Individual participant data meta‐analysis versus aggregate data meta‐analysis: A case study in eczema and food allergy prevention

Abstract Introduction Meta‐analysis traditionally uses aggregate data from published reports. Individual Participant Data (IPD) meta‐analysis, which obtains and synthesizes participant‐level data, is potentially more informative, but resource‐intensive. The impact on the findings of meta‐analyses using IPD in comparison with aggregate data has rarely been formally evaluated. Methods We conducted a secondary analysis of a Cochrane systematic review of skincare interventions for preventing eczema and food allergy in infants to identify the impact of the analytical choice on the review's findings. We used aggregate data meta‐analysis only and contrasted the results against those of the originally published IPD meta‐analysis. All meta‐analysis used random effects inverse variance models. Certainty of evidence was evaluated using GRADE. Results The pooled treatment effects for the Cochrane systematic review's co‐primary outcomes of eczema and food allergy were similar in IPD meta‐analysis (eczema RR 1.03, 95% CI 0.81, 1.31; I241%, 7 studies 3075 participants), and aggregate meta‐analysis (eczema RR 1.01 95% CI 0.77, 1.33; I253%, 7 studies, 3089 participants). In aggregate meta‐analysis, the statistical heterogeneity could not be explained but using IPD it was explained by one trial which used a different, bathing intervention. For IPD meta‐analysis, risk of bias was assessed as lower and more adverse event data were available compared with aggregate meta‐analysis. This resulted in higher certainty of evidence, especially for adverse events. IPD meta‐analysis enabled analysis of treatment interactions by age and hereditary eczema risk; and analysis of the effect of treatment adherence using pooled complier‐adjusted‐causal‐effect analysis, none of which was possible in aggregate meta‐analysis. Conclusions For this systematic review, IPD did not significantly change primary outcome risk ratios compared with aggregate data meta‐analysis. However, certainty of evidence, safety outcomes, subgroup and adherence analyses were significantly different using IPD. This demonstrates benefits of adopting an IPD approach to meta‐analysis.


| INTRODUC TI ON
Meta-analysis is fundamental to evidence-based decision making, combining quantitative outcomes across multiple related studies to summarize all the available evidence on a particular clinical question. Traditional meta-analysis combines aggregate data (e.g. mean difference, risk ratio, etc.) obtained from individual trial publications or trial authors. Individual Participant Data (IPD) refers to raw, original participant level data. In an IPD meta-analysis, IPD is collected from all possible eligible trials and included in the synthesis. Data sharing has become more common in recent years meaning easier implementation and increased use of IPD meta-analysis. 1,2 Individual Participant Data meta-analysis is considered 'the gold standard of systematic reviews'. There are many well-known potential advantages of IPD meta-analysis over aggregate meta-analysis. [2][3][4] These include the ability to standardize participant inclusion and exclusion criteria and outcome definitions across studies, use of uniform statistical methods, such as consistent adjustment for baseline characteristics, analysis models and methods of handling missing data. Other benefits include the ability to explore how a treatment effect is modified by participant factors, and more in-depth risk of bias assessments. IPD from unpublished studies can also be included, reducing publication bias. Reviewers can independently check the trial data set for errors and recalculate poorly reported outcomes from published studies. However, it is more resource intensive, requiring large collaborations, trusting relationships and data sharing in addition to more complex data preparation and analysis. 3,5,6 In this article, we investigated the potential added value of conducting IPD meta-analysis for a previously conducted IPD Cochrane systematic review which primarily assessed the effects of skincare interventions, such as emollients, for primary prevention of eczema and food allergy in infants. 7 Secondary objectives of the original systematic review were to identify features of the study population associated with greater treatment benefit or harm including age, hereditary risk and intervention adherence. Using the same search results, we conducted a secondary analysis of the review and metaanalysis using aggregate methods only and compared the results to those of the previously reported primary IPD meta-analysis. We sought to identify the impacts of the analytical choice on the review's findings, in order to help guide other researchers planning future meta-analysis or interpreting meta-analysis results.

| Outcomes
The main outcomes of interest were (i) differences in the quantity of studies included within IPD vs traditional aggregate data metaanalyses, (ii) differences between the pooled treatment effects sizes from IPD vs. traditional meta-analysis for the two co-primary outcomes (described below). Additional outcomes were (i) differences of evidence, safety outcomes, subgroup and adherence analyses were significantly different using IPD. This demonstrates benefits of adopting an IPD approach to meta-analysis.

