAVE score using the three STARWAVe classifications of low, normal or high risk. 39

Proportions followed up for key outcome measures

In the observational study, the duration of illness and illness severity in days 2–4 following the consultation were recorded for 232 out of 312 (74.4%) participants. Reconsultation was available for 271 (86.9%), progression of illness for 290 (92.9%) and side effects for 228 (73.1%) participants.

In the combined data set, 549 out of 744 participants (73.8%) reported the duration and severity of illness. Reconsultation was recorded for 672 (90.3%), progression of illness for 705 (94.8%) and side effects for 538 (72.3%) participants.

Clinical characteristics

As expected, the number of children with more severe clinical features (see Table 13) was much larger in the antibiotics group than in the no-antibiotics group. Children in the antibiotics group had more severe baseline symptoms (1.8 score and 1.5 score, respectively), more abnormal chest signs (81% vs. 24%), sputum production (87% vs. 70%), history of fever (91% vs. 64%), cases of feeling unwell (81% vs. 51%), shortness of breath (70% vs. 36%) and low oxygen saturation (21% vs. 7%) than children in the no-antibiotics group.

Propensity scores

The differences between the antibiotics group and no-antibiotics group are shown in Figure 5 before and after adjustment using propensity scores; this demonstrates that, although there was a major impact of adjustment, there is nevertheless likely to be some residual confounding.

FIGURE 5.

Standardised differences in baseline characteristics between antibiotics and no-antibiotics groups in the observational data set and after adjusting for confounding by indication using the propensity score.

Primary and secondary outcomes

For the combined trial and observational data set, the estimates for primary and secondary outcomes were similar to the trial outcomes (see Table 14). The effect of antibiotic treatment on severity of symptoms was not significant. The apparent non significant increase in the progression of illness in the antibiotics group is very likely to be due to inadequate control of confounding by indication. The only outcome for which there was a statistically significant comparison between the antibiotics and no-antibiotics groups was side effects, which were higher in the antibiotics group.

Subgroups

After controlling for confounding with propensity scores, none of the prespecified subgroups had statistically significant interaction terms in either the observational data set or the combined data set (see Table 15). When selecting each subgroup, we found a suggestion of benefit among both children with productive sputum and children with fever.

Prognostic model

Twenty-nine children (4%) in the cohort had illness progression necessitating attendance or admission to hospital (details of the illnesses are in Table 18). We tested the STARWAVe score using the three classifications. The calibration was excellent, with a Hosmer–Lemeshow test p-value of 0.9847, but the AUROC was modest, at 0.66 (95% CI 0.50 to 0.77). As STARWAVe was developed to predict the progression of illness requiring overnight hospital admission (11 of the children in the current data set), we separately tested test performance for this outcome; the AUROC for STARWAVe in predicting the need for overnight hospital admission was 0.70 (0.56 to 0.84; Hosmer–Lemeshow p-value 1.00).

Given the modest discrimination of the STARWAVe model, we developed new models. The predictors of the progression of illness are shown in Table 16. We retained 8 variables: baseline severity, difference in respiratory rate from normal for age, duration of illness prior to consultation, low oxygen saturation, sputum/rattly chest, passing urine less often and diarrhoea as variables in the model. This model had an AUROC of 0.83 (95% CI 0.74 to 0.92); model calibration was good, with a Hosmer–Lemeshow test p-value of 0.22. A reduced model using only the three significant predictors from the eight variate model (difference in respiratory rate from normal for age, sputum and low oxygen saturation) had an AUROC of 0.81 (95% CI 0.71 to 0.91; Hosmer–Lemeshow statistic p-value 0.86) (see Table 17). We did not include antibiotic prescription in the first predictive model owing to the observed inverse association between antibiotic prescribing and the progression of illness (very likely to be due to confounding by indication); nevertheless, we also assessed the model after including antibiotic prescription and the same variables were included. We also looked at the discrimination of the five-item model for the 11 children requiring overnight admissions: the AUROC was 0.88 (95% CI 0.78 to 0.97; Hosmer–Lemeshow p-value 0.74).

We also converted the three-item model [respiratory rate (breaths per minute), oxygen saturation < 95%, sputum] to a score by multiplying the beta-coefficients by 10 and rounding to the nearest integer (score = 46 + difference in respiratory rate from normal for age + 14 × low oxygen saturation + 18 × sputum). Scores range from 19 to 102, and the AUROC was 0.78 (95% CI 0.68 to 0.88) for progression of illness and 0.86 (95% CI 0.72 to 1.00) for hospital admission.

