Meta-analysis is usually performed using aggregate data. Aggregate data refer to the summary of individual participant data (IPD) on the investigated outcomes and related characteristics obtained from the trial reports. In the case of binary outcome data, the aggregate data include the number of participants who experienced the event (eg, bond failures per treatment group) from the total randomized in the trial. For continuous outcomes, such as overjet, intercanine, and intermolar width or space closure in millimeters, the researcher extracts the number of participants, the mean and standard deviation of the outcome measured in each trial arm. In terms of characteristics, aggregate data can include the summary of participant- and clinical-related characteristics, such as the mean age, gender distribution across arms, and the mean follow-up period.
Although aggregate data allow for a straightforward analysis, they cannot always uncover the true associations between treatment effects and characteristics. Over time, the well-documented limitations of aggregate data analysis have initiated a paradigm shift for collecting and re-analyzing IPD, whenever possible, after the principles and procedures of conventional systematic reviews. , Systematic reviews of IPD have gained popularity in the therapeutic, diagnostic, and prognostic research in numerous health care fields.
Advantages of IPD meta-analysis
The inclusion of IPD may resolve several shortcomings inherent to the synthesis of trials with aggregate data. For example, when information on the outcome is missing (eg, because of lost-to-follow-up or protocol violations) and important characteristics of participants are not reported in the trial, IPD allows for the application of more sophisticated methods, such as multiple imputations, to properly account for the missing outcome data. In this way, we can investigate the implications of losses to follow-up more thoroughly. In contrast, methods that use the aggregate outcomes to address missing outcome data may suffer from meta-confounding. , Furthermore, access to IPD for outcomes not reported in the journal article (but mentioned in the study protocol) can guard against selective outcome reporting.
Importantly, the provision of IPD for characteristics that naturally differ among the participants, such as age, type of malocclusion, amount of crowding, and overjet, can protect against the effects of ecological bias. Ecological bias arises when associations between the aggregate characteristics (eg, average age) and the treatment effects do not represent the true associations between patient characteristics (eg, individual age) and treatment effects. For instance, several trials have demonstrated the participants’ age to be an important predictor of treatment effectiveness. However, when the average age is similarly distributed across the trials, the meta-regression will fail to demonstrate the true association between age and treatment effectiveness.
Ηaving access to IPD enables the researcher to deliver more precise recommendations. This is accomplished by adjusting the trial inclusion and exclusion criteria to target the population of interest in the systematic review at hand. Similarly, the researcher may explore whether the intervention effectiveness differs among specific subgroups of participants and provide more efficient care on the basis of the participant characteristics. , In the meta-analysis of aggregate data, personalized recommendations are problematic because of ecological bias.
Heterogeneity in the extracted aggregate data because of inconsistent reporting across trials is prevalent in conventional systematic reviews and can make the analysis challenging. For instance, some trials may report the mean of the outcome and the standard deviation per arm, whereas other trials may report only the effect measure together with the P value. Such inconsistencies in the reporting require extra work to standardize the data and make them suitable for meta-analysis. In some cases, inconsistencies in the reporting may render the meta-analysis impossible. Access to IPD enables the standardization of the data across the trials to perform a consistent analysis which may differ from the analyses of the included trials. Furthermore, the availability of IPD facilitates the assessment of model assumptions for their plausibility (eg, the proportionality of hazards in the Cox regression model) and the correction of any errors and inconsistencies in the extracted data. , In split-mouth designs, in which the outcomes are correlated, the variances are reduced as 2 dental quadrants within patients can be randomized to the treatment and control, respectively. In this design, patients act as their own controls; therefore, to correctly combine those trials with parallel trials, the between quadrant correlations are required. This correlation is usually missing from the reported aggregate data but can be easily calculated when IPD are available.
It is often not appropriate to combine randomized trials with observational data in a meta-analysis. However, this is more feasible with IPD because adjusted and possibly unbiased estimates can be obtained for the observational studies allowing for meta-analysis that includes both observational studies and randomized trials.
IPD meta-analysis can either be performed in 1 or 2 stages. In the 1-stage approach, the analysis is conducted, as the name states, using a single analysis. In the 2-stage IPD meta-analysis, the treatment effects are initially estimated for each trial separately via a regression model. The estimated treatment effects from each trial are combined using conventional meta-analysis at the second stage. For a discussion on selecting between a 1-stage and 2-stage meta-analysis of IPD, the reader may refer to a comprehensive review of the methodology.
Challenges of IPD meta-analysis
Depending on the number of trials and the complexity of the planned analysis, the collection and management of IPD can be time-consuming, resource-intensive, and expensive. , The number of collaborators involved, the skills and the software, and the working hours required to obtain the dataset in the appropriate format for the analysis should be considered carefully before initiating a systematic review with IPD. Furthermore, establishing collaborations with trialists to provide the required IPD can be challenging.
The conduct and reporting quality of the included trials may pose further challenges in obtaining IPD. For instance, trials apt to selection biases because of poor planning and conduct may not have recorded important outcomes and clinical characteristics. Running and not yet completed clinical trials can be included in a prospective meta-analysis, can tackle the challenges mentioned above that are typically encountered in completed clinical studies. This prospective meta-analysis will ensure consistency in the recorded outcomes and characteristics and data collection. The Table summarizes the advantages and disadvantages of IPD in meta-analysis.
