A guide to prospective meta-analysis

Step 0: deciding if a PMA is the right methodology

PMA methodology should be considered for a high priority research question for which new studies are expected to emerge and limited previous evidence exists (fig 3):

Priority research question—PMAs should be planned for research questions that are a high priority for healthcare decision makers. Ideally, these questions should be identified using priority setting methods within consumer-clinician collaborations, and/or they should address priorities identified by guideline committees, funding bodies, or clinical and research associations. Often these questions are in areas where important new treatment or prevention strategies have recently emerged, or where practice varies because of insufficient evidence.

New studies expected—PMAs are only feasible if new studies are likely to be included—for example, if the research question is an explicit priority for funding bodies or research associations. Some PMAs have been initiated after researchers learnt they were planning or conducting similar studies, and so they decided to collaborate and prospectively plan to combine their data. In other cases, a research question is posed by a consortium of investigators who then decide to plan similar studies that are combined on completion. A research team planning a PMA can play an active role in motivating other researchers to conduct similar studies addressing the same research question. A PMA can therefore be a catalyst for initiating a programme of priority research to answer important questions.8 Initiating a PMA rather than conducting a large multicentre study can be advantageous as PMAs allow flexibility for each study to answer additional local questions, and the studies can be funded independently, which circumvents the problem of funding a mega study.

Insufficient previous evidence—A PMA should only be conducted if insufficient evidence exists to answer the research question. If sufficient evidence is available (eg, based on a retrospective meta-analysis), no further studies and no PMA should be planned, to avoid research waste.

Fig 3

When to conduct a prospective meta-analysis (PMA)

• Download figure
• Open in new tab
• Download powerpoint

If evidence is available, but is insufficient for clinical decision making, a nested PMA should be considered. A nested PMA integrates prospective evidence into a retrospective meta-analysis, making best use of existing and emerging evidence while also retaining some benefits of PMAs. A nested PMA allows the assessment of publication bias and selective reporting bias by comparing prospectively included evidence with retrospective evidence in a sensitivity analysis. Studies that are prospectively included can be harmonised with other ongoing studies, and with previous related retrospective studies, to optimise evidence synthesis (see step 5).

Step 1: defining the research question and the eligibility criteria

At the start of a PMA, a research question needs to be specified. Research questions for PMAs should be formed in a similar way to traditional retrospective systematic reviews. Guidance for formulating a review question is available in the Cochrane Handbook for Systematic Reviews of Interventions.15 For PMAs of interventional studies, the PICO system (population, intervention, comparison, outcome) should be used. To avoid selective reporting bias, the PMA research question and hypotheses need to be specified before any study results related to the PMA research questions are known.

PMAs are possible for a wide range of different study types—their applicability reaches beyond randomised controlled trials. An interventional PMA includes interventional studies (eg, randomised controlled trials or non-randomised studies of interventions). For interventional PMAs, the key inclusion criterion of “no results being known” usually means that the analyses have not been conducted in any of the trials included in the PMA.

An observational PMA includes observational studies. For observational PMAs, “no results being known” would mean that no analyses related to the PMA research question have been done. As many observational studies collect data on different outcomes, a meta-analysis can be classified as a PMA if unrelated research questions have already been analysed before inclusion in the PMA. For instance, for a PMA on the risk of lung cancer for people exposed to air pollution, observational studies where the relation between cardiovascular disease and air pollution has already been analysed can be included in the PMA, but only if the analyses on the association between lung cancer and air pollution have not been done. In this case, however, little harmonisation of outcome collection is possible (unless the investigators agree to collect additional data).

Step 2: writing the protocol

Key elements of the protocol need to be finalised for the PMA before any individual study results related to the PMA research question are known. These include specification of the research questions, eligibility criteria for inclusion of studies, hypotheses, outcomes, and the statistical analysis strategy. The preferred reporting items for systematic reviews and meta-analyses extension for protocols (PRISMA-P)16 provides some guidance on what should be included. As these reporting items were created for retrospective meta-analyses, however, key adaptations need to be made for PMA protocols (see box 2).

An initial PMA protocol should be drafted before the search for eligible studies, but it can be amended after searching and after all studies have been included if the results of the included studies are not known when the PMA protocol is finalised. The investigators of the included studies can agree on the collection and analysis of additional rare outcomes and these outcomes can be included in a revised version of the protocol.

The final PMA protocol should be publicly available on the international prospective register of systematic reviews, PROSPERO17 (which supports registration of PMAs), before the results (relating to the PMA research question) of any of the included studies are known. A full version of the PMA protocol can be published in a peer reviewed journal or elsewhere.