Limitations

The method of controlling for confounding by indication needed to be adapted to improve the balance between the groups, and there was evidence of some residual confounding by indication for some of the outcomes, particularly progression of illness. It is also likely that residual confounding by indication contributes to the much higher rate of apparent ‘side effects’ seen in the antibiotics group, as we know that diarrhoea, vomiting and skin rash occur commonly in the placebo group (i.e. as part of the illness for both children and adults). 25 Therefore, the more common ‘side effects’ in the antibiotics group in the observational data may reflect illness severity rather than side effects per se. It is also possible that there is greater attribution and monitoring of known side effects when parents know that their child is receiving antibiotics. The differences in clinical characteristics between observational and trial data sets cannot be attributed just to clinical decision-making, as the range of primary care settings was different: many of the observational patients came from primary care sites that were less typical of routine general practice and were not able to recruit to the trial (e.g. A&E). The prognostic model was limited by the relatively few children in whom illness progressed (and so we may not have had sufficient power to include all of the relevant variables), although having more significant variables would increase the discriminatory power of the model. Although bootstrapping was used to limit the problem of overfitting, and the reduced model with five or three variables also provided reasonable estimates of discrimination, external validation of these models in a separate sample will be needed.

Main results in context of other literature

Children given antibiotics in the observational study had much more severe clinical presentations than children not given antibiotics, matching the trends in the much larger STARWAVe cohort. 12,39 The children in our trial cohort were more severely affected than the children in the STARWAVe cohort, and that trend is even more apparent in the children contributing to the observational data: there were high percentages of children given antibiotics with sputum production (87% of children, compared with 63% in STARWAVe), fever (91% of children, compared with 75% in STARWAVe) and shortness of breath (70% of children, compared with 46% in STARWAVe).

Despite major differences in clinical presentation between children given and children not given antibiotics, we found that, when controlling for the propensity to prescribe antibiotics, the main outcomes of the combined trial and observational data were very similar to those of the ‘pure’ trial data; overall, there were 1–2 days’ difference between the groups, both for the duration of moderately bad or worse symptoms and the duration of symptoms until very little or no problem. In the subgroup analyses, although there was some suggestive evidence of differences for some subgroups, the interaction terms were not statistically significant. Among children with productive sputum or fever, the lower CIs for the HRs were only just above the null, so it is possible that these are chance findings. The estimates of benefit for both of the above subgroups were also not clinically important (neither subgroup had a difference in symptom duration of more than 2 days).

The final prognostic model to predict the progression of illness included eight variables and demonstrated excellent discrimination (AUROC 0.85) and good calibration. However, it included rather different variables from the STARWAVe model:12,39 the models had only age and prior illness duration in common (STARWAVe included age < 2 years, current asthma, illness duration of ≤ 3 days, parent-reported moderate or severe vomiting in the previous 24 hours, parent-reported severe fever in the previous 24 hours or a body temperature of ≥ 37.8 °C at presentation, clinician-reported intercostal or subcostal recession, and clinician-reported wheeze on auscultation). Thus, it is perhaps not surprising that the STARWAVe model had lower discrimination in the current population (AUROC 0.65), and the difference probably reflects the more severely affected group of children recruited to the current study. Another difference from STARWAVe is that data on pulse oximetry were available for only < 50% of children in the STARWAVe cohort, and if more had been available pulse oximetry might have been included in STARWAVe. In addition, intercostal recession was not measured in the current study, and so it is possible the estimated discrimination might be improved further; and if the outcome is overnight hospital admission the discrimination of both the new model and STARWAVe improve further. The prognostic model could have clinical utility in that all of the variables are easily documented in routine consultations, so they could possibly be used by clinicians to predict adverse outcome when seeing more unwell children, potentially guiding management decisions for antibiotic prescribing and/or for follow-up. 57

Conclusion

The observational study included children from a slightly different range of settings from the trial so direct comparisons are not possible. Although children given antibiotics in the observational settings were much more severely affected, there was no evidence of greater effectiveness of antibiotics in modifying symptom resolution. We have developed simple clinical scores to identify the minority of children whose illness is more likely to progress sufficiently to require hospital assessment. A simple clinical score requires external validation but could potentially be used to guide clinical management.

Introduction

A large proportion of consultations in primary care involve parents consulting with children with symptoms of a RTI, and antibiotics are frequently prescribed. 45 The overuse of antibiotics is a cause for concern in terms of both the growing threat of antimicrobial resistance (AMR)2 and the cost burden on the NHS. 11,13,34

Research has shown that parents vary in their confidence and self-efficacy when it comes to assessing the symptoms of a RTI58 and consult a primary care service owing to perceptions of illness severity, uncertainty about appropriate management and a desire for a clinical examination and reassurance from a trusted medical professional. 19,20,58–60 Although some parents may consult with an expectation for antibiotics, studies have shown that these expectations tend to be overestimated by clinicians during the consultation and are a significant contributor to unnecessary prescribing. 61,62 Furthermore, many parents are said to be accepting of a ‘no-antibiotic’ treatment outcome as long as they leave the consultation feeling reassured and have a plan for symptomatic relief at home. 59,63,64 Clinician decision-making to prescribe antibiotics is influenced by perceived clinical need,19,62,65 but more significantly by clinical uncertainty about future illness deterioration and fears of missing something more serious in a child. 19,20,58,66–68

Two qualitative studies with parents and clinicians were incorporated into ARTIC-PC to explore their views and experiences of managing a child with symptoms of LRTI, as well as their views of participating overall.

