Читать книгу Individual Participant Data Meta-Analysis - Группа авторов - Страница 46

Unlike a conventional systematic review, the pace and progress in an IPD meta‐analysis project is not entirely under the control of the research team. The time required for some activities can be reasonably estimated, such as producing a protocol, running searches, eligibility screening, and preparation of a data dictionary (Section 4.2). However, the time taken to assemble, code and check the quality and applicability of received IPD is highly dependent on the actions of the data providers (Sections 4.4 to 4.6), and so is often difficult to estimate in advance.

There is inevitably a lag between inviting trial investigators to participate and receiving IPD. After having been persuaded personally of the value of the proposed project (which may itself take some time), some trial investigators will need to obtain institutional or other approval to release data. This can take several months, particularly if granted at institutional review board or committee meetings, which may be scheduled infrequently. Data and documentation may then have to be located in stores and archives, and prepared for release.

After data have been received, the central research team then need to check the trial IPD and send queries to trial investigators and resolve them. Again, this can take considerable time, particularly if the lead trial investigator needs to liaise with others such as the trial data manager or statistician, to determine exactly how data have been coded or which variables were used to define outcomes in the original trial analyses. It is important to remember that whilst the IPD meta‐analysis may be a top priority for the research team leading it, the same may not be true for trial investigators, who are likely to have many competing demands for their time.

Another important consideration when estimating timing and resource requirements is whether those responsible for included trials would prefer to prepare and code their IPD in the format required for the IPD meta‐analysis project, or whether they prefer to send it in its original format and for the central research team then to do any recoding required. Even when trial investigators prepare data, the central research team still need to spend some time assimilating it. An estimated minimum of around a day should be allowed for processing the data for each trial, and longer if the suggested coding has not been followed as intended. Dealing with data that have not been prepared by the trial investigators can be very time consuming, depending on the complexity of the data and the way that it is structured. For a very large and complex trial dataset, with variables stored across many electronic forms and with, for example, repeated follow‐up and measurements, it may take a team member several weeks to fully understand it, write code to extract and transform data to meta‐analysis format, and to check thoroughly that this has generated the correct values. It can be difficult to estimate what proportion of trials this will apply to. Our experience is that most trial investigators are willing to prepare and recode data, but that it would be wise to factor in time for the research team to prepare data from at least 20% of trials.

Unless there are very specific assurances that IPD will be released rapidly and that any issues encountered during data checking will be resolved promptly by trial investigators, as a rule of thumb, upwards of a year should be planned for collecting and checking the IPD from the full set of eligible trials. IPD collection, cleaning and harmonisation for large projects involving many trials may take much longer than a year, and typically 18 to 24 months is needed prior to the IPD meta‐analysis itself. Much of the elapsed time is taken up by communications and by trial investigators gaining approvals. As older trials may be difficult and time‐consuming to trace, and agreement more difficult to reach for controversial topics, these sorts of issues should also be factored into planning project timelines.

Likewise, there will usually be a lag between requesting IPD from a data repository and subsequent provision of that IPD. There will usually be a process for approving project proposals prior to the release of a trial’s IPD that may take several months to complete. As noted earlier, experience to date has been mixed; whilst some teams have found that communication with the data providers has become more streamlined and that pre‐coded datasets can reduce the time taken to prepare data for analysis,⁷⁴ others have found that obtaining permissions when multiple data owners are involved has been difficult.⁷⁵ Based on the limited experience so far, it is sensible to still factor in at least a year to obtain the necessary approvals and gain access to IPD from data repositories and data‐sharing platforms.

As trial data usually arrive sporadically over a period of time, data checking is usually done concurrently with data collection and coding. Data are checked as soon as possible after receipt, and any issues discussed and resolved with trial investigators, usually with a series of iterations (Section 4.5.4). This often includes analysing each trial individually and comparing with any published analyses, both to understand any differences that may arise as a result of, for example, using different outcome definitions, and to ensure that the central research team has understood the data correctly. The time and resource needed for checking will depend on how clean the data are on arrival and the extent of checking required. Allowing about three or four days per trial for carrying out data checking is a good starting point. It is useful to allow extra time beyond when the last dataset might be received, in order to complete the checking processes.

Sufficient time must also be factored into timelines for the meta‐analyses, and it is important that this critical phase does not get squeezed if data collection takes longer than anticipated. There is sometimes a misconception that statistical analyses are straightforward and can be done at the click of a button, but this is never the case. Assuming data are cleaned and ready for analysis, usually at least six months will be required for the statistical phase of an IPD meta‐analysis of randomised trials to evaluate treatment effects. Often, multiple analyses will be needed, for example to analyse multiple outcomes, subgroups, and participant‐level covariates, as well as sensitivity analyses and production of associated summary tables and graphs. If complex modelling is planned, such as using multiple imputation to deal with missing data (Chapter 18), analysing non‐linear relationships (e.g. for treatment‐covariate interactions; Chapter 7) or network meta‐analysis (Chapter 14), then six to 12 months will be a more sensible time frame. Furthermore, although most problems with data should be identified when checking data, issues may still arise during synthesis that require further communication with the trial investigators.

Generally, IPD meta‐analysis projects will take upwards of two years to complete, and sometimes longer depending on how many trials are involved and the complexities of negotiating collaboration, data coding, checking, cleaning and analysis. Delays may be outside the control of the project team. Flexibility in staffing and scheduling is needed to accommodate this, which can make setting and meeting funder milestones challenging. In our experience, extensions to the original timelines agreed with funders may be required, particularly if key trials delay sending their IPD. Work may be most intense at the beginning and end of a project, and this can be borne in mind when planning resourcing and staffing. Nonetheless, it is important that projects are actively managed at all times and that the research team keep on top of projects and moving them forward.

Prospective IPD meta‐analysis projects will additionally need to run to timelines that accommodate those of the participating trials (Section 3.12).