Step 3: searching for studies

After the PMA protocol is finalised, a systematic literature search is conducted, similar to that of a systematic review for a high quality meta-analysis. The main resources available for identifying planned and ongoing studies are clinical trial registries. Currently, 17 global clinical trial registries provide data to the World Health Organization’s International Clinical Trials Registry Platform.19 Views on the best strategies for searching trial registries differ.20 Limiting the search by date can be useful (eg, only studies registered within a reasonable time frame, taking into account the expected study duration and follow-up times) to reduce the search burden and exclude studies registered earlier that would likely be completed and thus ineligible for a PMA. Ideally, searches should be repeated on a regular basis to identify new eligible studies.

Prospective trial registration is mandated by various legislative, ethical, and regulatory bodies but compliance is not complete.212223 Observational studies are not required to be registered. Hence additional approaches to identifying planned and ongoing studies should be pursued, including searching bibliographic databases for conference abstracts, study protocols, and cohort descriptions, and approaching relevant stakeholders. The existence and possibility of joining the PMA can be publicised through the publication of PMA protocols, presentations at relevant conferences and research forums, and through an online presence (eg, a collaboration website).

Step 4: forming a collaboration of study investigators

Ideally, PMAs are conducted by a collaboration or consortium, including a central steering committee (leading the PMA and managing the collaboration), a data analysis committee (responsible for data management, processing, and analysis), and representatives from each study (involved in decisions on the protocol, analysis, and interpretation of the results). Regular collaboration meetings can be beneficial for achieving consensus on disagreements and in keeping study investigators involved in the PMA process. Transparent processes and a priori agreements are crucial for building and maintaining trust within a PMA collaboration.

Investigators might refuse to collaborate. Refusal to collaborate is less likely in a PMA than in a retrospective individual participant data meta-analysis as reaching agreement to share data is easier if studies are in their planning phases and can still be amended and harmonised after internal discussions. Aggregate data can be included in the PMA even if investigators refuse to collaborate, if the relevant summary data can be extracted from the resulting publications when the studies are completed. The ability to harmonise studies (step 5), however, may be limited if eligible investigators refuse to participate.

Step 5: harmonisation of included study population, intervention/exposure, and outcome collection

When a collaboration of investigators of planned or ongoing studies has been formed, the investigators can work together to harmonise the design, conduct, and outcome collection of the included studies to facilitate a meta-analysis and interpretation. A common problem with retrospective meta-analyses is that interventions are administered slightly differently across studies, or to different populations, and outcome collection, measurement, or reporting can differ. These differences make it difficult, and sometimes impossible, to synthesise results that are directly relevant to the study outcomes, interventions, and populations. In a PMA, studies are included as they are being planned or are ongoing, allowing researchers to agree on how to conduct their studies and collect common core outcomes. The PMA design enables the generation of evidence that is directly relevant to the research questions and thus increases confidence in the strength of the statements and recommendations derived from the PMA.

The ability to harmonise varies depending on the time when the PMA is first planned (fig 4). In a de novo PMA, studies are planned as part of a PMA. For PMAs of interventional studies, a de novo PMA is similar to a multicentre trial: the included trials often share a common protocol, and usually the study population, interventions, and outcome collection are fully harmonised. In contrast, some PMAs identify studies for inclusion when data collection has already finished but no analyses related to the PMA research question have been conducted (outside of data safety monitoring committees). These types of PMAs allow little to no data harmonisation and are more similar to traditional retrospective meta-analyses. Yet they still have the advantage of reducing selection bias as the studies are deemed eligible for inclusion before their PMA specific results are known.

Fig 4

Different scenarios and time points when studies can be included in a prospective meta-analysis (PMA)

• Download figure
• Open in new tab
• Download powerpoint

Harmonisation of studies in a PMA can occur for different elements of the included studies: study populations and settings; interventions or exposures (that is, independent variables); and outcomes collection. For study populations, settings, and interventions/exposures, harmonisation of studies to some degree is often beneficial to enable their successful synthesis. But some variation in the individual study protocols, populations, and interventions/exposures is often desirable to improve the generalisability (that is, external validity) of the research findings beyond one study, one form of the intervention, or narrow study specific populations. The variation in populations also enables subgroup analyses, evaluating if differences in populations between and within the studies leads to differences in treatment effects. If particular subgroups appear in more than one study, additional statistical power for subgroup analyses is also achieved.