Methods

Two qualitative studies using semistructured telephone interviews were carried out with parents and clinicians who participated in ARTIC-PC. Parents who had given written informed consent to take part in an interview at the time of recruitment to ARTIC-PC were invited to participate in a telephone interview. Participants were purposively sampled by their participation either in the randomised controlled trial (RCT) or in the observational study, to represent a range of views/experiences following participation, to explore parental reasons for taking part in both aspects of the research and use the findings to optimise study procedures and improve recruitment as part of the ongoing trial. The topic guide included questions about parental motivations for taking part in the research.

Parent interviews took place between November 2016 and July 2017. Clinicians taking part in recruiting participants into the study were also invited to take part in a telephone interview. We aimed to recruit 10–15 clinicians from different practices taking part in the trial. The interviews took place between August 2017 and February 2018.

Catherine Woods conducted semistructured telephone interviews with parents and clinicians. Telephone interviews were deemed suitable for both types of participants, and it was thought that this format would encourage participation by making it easier for parents and clinicians to fit the interview around childcare responsibilities and clinical practice, respectively. The parent interview guide explored views and experiences of managing a child with RTI, reasons for consulting, and views of the study procedures and why they chose to take part in the trial or the observational study. Interviews with parents lasted between 20 and 50 minutes, with an average duration of 33 minutes. The clinician interview guide explored views and experiences of managing children with symptoms of a RTI in primary care as well their views of taking part in the trial. The interviews with clinicians lasted between 25 and 50 minutes, with an average duration of 35 minutes.

All interviews were audio-recorded, transcribed verbatim and analysed using inductive thematic analysis. 69 The parent interviews were analysed by Catherine Woods and second-coded by a medical student (Zöe Morrice) on a 6-week summer placement in the Primary Care Research Centre in Southampton. The codes were compared and refined in meetings with Catherine Woods and Zöe Morrice and the qualitative work package lead Geraldine Leydon; and the final thematic framework which captured salient themes across the entire data set was developed by Catherine Woods. The clinician interviews were coded by Catherine Woods. Analyses were facilitated by NVivo 11 software.

Theme 1: reasons for consulting

Theme 2: parent perspectives and understandings about antibiotic use and antibiotic resistance

Theme 3: views on taking part in the trial and its procedures

Theme 1: antibiotic prescribing for children with lower respiratory tract infection and other common infections

Theme 2: views about antimicrobial resistance

Theme 3: positive views of taking part in the trial and its procedures

The trial procedures were clear and easy to follow

The clinicians thought that all of the trial procedures, including the study packs, paperwork (i.e. participant information sheets and consent forms) and criteria for identifying suitable children, were clear and easy to follow. The clinicians also felt comfortable explaining the concept of a placebo to parents at the time of recruitment:

Interviewer: How have you found things like the information sheets and things like that that we’ve given?

Nurse 4: I think they’re fine. I’ve not had a problem. I think they do explain things clearly.

Nurse 4

Interviewer: How easy is it to explain this [concept of placebo] to patients?

GP 1: I haven’t really found a particular problem, you just have to explain that, in order to compare whether something’s working or not, you have to compare it with something else, and that something else has to not be influenced by either them or me, making any judgements because we know for sure what they’re having. I think most people grasp that concept quite quickly.

GP 1

Clinicians felt supported by the study team

The clinicians felt that the study team were supportive and available to answer queries when needed. They also found the reminders about recruitment to the trial in e-mails and newsletters useful:

I think the amount of support from you guys has been really good, compared to other trials that we’ve been involved in.

GP 5

Interviewer: No? OK, and is there anything that we can be doing to help support you as you carry on in the trial, do you think?

GP 9: No, I mean, again, I’d like the updates, not quite weekly, but the fortnightly update e-mails about how the trial is going I think is useful to remind people that you’re quite a long way off getting your total numbers and any effort is still appreciated. I mean really and truly the input from you guys is fantastic.

GP 9

Many parents were willing to take part

Some clinicians reported that the parents approached had been willing for their child to be recruited to the trial and randomised and that they seemed to understand the concept of a placebo:

I think most people are seemingly quite amenable. Once you’ve seen them and examined them, made sure that they’re not clinically ill enough to need treatment immediately, I find the majority are amenable to having a 50/50 shot at antibiotics or the randomisation shot at antibiotics. They usually accept it.