K E Y W O R D S
atopic dermatitis, food allergy, individual participant data Introduction, meta-analysis, prevention

G R A P H I C A L A B S T R A C T
Benefits of an IPD approach to meta-analysis were demonstrated. Risk ratios were similar, but IPD allowed a conclusion that skincare interventions probably do not affect eczema risk, compared with a conclusion that they may not affect eczema risk when using aggregate data meta-analysis, due to unexplained statistical heterogeneity in the latter. GRADE certainty of evidence ratings, adverse events analyses, subgroup and adherence analyses were either significantly different with IPD or only possible with IPD compared with aggregate data meta-analysis.

Key messages
• We compared IPD and aggregate meta-analysis using data from a Cochrane review of skincare interventions.
• Primary outcome estimates were similar, but certainty of evidence increased for Individual Participant Data meta-analysis.
• Safety, subgroup and adherence analysis and heterogeneity investigation were also facilitated with Individual Participant Data.
between the pooled treatment effects sizes from IPD and traditional meta-analysis for secondary outcomes (described below), (ii) pooled effect sizes that were only possible with IPD (iii) differences in risk of bias assessments, and (iv) differences in quality of the evidence. We also summarize the costs and resources involved for the IPD-analysis.
The co-primary Cochrane review outcomes were (i) eczema by 1-3 years assessed using the Hanifin and Rajka criteria 8 or the UK Working Party refinement of them 9 and (ii) food allergy by 1-3 years assessed using a combination of parental history, skin prick tests and if needed, oral food challenge. If (i) was not available then doctor diagnosis of eczema could be used; if no doctor diagnosis was available, then parent report of eczema was used. Secondary outcomes were: skin infections; stinging or application site reactions to moisturizers; slippage accidents; serious adverse events; clinicallyassessed eczema severity at 1-3 years; parent report of eczema severity at 1-3 years; time to eczema onset; parent report of immediate (<2 hours) reaction to a known food allergen at 1-3 years; and allergic sensitization to foods and inhalants at 1-3 years.

| Types of studies, participants and interventions
The types of studies, participants and interventions eligible for inclusion within the review which we re-analyse here have previously been reported. 10 To summarize, eligible studies were parallel-group, or factorial randomized trials using individual or cluster randomization. Participants were infants aged from birth to 12 months, excluding study populations defined by pre-existing disease or illness.
Interventions were skin care interventions that could potentially enhance the skin barrier function, reduce dryness or subclinical inflammation such as moisturizers/emollients, bathing products, advice regarding soap exposure and bathing frequency. It is thought that disruption of the skin barrier in early life can lead to eczema and subsequent food sensitization and allergy. Comparators were no treatment intervention or standard care in the study setting.

| Search strategy
We used the same search of MEDLINE, Embase, CENTRAL and trial registers (performed July 23rd 2020) and study selection as the Cochrane review previously reported. 10 Two review authors independently screened study titles and abstracts for eligibility, with arbitration by a third author where necessary. The full texts of all potentially eligible studies from the search were obtained to confirm eligibility. No language restrictions were imposed.

| Data extraction
Aggregate data were extracted from trial publications of the eligible studies and did not include any author correspondence. Data were extracted into excel in duplicate by EVV and SC. Any differences in extraction were discussed and resolved.

| Statistical analysis
Aggregate meta-analyses included eligible trials providing aggregate data for the relevant outcome. For binary outcomes, we calculated a pooled Risk Ratio (RR), for continuous outcomes a pooled standardized mean difference (SMD), and for time to event outcomes a pooled Hazard Ratio (HR). All estimates had associated 95% Confidence Intervals (CI) calculated.
Aggregate meta-analyses were carried out using inverse variance random effects models. As the majority of trials only reported unadjusted outcomes, we combined unadjusted estimates for the primary aggregate analysis. Aggregate meta-analysis, including adjusted RR estimates, was conducted for sensitivity analysis where possible. Across meta-analyses, the I² statistic and Chi 2 test quantify the degree of statistical heterogeneity of trials judged as clinically homogeneous. 11 Analysis was carried out in STATA 15 and Revman.