Harmonisation of outcome collection requires careful consideration of the amount of common data needed to answer the relevant research questions. These discussions should aim to minimise unnecessary burden on participants and reduce research waste by avoiding excessive data collection, while increasing the ability to answer important research questions. Researchers can also agree to collect and analyse rare outcomes, such as severe but rare adverse events, that their individual studies would not have had the statistical power to detect. Collaborations should be specific on exactly how shared outcomes will be measured to avoid heterogeneity in outcome collection and difficulties in combining data. The COMET (core outcome measures in effectiveness trials) initiative (www.comet-initiative.org/) has introduced methods for the development of core outcome sets, as detailed in its handbook.24 These core outcome sets specify what and how outcomes should be measured by all studies of specific conditions to facilitate comparison and synthesis of the results. For health conditions with common core outcome sets, PMA collaborators should include the core outcomes, and also consider collecting other common outcomes that are particularly relevant for the specific research question posed. Not all outcomes have to be harmonised and collected by all studies: individual studies in a PMA have more autonomy than individual centres in a multicentre study and can collect study specific outcomes for their own purposes.

The improved availability of common core outcomes in a PMA has recently been shown in a PMA of childhood obesity interventions.25 Harmonisation increased from 18% of core outcomes collected by all trials before the trial investigators agreed to collaborate, to 91% after the investigators decided to collaborate in a PMA.

Step 6: synthesising the evidence and assessing certainty of evidence

When all the individual studies have been completed, data can be synthesised in a PMA. For aggregate data PMA, results are extracted from publications or provided by the study authors. For individual participant data PMA, the line by line data from each participant in each study must be collated, harmonised, and analysed. This process is usually easier for PMAs than for traditional, retrospective individual participant data meta-analyses because if outcome collection and coding were previously harmonised, fewer inconsistencies should arise. If possible, plans to share data should be outlined in each study’s ethics application and consent form. For PMAs that are planned after the eligible studies have commenced, amendments to ethics applications may be necessary for data sharing. To assure independent data analysis, some PMAs appoint an independent data manager and statistician who have not been involved in any of the studies. The initial time intensive planning and harmonisation phase is followed by a waiting period when all the individual studies are completed before their data are made available and synthesised. During this middle period, PMAs usually demand little time and can run alongside other projects.

For studies where data safety monitoring committees are appropriate, it might be sensible for the committees to communicate and plan joint interim analyses to take account of all the available evidence when making recommendations to continue or stop a study. The PMA collaboration should consider establishing a joint data monitoring committee to synthesise data from all included studies at prespecified times. Methods for sequential meta-analysis and adaptive trial design could be considered in this context.26

When all studies have been synthesised, the methodological quality of the included studies needs to be appraised with validated tools, such as those recommended by Cochrane.2728 The certainty of the evidence can be assessed with the grading of recommendations assessment, development and evaluation (GRADE) approach.29

Step 7: interpretation and reporting of results

Generally, the quality of the evidence derived from a PMA, and the extent to which causal inferences can be made, directly depend on the type and quality of the studies included in a PMA. The prospective nature of interventional PMAs make them similar to large multicentre trials, allowing for causal conclusions to be drawn rather than only associations, as sometimes suggested for traditional retrospective meta-analyses. The results of observational PMAs should generally be interpreted as providing associations, not causal effects, as only the results of observational studies are included. But with modern methods for causal inference from observational studies, justification for supporting conclusions about causality can sometimes be found.33

Currently no PMA specific reporting standards exist, but where applicable, PMA authors should follow the PRISMA-IPD (PRISMA of individual participant data) statement34 if they are reporting an individual participant data PMA, or the PRISMA statement35 if they are reporting an aggregate data PMA. As well as the PRISMA items, authors of PMAs need to report on identification of planned and ongoing studies, the PMA timeline, collaboration policies, and outcome harmonisation processes.

Discussions about methodology and interpretation of the results among all collaborators can sometimes be difficult to navigate, particularly if the results from the combination of the studies contradict the results of some of the individual studies. Although these discussions can be demanding and time consuming, robust discussion among experts can lead to well considered and high quality publications that can directly inform policy and practice.

For the successful management of a PMA collaboration, an explicit authorship policy should be in place. One model is to offer authorship to each member of the secretariat, and one investigator from each included study, for the main PMA publication, assuming they fulfil the authorship criteria of the International Committee of Medical Journal Editors (ICMJE). This model incentivises ongoing involvement and allows for multiple viewpoints to be integrated in the final publication. The collaborators usually agree that the final PMA results cannot be published until the results of each study are accepted for publication, but this is not essential.