GP 1

Some parents, most parents we’ve been involved with and people put in have been super, really receptive and understand why it needs to be done. Also, they’ve all been happy to sign up, although it takes ages to go through it all, sign up for the interviews as well.

GP 2

However, although many of the clinicians acknowledged that the trial was important and well designed, the majority reported numerous factors that affected recruitment to the trial.

Theme 4: factors affecting recruitment

Strengths and limitations

The two interview studies provided insights into how perceptions of illness severity and management can affect parent and clinician decision-making regarding the appropriate treatment for a child with an LRTI and whether or not that child should participate in a placebo-controlled trial. However, the sample of each type of participant was small. All of the parents interviewed apart from one were mothers, and so the views of fathers have not been fully captured. Additionally, we did not capture the views of children. The parents interviewed were those who agreed to take part in the trial or observational study, and so we did not explore the views of parents who declined to take part in ARTIC-PC and how these compared. It should also be noted that parents might have expressed less concern about their child not receiving antibiotics because they were enrolled in a research study and would have ready access to care if their child’s illness deteriorated. This attitude may differ in usual consultations, although the clinicians in the study did report a decline in parents’ expectations for antibiotics.

Comparisons with existing literature

A study by Ingram et al. 58 found that parents judged the threat from their child’s cough through a combination of the severity of the illness, whether the cough appeared alongside other symptoms such as fever or a croupy cough, and whether or not the child had experienced a RTI previously. Parents consulted primary care because they did not feel confident caring for a child with a persistent cough or assessing the severity of the illness at home. 58 A more recent study, by Halls et al. ,59 found that although several parents reported at the consultation that their child had a cough, they were more concerned if their child had breathing difficulties and placed great value on obtaining a chest examination in that situation. Our study adds that parents judge the severity of a cough by its sound, in addition to whether or not it appears in combination with other symptoms and the child’s susceptibility to RTIs. Similar to Halls et al. ,59 our findings show that parents consult primary care because they do not feel able to assess illness severity at home, and they place great value on a clinician conducting a chest examination to rule out a chest infection or indications of serious illness. 58,60,61 The persistence of these findings over several years may indicate a need for better information and support for parents that will enable them to assess illness severity at home.

In terms of clinicians’ decisions to prescribe immediate antibiotics for symptoms of RTI, research has shown that perceived clinical need,19,62,65 uncertainty about future illness deterioration and fear of missing something more serious are significant drivers of prescribing antibiotics for children. 19,20,58,61,66,67,70 Additionally, clinicians’ perceptions of parents’ preferences for antibiotics have been found to impact prescribing more than parents’ actual wishes or communicative behaviours during the consultation. 61,62,71 The clinicians in our study displayed confidence both in their decisions not to prescribe antibiotics and their ability to communicate ‘no antibiotic’ prescribing outcomes to parents. They emphasised the importance of ‘clear’ communication during the consultation, especially in terms of safety-netting advice. The provision of safety-netting information is recommended in clinical guidelines,72 but there is evidence that such information might not always be delivered in useful ways that parents can understand. 73–75 Further research on the optimum delivery of safety-netting advice when implementing ‘no antibiotic’ prescribing strategies may, thus, be warranted.

Methods

For description of participant recruitment, see Chapters 3 and 4.

When parents and children were willing for a sample to be taken, a single sweep dual viral/bacterial throat swab was taken and analysed using multiplex PCR. A throat swab was chosen because it has high yield and was acceptable in previous large primary care cohorts. 32

Sample processing for multiplex polymerase chain reaction

Throat swabs in viral transport medium were stored at –80 °C until they were required for testing. Batches of samples for extraction were allowed to thaw at room temperature and 200 µl of viral transport medium was extracted using the QIAsymphony SP (QIAGEN, Manchester, UK) along with an internal process control containing bacteriophages T4 and MS2. Extraction was carried out using the QIAsymphony DSP Virus/Pathogen Mini Kit (QIAGEN) and the 60 µl elution protocol. Real-time PCR samples were amplified and analysed using a Life Technologies Custom TaqMan^TM Low Density Array (TLDA) system on the ViiA™ 7 Real-Time PCR System. Reaction mixes (104 µl) were prepared containing 26 µl of TaqMan Fast Virus 1-Step Master Mix (Life Technologies), 58 µl of molecular grade water and 20 µl of nucleic acid extract. Samples were vortex mixed and pulse centrifuged briefly before 100 µl of the reaction mix was loaded into the chamber of the TLDA card. TLDA cards were centrifuged twice at 1200 r.p.m. (336 x g) for 1 minute to load the wells with reaction mix. TLDA cards were sealed twice with the staking device before they were loaded into the ViiA 7 and amplification was initiated (50 °C for 5 minutes and 95 °C for 20 seconds followed by 45 cycles of 95 °C for 1 second and 60 °C for 20 seconds). On completion of the amplification reactions, fluorescent traces were inspected and analysed for sigmoidal curves. Baselines and thresholds were set automatically using the software algorithms or, where necessary, by manual adjustment to avoid background fluorescence noise. The cycle threshold (Ct) of positive samples was recorded as the point at which the fluorescent trace rises above background and passes through the threshold.