| Risk of bias assessment and quality of evidence
Two authors (SC and RJB) re-assessed risk of bias for outcomes with risk of bias assessments in the IPD meta-analysis using the aggregate data to mimic a traditional meta-analysis. Aggregate data risk of bias assessments were compared to original IPD assessments. The Cochrane 'Risk of bias 2' tool was used. The GRADE approach was applied to the main outcomes of the Cochrane review by two authors (SC and RJB), using aggregate data information to compare to the IPD assessments. Outcomes were graded as high, moderate, low or very low quality.

| Additional methods of previously conducted IPD analysis
The protocol for the original Cochrane IPD review has previously been published. 7 A formal protocol was not prepared for the aggregate data secondary analysis. In the Cochrane IPD review, all trial authors of eligible studies were contacted and asked to provide IPD.
Data were de-identified, transferred, then cleaned and coded for analysis. Consistency checks against published results were carried out on the data, any errors or extreme values were queried with trial authors where necessary.
For the IPD meta-analysis, a pre-specified statistical analysis plan (SAP) was previously followed. 12 No new IPD meta-analysis was performed for this article. In brief, the previously conducted IPD meta-analysis used a two-stage approach. In stage 1, for binary/continuous/time-to-event outcomes, a binomial/linear/binomial with a complementary log-log link regression model was, respectively, used. All stage 1 models included sex and where relevant (trial not exclusively in a high risk population) family history of atopic disease and resulted in adjusted RR. In the second stage, inverse variance random-effects models were used to obtain a pooled treatment effect. As the aggregate meta-analysis predominately used unadjusted treatment estimates, we did not expect aggregate and IPD metaanalysis results to be identical.

| Number of studies in aggregate and IPD meta-analysis
The search identified 33 eligible RCTs, with 25,827 participants of which 11 (5217 participants) had outcomes qualifying for inclusion in one or more meta-analysis (see Figure S1). Of these 11 studies, 10 (5163 participants) were included in aggregate meta-analysis. One trial, identified as completed via its trial registry record, had not yet been analysed or reported when the review was carried out 13 so could not be included in the aggregate data meta-analysis. This trial could be included in the previously conducted IPD meta-analysis, which included 10 studies (5154 participants). In the IPD metaanalysis, all 11 studies eligible for meta-analysis were contacted and asked to provide IPD; one study did not respond nor provide IPD, Migacheva 2018. 14 This study was not included in the primary IPD analysis; but was included in a sensitivity analysis which combined IPD with aggregate data. Figures 1-2 and Table 1 show that for most outcomes, IPD enabled a greater number of studies to be included within pooled estimates.

| Aggregate versus IPD meta-analysis for eczema
We extracted unadjusted aggregate data on the primary eczema outcome for 7 trials (3089 participants) which differed to the 7 trials (3075 participants) in the IPD meta-analysis. Unlike the IPD metaanalysis, study NCT03376243 could not be included as no aggregate data were available, but Migacheva 2018 15 had aggregate data on eczema and was included. The unadjusted trial treatment estimates naturally differed slightly to the results used in the IPD meta-analysis which were adjusted for sex and family history of atopy. In addition, for one trial, IPD enabled us to include 156 in the analysis and use the preferred, more robust, outcome of eczema by 2 years as diagnosed by a physician only (rather than diagnosis by a physician or a parental report) following the pre-specified hierarchy of eczema diagnosis.
Collected data on eczema from one participant had been excluded from the published analysis as they had incomplete baseline data on parental atopy, which was a covariate in the trials original published analysis. In another trial IPD revealed 86 cases of eczema as diagnosed by UKWP out of a denominator of 455, which differed slightly to the figures reported in the study authors publication of 86/459. and reduction of statistical heterogeneity in the IPD meta-analysis.
Moreover, in IPD meta-analysis the statistical heterogeneity (41%) could be entirely explained by one trial with a RR favouring standard care (Skjerven). When this trial was excluded in IPD meta-analysis the I 2 reduced to 0. This may be explained by the intervention type (bathing with oil and moisturizer applied to the face only) and/or timing of intervention initiation (initiated from 2 weeks) in the Skjerven trial. In contrast in the aggregate meta-analysis Skjerven did not alone explain the heterogeneity. After excluding Skjerven from the aggregate data, the pooled adjusted RR was 0.93, 95% CI [0.72, 1.20] I 2 = 31%. Sensitivity analysis results using adjusted aggregate data where available were similar to primary analysis (see Table S1).
Results of aggregate data and IPD subgroup analyses (see Table S2) were similar, but also with reduced statistical heterogeneity in IPD meta-analysis.