Statistical analyses

The initial plan was to use Cox regression for the primary outcome (the duration of symptoms rated moderately bad or worse). However, for some subgroups (particularly the dual bacterial/viral subgroup) the proportional hazards assumption was not met, so quantile regression was used. Linear regression was used for symptom severity, and logistic regression was used for reconsultations with ongoing, new or worsening symptoms. We assessed whether or not antibiotics had an effect among subgroups with bacterial infections where there was a potentially sensitive organism (H. influenzae, M. catarrhalis, S. pneumoniae) or a viral or dual infection. We report both adjusted analyses (adjusting for age, duration of illness, baseline severity and comorbidity) and unadjusted analyses (as microbiological status may be linked to prior duration and severity and so controlling for these could be controlling for microbiological status).

Potential pathogens were categorised according to Ct value as bacterial pathogens, or viruses, that could potentially respond to amoxicillin. When detected, Neisseria meningitidis and coagulase-negative staphylococci were classified as commensal carriage organisms. 83 We assumed that indeterminate Ct values were negative, but we also performed a sensitivity analysis assuming that indeterminate Ct _t values were positive. We also looked at the interaction between the mean Ct value (inversely related to bacterial/viral load) and antibiotic group. We performed a Cox regression of symptom duration on antibiotic group, bacterial Ct value and their interaction, adjusting for baseline covariates. We repeated this analysis using mean viral Ct value. We then repeated these analyses for symptom severity and reconsultation outcomes.

These analyses were repeated when the observational data were included to increase power, and propensity scores were used to control for confounding by indication, but, as noted with the effectiveness data, there is still likely to be residual confounding by indication.

Trial results

The flow diagram of samples in both the trial data and the observational data is shown in Figure 6.

FIGURE 6.

Flow diagram of throat swabs.

A total of 306 throat swabs were analysed in the RCT. Potential pathogens categorised as bacterial pathogens, viruses or carriage organisms were balanced between the treatment groups (see Table 20; for individual organisms, see Table 30). Bacterial pathogens that were potentially sensitive to amoxicillin (H. influenzae, M. catarrhalis, S. pneumoniae) were found in 51% (76/150) of the trial placebo group and 49% (76/156) of the trial antibiotics group.

Results assuming that intermediate cycle threshold values were negative

There were no statistically significant interactions of bacterial, viral or dual bacterial/viral subgroups with antibiotic group for any of the outcomes, and the adjusted effect of antibiotics on the primary outcome (duration of symptoms rated moderately bad or worse) for those with potential bacterial pathogens was similar to that for those without, as it was for viral infections (see Table 21). The only subgroup with potentially important differences in duration of illness with antibiotics was the dual bacterial/viral subgroup, where quantile regression suggested that the estimate of the effect of antibiotics in this group was a non-significant reduction in the median duration of illness of 3 days (99% CI –9.9 to 3.9 days).

The adjusted effect of antibiotics on symptom severity was also similar whether there were bacteria, viruses or dual infections, albeit non-significantly greater when no viruses were detected or with dual infections (see Table 22). The adjusted effect of antibiotics on reducing reconsultation documented a non-significantly greater effect when no viruses were present (see Table 23).

The regression of symptom duration on antibiotic group, bacterial Ct value and their interaction, adjusting for baseline covariates, is shown in Table 24. There was no evidence that antibiotics had a greater effect on symptom resolution when high bacterial loads were present; if anything, a non-significant trend was seen for the impact of antibiotics to increase as bacterial load decreased, i.e. as Ct value increased (HR 1.05, 95% CI 0.96 to 1.13), and no evidence of an interaction with virus Ct value. There was also no clear evidence of treatment benefit from the interaction with Ct value for symptom severity or reconsultations for either bacteria or viruses (see Table 24, and shown graphically for symptom severity in Figures 7 and 8).

FIGURE 7.

Symptom severity at days 2–4 vs. bacterial Ct value.

FIGURE 8.

Symptom severity at days 2–4 vs. viral Ct value.

Results assuming that intermediate cycle threshold values are positive

Potential bacterial, viral pathogens and carriage organisms were balanced between the trial groups (see Table 25).

There were no statistically significant interactions between any of the pathogen subgroups and antibiotic group for any of the outcomes, and no consistent trends. There was no evidence of an increased effect of antibiotics on the duration and severity of symptoms or on reconsultations for those with potential bacterial pathogens (see Tables 26–28). When viruses were not present, there was a trend for increased impact on symptom severity, reconsultation and duration of symptoms. For dual infections, there was a trend for an impact on symptom severity and duration but not on reconsultation.

Analyses including the observational data

A total of 488 throat swabs were analysed, with a further 182 in the observational cohort. Both in the observational cohort and overall these potential pathogens were fairly balanced between the antibiotics and no-antibiotics groups in the observational, combined and trial samples (see Tables 29 and 30).

There were no statistically significant interactions of any of the pathogen subgroups with antibiotics for any of the outcomes, no evidence of important clinical benefits from antibiotics in subgroups (see Tables 31–34) and no impact of Ct values (see Figures 9 and 10).

FIGURE 9.

Symptom severity at days 2–4 vs. bacterial Ct value in combined trial and observational data.

FIGURE 10.

Symptom severity at days 2–4 vs. viral Ct value in combined trial and observational data.

Limitations

The analyses were not powered specifically to assess the interaction of microbiological diagnosis with outcome, and the power was reduced by some children or their parents not wanting swabs to be taken. Nevertheless, the estimates largely indicate no clear evidence of differential impact of antibiotics on bacterial or viral subgroups. In cases in which microbiological diagnosis may demonstrate an impact (the impact of antibiotics where viral pathogens were absent in the combined data set on symptoms duration), given the number of secondary analyses, a chance effect is quite possible, and the estimate was of only borderline significance. The detection of Chlamydophila pneumoniae, Bordetella pertussis and Mycoplasma pneumoniae implies a role of those micro-organisms in the disease, as they are seldom carried asymptomatically. However, pathobionts such as S. pneumoniae, M. catarrhalis and H. influenzae may also be detected by PCR testing, but because these organisms are frequently detected in throat swabs from normal children,84–86 their detection does not necessarily imply a role in the disease process, and the presence of asymptomatic carriage will dilute the apparent effectiveness of antibiotics in bacterial subgroups. Similarly, we did not assess the antibiotic resistance of individual strains, and the presence of resistant strains will also dilute the apparent effectiveness of antibiotics, but most strains are susceptible to amoxicillin in community samples of children. 84–86 Similarly, the study is not powered to detect the impact of antibiotics on infection with particular viruses such as influenza. 87 It is also possible that lower airway samples would be better but are not possible in routine primary care.

Relationship to other studies

The trial data set shows that the prevalence of both potential viral and bacterial pathogens using similar sampling methods is similar to that in very large observational data sets of children presenting in primary care,32 suggesting that the impact of microbiological diagnosis should be generalisable. As in a previous study, we have not shown that microbiological diagnosis improves the estimation of prognosis in children79 and we found a similar finding to studies in adults of no impact on symptom duration. 80

For the dual infection subgroup, we found no clear impact of dual infections on reconsultations, unlike in the trial in adults,30 but the current study had much less power than the adult trial. We did find a potentially important, albeit non-significant, 3-day reduction in the median duration of symptoms with antibiotics for the dual infection subgroup, but the analysis was underpowered and, given the exploratory nature of subgroup analyses, should be viewed with some caution. Assuming that the effect in this subgroup might be real, three children would have to be assessed to identify one child who might benefit, and the cost-effectiveness of such an approach would need to be demonstrated.

Conclusion

Both viral and bacterial pathogens are frequently detected among children presenting with uncomplicated LRTI in primary care, but there is no clear evidence that the impact of antibiotics is likely to be different according to the presence of bacteria or viruses. This calls into question if microbial diagnosis using point-of-care tests to detect pathogens in the upper airway should be used for children presenting with uncomplicated LRTI in primary care and suggests that, prior to introduction, microbiological point-of-care tests should be subject to rigorous trials of effectiveness.

Intervention cost

In the trial, both amoxicillin and placebo were specially prepared in identical formats. This meant that the cost of the intervention drug did not reflect its cost in normal practice, in which it is priced as a prescription-only drug. Among the amoxicillin variants available, the NHS-funded drug most similar to the intervention drug was 100 ml oral suspension of amoxicillin in 20-dose vials, which is priced to the NHS at £1.75. A dispensing cost of £1.26 was added, giving a cost of £3.01 per patient in the intervention group.

The placebo cost was set at zero as this was specific to the study. The cost of the initial consultation was also set at zero as it was specific to the study and the same in both groups.

Primary care consultations

A total of 129 patients in the trial had a primary care consultation in the 4 weeks following randomisation. Consultations were classified by who they were with (GP, practice nurse) and how they took place (face to face or over the telephone). Most patients saw a GP and most consultations were face to face. A few patients (< 5) who did not have a record of who saw them were assumed to have been seen by a GP and costed accordingly. All were costed using PSSRU 2019/20 financial year values (see Table 35).