| Aggregate versus IPD meta-analysis for food allergy
One trial (Chalmers 2020) collected data on food allergy as confirmed by oral food challenge ( Figure 2). A total of 15/482 confirmed food allergy cases were identified in the intervention group versus 6/494 in the control group in both IPD and aggregate data analysis.
The aggregate data analysis resulted in an unadjusted RR = 2.56, 95% CI [1.00, 6.55]; The lower limit of the 95%CI included no difference (RR = 1). The IPD analysis resulted in an adjusted RR = 2.53, 95% CI [0.99, 6.47]. The use of adjustment in the IPD meta-analysis provides a more powerful analysis, and resulted in a lower limit for the 95%CI which spanned down to 0.99, including no difference.

| Aggregate versus IPD meta-analysis for secondary outcomes
Aggregate data on skin infections were available for three studies   Table 1 and were typically small.
Using the IPD, we could calculate time to eczema in a consistent manner across nine trials (3349 participants) using obtained visit dates and eczema outcomes, pooled adjusted HR 0.86, 95% CI [0.65, 1.14].

| Analyses only possible using IPD
We could not conduct participant-level treatment interaction metaanalyses using aggregate data and were limited to individual trial

Number of studies (I 2 )
Adverse event: skin infection   Table 2).

Meta-analysis for the Complier Average Causal Effect (CACE)
was not possible using aggregate data. Only one trial previously reported the CACE for eczema as an adjusted OR (adjusted for centre and parental atopy) for intervention use ≥3 days/week over the first 3 months, adjusted OR = 0.88 (0.50-1.56). IPD enabled computation of compliance indicators, using consistent thresholds across trials supplying compliance data, and conduction of additional CACE analysis that was not previously performed. The pooled CACE estimates did not provide evidence that treatment adherence impacted risk of eczema (see Table 2).

| Risk of bias
IPD beneficially impacted risk of bias assessments (see Table 3). For the co-primary eczema outcome, 3/7 studies were judged low risk of bias (4/7 some concerns of risk of bias) in the IPD meta-analysis, compared to only 1/7 in aggregate meta-analysis (6/7 some concerns of risk of bias). This was due to: (1) IPD enabled additional missing data sensitivity analysis that had not previously been reported, (2) In IPD meta-analysis, the risk of bias for the selection of the reported results was low as a pre-specified SAP, finalized before any unblinded data was available, was used for meta-analysis. In the aggregate meta-analysis, three studies did not provide adequate prespecified information on a trial registry or whether a pre-specified SAP was followed, resulting in some concerns on this domain.

| Quality of the evidence
The quality of the evidence for the aggregate meta-analysis is summarized in Table 4 and for the IPD analysis in Table 5. For the coprimary eczema outcome, due to increased unexplained statistical heterogeneity and more trials rated as some concerns of risk of bias in the aggregate analysis, there is low certainty evidence; in IPD meta-analysis there was moderate certainty evidence. For two other secondary outcomes, the certainty of evidence is lower by one grade in the aggregate meta-analysis versus IPD, with reasons given in Table 3. Table 6 summarizes the impact of these differences for making judgments on the review outcomes.

| Resources required for the IPD meta-analysis
IPD meta-analysis can be retrospective, where previously published trial data is shared by collaborators, or prospective where collaborating groups agree prior to trial publication to share data. Our IPD meta-analysis included prospective IPD meta-analysis, though some of the smaller pilot studies had been published previously. This necessitated open, trusting relationships with collaborators who shared data prior to its publication, which took effort and time, making it slower and more costly than retrospective IPD meta-analysis.
IPD analysis also required data sharing agreements between institutions, as de-identified data is shared by the collaborating groups.
The IPD meta-analysis needed increased time for the statistical analysis (50% FTE statistician for 2 years), and costs for project management which included searching, data extraction, data sharing agreements, data transfer and data cleaning (100% full time equivalent (FTE) for 2 years), collaborator meetings and international travel beyond the typical costs for an aggregate systematic review and meta-analysis.