The PSSRU cost of a GP was for an average consultation lasting just over 9 minutes. The only cost available from PSSRU for nurse consultations was per hour. The duration of nurse consultations was assumed to be the same as that of consultations with GPs, giving an estimated cost of £5.99 per consultation. The cost of a telephone consultation with a GP is given by PSSRU as £8, and it was assumed that a telephone consultation with a nurse cost the same. The alternative of pro-rating the nurse cost to that of the GP would have led to an unreasonably low unit cost. As the number of telephone consultations with nurses was very small, this assumption makes minimal difference to the total costs.

Drugs prescribed

A total of 51 antibiotic drug prescriptions were recorded (see Table 36) in primary care consultations, classified as penicillin, macrolides and other antibiotics. These were costed using the National Drugs Tariff for January 2020. 89

The unit cost of penicillin was £1.75, based on a 125 mg/5 ml oral solution of phenoxymethylpenicillin.

The unit cost of macrolides was £3.94, based on the average of the three drugs and so defined based on the same oral solution. The cost of other antibiotics was that of the named drug, with the dose assumed to be that of the most commonly prescribed.

Referrals

The note review provided data on 24 patients referred by primary care to one of the following services: A&E, out of hours, NHS Direct, and outpatients and other.

Of the two under ‘other’, based on the text, one was allocated to outpatients (‘hospital’ was recorded) and one had zero cost attributed (‘no referral’).

These were costed using the unit costs in Table 37, taken from the NHS Reference Costs. 90

Admissions

Data collected included hospital admissions by reason. These are shown in Table 38, which indicates that six of the seven were directly related to respiratory problems and the seventh, ‘gastrointestinal’, was possibly related to respiratory problems. The dates of admission and discharge were used to distinguish between inpatient and day cases.

These cases were costed using the NHS National Tariff (https://improvement.nhs.uk/documents/6486/2_-_National_schedule_of_NHS_costs_V2.xlsx), which distinguishes between elective and non-elective and by Healthcare Resource Group (HRG). If referral was indicated in primary care notes, then the elective cost was used; otherwise, the non-elective was employed. The four admissions were assumed to fall under a paediatric respiratory HRG.

The HRGs used (see Table 39) were as follows. PD15D Paediatric Acute Bronchiolitis (with CC Score 0, that is with zero complications) was assigned to those two admissions that mentioned bronchiolitis. PD12C Paediatric Lower Respiratory Tract Disorders without Acute Bronchiolitis (with CC Score 0) was assigned to those with wheezing or cough as well as those not fitting readily into the other two HRGs below. PD14F Paediatric, Asthma or Wheezing (with CC Score 0) was assigned to the patient admitted with gastrointestinal, as the most appropriate respiratory related condition paediatric HRG. In a sensitivity analysis this patient was omitted.

The hospital unit costs ranged from £396 to £562. The relatively small number of patients and cost differences involved imply that the allocation to a particular HRG has little effect on the overall costing.

Remedies

Thirty-two participants reported buying over-the-counter remedies for their children. Most of these remedies were low-priced items such as Calpol, but a few were more expensive; notably, one parent reported the nightly use of a humidifier. Over-the-counter remedy costs were based on information from Boots online, the largest UK retail pharmacy. A small number of prescription drugs recorded under remedies were costed using the National Drugs Tariff. 89

Time off work

Both time off work and additional care purchased because of the child’s illness were recorded weekly in patient diaries; 79 parents reported having to take time off work or college or to pay for child care. Most parents reported this in terms of days but a few reported it in terms of hours. These items were aggregated and costed at minimum wage for adults of £8.21 per hour (the value in April 2019).

Additional care purchased

Nineteen parents reported that they had had to purchase additional care for their children. This was again recorded mainly in days but sometimes in hours. As with time off work, this was costed at the national adult minimum wage.

The cost of remedies plus that of time off work and any additional care comprised the non-NHS costs which, added to the NHS costs, comprised the societal costs.

Cost results

Table 40 summarises the NHS cost differences between the treatment groups. The mean NHS cost per patient was slightly higher in the antibiotics group at £29.40, than in the placebo group at £25.80: a difference of £3.50. This was largely because of the cost of the intervention (£3.01). The overall cost difference was not statistically significant and neither were any of the more detailed elements shown in Table 40.

Societal costs, that is NHS costs plus remedies purchased and time off work and care, were, on average, roughly double NHS costs, mainly owing to time off work. Societal costs were slightly lower in the placebo group at £58.60, than in the antibiotics group at £62.30. This difference of £3.70 was almost entirely owing to the NHS cost difference, indicating that the groups incurred fairly similar non-NHS costs.

Quality of life and quality-adjusted life-years

The EQ-5D-Y questionnaire scores were translated into utility scores based on the UK tariff (EQ-5D-Y user guide 2015). QALYs were derived using area under the curve based on five points of utility scores (baseline and weeks 1–4).

The means of the overall EQ-5D scores and VAS scale of all patients at baseline and follow-up are shown in Table 41. The completion level was poor, with around two-thirds of the data missing. Completion rates fell from just under 40% at baseline to just over 20% at week 4. There were no differences in completion rates and mean EQ-5D scores between the groups at baseline or at follow-up. The EQ-5D scores and VAS scale increased over time in both groups. The completion rates and scales for VAS in the antibiotics group are similar to those in the placebo group at baseline and follow-up (see Table 41). The utility scores are shown in Table 42; these increased over time in both groups and in both children aged ≥ 5 years and children aged < 5 years (see Figure 11).

FIGURE 11.

Utility scores at different time points for antibiotics (dark blue) and placebo (light blue) groups for children aged > 4 years.

Table 43 and Figure 12 document the VAS results. In line with the EQ-5D results, these indicated only minimal differences between the groups.

FIGURE 12.

The VAS score at different time points for antibiotics (dark blue) and placebo (light blue) for all patients in the treatment groups.

Incremental cost-effectiveness analysis

In terms of cost per illness-day (i.e. a day rated moderately bad or worse), the incremental cost of £3.50 in the antibiotics group combined with a difference in symptom-free days of 0.5 yielded an incremental cost per illness-day of £22 (95% CI –£63 to £48). The lack of a willingness-to-pay threshold makes the interpretation of this estimate difficult.

Incremental cost, incremental quality-adjusted life-years and incremental cost-effectiveness analysis

Using the EQ-5D data based on the mean scores between the groups based on the available EQ-5D scores at baseline and follow-up (see Table 44) led to a QALY difference that was very small, at 0.0001, and not significant. The incremental cost-effectiveness ratio (ICER) was estimated at £30,851, with a wide credible interval from –£73,639 to £109,429 (see Table 45). However, as noted above, EQ-5D scores were available for less than half of all participants.

Multiple imputation was explored as a means of dealing with the missing EQ-5D data. Applying the same imputation method as for the primary outcome led to a larger but not statistically significant QALY difference of 0.002 and an ICER of £6417 (95% CI £12,240 to –£20,535) (see Table 46). The exploration of other imputation methods led to a wide variety of estimates that were not considered reliable. Details of these are available from the authors.

Exploration of the uncertainty around this estimated incremental cost per QALY can be shown in terms of a scatterplot (see Figure 13) and cost-effectiveness acceptability curves (see Figure 14).

FIGURE 13.

Scatterplot of joint distribution of incremental mean cost from NHS perspective and mean QALY over 1 month: 95% confidence ellipse of antibiotics vs. placebo.

FIGURE 14.

Cost-effectiveness acceptability curve of the antibiotics and placebo groups based on QALY over 1 month.

Figure 13 shows a wide spread around the points of zero difference for both cost and QALY increments. The cost-effectiveness acceptability curves (see Figure 14) show that the probabilities that the intervention is cost-effective against the placebo were 75% and 76% at the thresholds of £20,000 and £30,000, respectively.

Non-NHS costs based on private purchases of remedies and on time off work or caring were fairly equally distributed between the groups but led to a higher cost increment in the antibiotics group, with a resulting ICER of just over £40,000.

Limitations

No analysis of cost-effectiveness in subgroups was carried out for several reasons, including the lack of any relevant differences in the primary outcome, the extent of missingness in the QALY data, and the lack of cost differences between the groups.

The major limitation of this study is the high (two-thirds) proportion of missing data for the EQ-5D-Y. With so many missing data, imputation involved such major assumptions as to make bias likely.

Conclusion

In conclusion, the intervention is unlikely to be cost-effective as a result of the small and non-statistically significant differences observed, the difficulty of imputing the majority of the QALY data, and the lack of consideration of the societal costs of antibiotics in relation to resistance. 105,106

Implications for health care

1. Clinicians can reduce their prescribing of antibiotics for children presenting with uncomplicated chest infections without this leading to poor symptom control or high societal costs. The perception that parents’ attitudes to antibiotics are changing should help in antimicrobial stewardship initiatives.

2. A ‘simple’ score that incorporates respiratory rate and low oxygen saturation could potentially identify the minority of children who will develop complications. To be clinically useful it could be developed as an app with automated outputs.

3. Parents need better access to information to support them in their decision to consult and in the self-management of their child’s illness.

Recommendations for research

1. The data can be incorporated in the Cochrane review and in individual patient data meta-analysis.

2. Further work on the incremental QALY gain from antibiotics is needed, assessing a range of models and their implications when imputing missing QALY data, as is better evidence about how to incorporate AMR resource implications in modelling.

3. Further research is necessary to determine if upper respiratory tract microbes mediate the effects of antibiotics and are, therefore, useful targets for future microbial point-of-care tests.

4. The prognostic score and other recently developed clinical prediction rules need to be externally validated, and the score could be developed as an app with automated outputs as a tool to reduce antibiotic prescribing.