| Main Findings
In this evaluation, we explored differences in outcomes from a Cochrane systematic review depending on the approach to data collection and analysis. We compared IPD with aggregate data metaanalysis. We found that while risk ratios for primary outcomes were very similar, GRADE certainty of evidence ratings, adverse events analyses, subgroup and adherence analyses were all either significantly different with IPD or only possible with IPD compared with aggregate data meta-analysis. IPD allowed for a better understanding of the primary eczema outcome and explanation of the statistical heterogeneity. Overall, use of IPD allowed a conclusion that the skincare interventions studied probably do not affect eczema risk; compared with a conclusion that they may not affect eczema risk when using aggregate data meta-analysis.
There were significant differences in findings for secondary outcomes, with IPD generally able to offer more reliable and precise infor- Overall the IPD meta-analysis was more resourceful in terms of time, research staff and associated financial costs and was only possible due to trusting relationships across a large collaborative group of trialists who were willing to share anonymized data; some before the primary publication of their results. Barriers to IPD metaanalysis may include an unwillingness to share data. To facilitate data sharing within this review, data sharing agreements between each data provider and the SCiPAD research group included the conditions for secure data transfer and holding and that any results were not to be published without approval from all data suppliers. Data suppliers could nominate up to three representatives to serve on the authorship group for the main IPD meta-analysis.
The previously conducted IPD meta-analysis was performed using a two-stage analytical approach. This approach was originally chosen since individual study effects are immediately available for examination in forest plots alongside pooled results and it enables IPD and aggregate data for any trials not providing IPD to be readily combined in the same analysis (a sensitivity analysis conducted within the original Cochrane review combined IPD and aggregate data from one trial that did not supply IPD). A two-stage approach also proved valuable as one trial providing IPD gave access to their data in an online secure platform and it could not be exported to be i.e. the interaction predicts eczema perfectly. In the standard care group 5/22 (0 mutations) and 1/1 (1 or 2 mutations) had eczema. In the skin care intervention group 3/21 0 mutations) and 1/3 (1 or 2 mutations). ITT = Intention-to-treat. CACE estimates represent the relative risk of the outcome (eczema or food allergy) for skin care intervention use versus standard care or no skin care intervention among those who would comply with the allocated interventions. e CACE estimates for the one trial reporting food allergy were accompanied by wide 95% CI's suggesting a problem with the estimation procedure.

TA B L E 3 Risk of bias assessments for aggregate data and IPD meta-analysis
Outcome Study

Moderate
Same grading as IPD. In aggregate meta-analysis: Downgraded by one level for imprecision due to small numbers of events, with wide confidence intervals, which include both a harmful effect and a beneficial effect (3 trials).
In IPD meta-analysis: Downgraded by one level for imprecision due to wide confidence intervals, which include both a harmful effect and no effect (6 trials)

Outcome
Assumed risk

No. participants (studies)
Quality of the evidence (GRADE)

very low
Downgraded from low (IPD) to very low (aggregate).
In aggregate meta-analysis: Downgraded one level for some concerns of risk of bias due to potential selective reporting, and two levels for imprecision due to small numbers of events from a single study, with wide confidence intervals, which include both a harmful effect and no effect.
In IPD meta-analysis: Downgraded by two levels for imprecision due to small numbers of events, with wide confidence intervals, which include both a harmful effect and a beneficial effect (4 trials).
Time to onset of eczema low Downgraded from moderate (IPD) to low (aggregate).
In aggregate meta-analysis: Downgraded one level for some concerns of risk of bias in the two included trials due to missing data or potential selective reporting, and one level for clinical heterogeneity (one trial had only 32 week follow-up versus the second had 2 years follow-up). Small numbers of events from two studies, with wide confidence intervals, although all estimates represent a clinically meaningful increase in time to onset of eczema.
In IPD meta-analysis: Downgraded one level for heterogeneity driven by more than one trial, for which review authors were unable to identify a plausible explanation. low Same grading as IPD.

TA B L E 4 (Continued)
In aggregate meta-analysis: Downgraded two levels for imprecision due to small numbers of events from a single study, with wide confidence intervals, which include both a harmful effect and no effect.
In IPD meta-analysis: Downgraded two levels for imprecision due to small numbers of events from a single study, with wide confidence intervals, which include both a harmful effect and no effect.
Allergic sensitization to a food allergen (at 1 to 2 years)

very low
Same as IPD meta-analysis. Downgraded one level for heterogeneity, for which the review authors were unable to identify a plausible explanation, and two levels for imprecision due to wide confidence intervals, which include both a harmful and a beneficial effect and some concern of risk of bias due to missing data across both included trials.
Note: All estimated are unadjusted. Shading indicates one grade lower than IPD meta-analysis.

TA B L E 4 (Continued)
TA B L E 5 Summary of findings and quality of evidence in the IPD meta-analysis  (2) VERY LOW g directly combined with the other trial data sets, which a one-stage analytical approach requires. Previous research indicates that in most cases similar results will be obtained from a one-stage and two-stage IPD meta-analysis. 19,20 Where differences are reported this is because (i) researchers have knowingly or unknowingly made different modelling assumptions and/or (ii) used different estimation methods for deriving point estimates or confidence intervals. Where assumptions and estimations do not vary the two approaches will give similar results, enabling analysts to choose the most convenient procedure.

| How our study compares with others
The results of this case study demonstrate advantages of IPD metaanalysis previously discussed by others. 2

| Strengths and limitations of this study
The statistical analysis of the aggregate data performed here was conducted by the same researchers who performed the original IPD analysis. The authors who previously conducted the IPD risk of bias assessments went back to the papers to conduct risk of bias assessments for this aggregate analysis, enabling consistency in judgement.
However, this could also be viewed as a weakness since they were already exposed to unique information from the IPD analysis and not Downgraded one level for heterogeneity driven by one trial contributing 21.8% of the weight of the analysis, for which the review authors were unable to identify a plausible explanation.
b Downgraded one level for overall high risk of bias due to missing data (29%), and two levels for imprecision due to small numbers of events from a single study, with wide confidence intervals, which include both a harmful effect and no effect.
c Downgraded by two levels for imprecision due to small numbers of events, with wide confidence intervals, which include both a harmful effect and a beneficial effect.
d Downgraded by one level for imprecision due to wide confidence intervals, which include both a harmful effect and no effect.
e Downgraded one level for heterogeneity driven by more than one trial, for which review authors were unable to identify a plausible explanation.
f Downgraded two levels for imprecision due to small numbers of events from a single study, with wide confidence intervals, which include both a harmful effect and no effect.
g Downgraded one level for heterogeneity, for which the review authors were unable to identify a plausible explanation, and two levels for imprecision due to wide confidence intervals, which include both a harmful and a beneficial effect.

TA B L E 5 (Continued)
blind to both bodies of data. The IPD and aggregate meta-analysis results are not directly comparable, as we are not comparing like with like (adjusted versus unadjusted results). Therefore, it was not possible to fully separate the differences that are accounted for by aggregate versus IPD methods as opposed to the adjustment factors used consistently across study datasets in the IPD meta-analysis (sex and family history of atopic disease). Adjustment for key prognostic factors was made in the IPD meta-analysis 12 to increase power in the analysis 21 ; since all included studied were randomized trials, these factors were not expected to vary across intervention groups.
In the aggregate data meta-analysis, we only included published data and did not include any author correspondence, as we had already obtained IPD from trial authors. In the original Cochrane IPD analy-

| CON CLUS ION
In this evaluation of a Cochrane systematic review of skincare interventions for preventing eczema and food allergy, we found significant advantages to using an IPD meta-analysis approach. IPD analysis resulted in higher certainty evidence, new adverse event information and enabled the exploration of treatment estimands not previously calculated by trialists, including those which quantify the impact of adherence and treatment interactions. Significant collaboration and cooperation between trialists and systematic reviewers is needed to achieve IPD meta-analysis, but the process is likely to reduce future research waste by reducing uncertainty of findings.

ACK N OWLED G M ENTS
This analysis compared a new aggregate meta-analysis to the previously published IPD meta-analysis. We wish to thank all Skincare interventions for the prevention of atopic dermatitis (SCiPAD) collaborators for their time, cooperation and enthusiasm for the IPD meta-analysis. We also wish to acknowledge the participants from the original trials, whose deidentified data contributed to the IPD analysis.

CO N FLI C T S O F I NTE R E S T
HCW was director of the NIHR Health Technology Assessment (HTA) Programme until 1 October 2020, which is part of the NIHR that also supports the NIHR systematic reviews programme from which this work is funded; was also chief investigator of the BEEP study, which was funded by NIHR HTA and is included in this review; and has received funds (Nottingham) from the National Institute EVV and SC wrote the first draft of the manuscript. EVV, SC, RJB and MK wrote the manuscript. All authors advised on data interpretation, critically reviewed and contributed to this manuscript.

E TH I C A L S TATEM ENT
Ethical approval was not required for this study which retrieved and

DATA AVA I L A B I L I T Y S TAT E M E N T
The data analysed within this new aggregate data meta-analysis is available from the corresponding author upon reasonable request.