Meta-Analysis of Randomised Trial Data: Approaches and Practice
Meta-Analysis Sample Work
Statistical methods and results from a stroke thrombectomy meta-analysis
This sample shows the analysis and results chapters of a meta-analysis, under the working title Meta-analysis of individual patient data from randomised trials: comparison of approaches and practice suggestions. The data relate to endovascular stroke treatment. Outcomes include final reperfusion on the modified Thrombolysis in Cerebral Infarction (mTICI) scale, the modified Rankin Scale (mRS) and change in National Institutes of Health Stroke Scale (NIHSS) score. Results are compared by occlusion site and by thrombectomy device. The chapter is aimed at stroke researchers and statisticians who need a model for setting out pooled estimates clearly.
Individual patient data versus published summaries
In an individual patient data (IPD) meta-analysis, investigators obtain participant-level records from each trial instead of extracting summary figures from publications. This allows outcomes to be defined consistently and subgroups to be analysed reliably. IPD can be analysed in two stages, where each study is analysed separately and the estimates are then pooled, or in one stage, where all participants are modelled together while accounting for clustering by study. The sample includes both levels of data. Its descriptive tables report patient-level distributions, and its forest plots pool study-level risk ratios in the way a two-stage approach does. Reporting for this design is set out in the PRISMA-IPD extension.
The statistical plan in plain terms
The PRISMA chart describes searches in PubMed, Embase and other databases, screening by two independent reviewers, and 50 articles in the final analysis. The analysis plan states that:
- risk ratios with 95% confidence intervals were calculated for dichotomous outcomes, and mean differences for continuous outcomes;
- heterogeneity was assessed with the Cochran Q statistic and I², with values below 50% treated as low;
- estimates were pooled with the DerSimonian and Laird random-effects model in Review Manager 5.3;
- a two-sided p-value of 0.05 marked statistical significance.
Logistic and linear regression are also described for modelling binary and continuous outcomes from patient characteristics.
What the tables and forest plots report
Table 1.1 shows final mTICI for 130 records: grade 3 in 48 (36.92%), 2b in 40 (30.77%), 2a in 13 (10.00%) and 0 in 6 (4.62%). The remaining 23 were either not reported or scored on another scale. Tables 1.2 and 1.3 give means, medians, standard deviations and confidence intervals for final mRS, NIHSS change and age, split by sex and by occlusion site (vertebrobasilar, M1, M2 and ICA terminus).
| Comparison | Studies | Risk ratio (95% CI) |
|---|---|---|
| IVT | 25 | 0.41 (0.25–0.66) |
| IAT | 25 | 0.40 (0.26–0.62) |
| Vertebrobasilar | 33 | 0.63 (0.43–0.94) |
| ICA terminus | 30 | 0.55 (0.37–0.82) |
| M1 | 27 | 0.73 (0.48–1.10), not significant |
| M2 | 31 | 0.38 (0.24–0.59) |
| NIHSS change, Solitaire | 9 | 3.17 (1.36–7.38) |
| NIHSS change, Trevo | 4 | 3.00 (0.85–10.63), not significant |
| NIHSS change, Penumbra aspiration | 9 | 1.59 (0.79–3.20), not significant |
Each forest plot is followed by a funnel plot and a paragraph giving the Z test, χ², I² and τ².
Reading the descriptive tables with cell sizes in mind
The patient-level tables show subgroup patterns but need careful reading. By sex, mean final mRS was 3.50 ± 2.37 in males and 3.83 ± 2.02 in females, and mean NIHSS change was 13.86 ± 7.88 and 11.83 ± 5.53 respectively. By occlusion site, vertebrobasilar cases had the largest mean NIHSS change (17.50 ± 9.25). The M2 row shows how small cells affect the estimates: its confidence interval for NIHSS change runs from −48.03 to 79.03, and no final mRS is given. Both point to very few observations. Giving the number of patients in each cell, next to the mean and interval, would let readers judge such rows immediately and stop a wide interval being mistaken for a meaningful difference.
Points to settle before submission
A reader adapting this chapter should define the "event" in each comparison precisely. Phrases such as "presence and absence of events" need to name the outcome and the comparator. The funnel plots should be interpreted, not only displayed, especially for subgroups with fewer than ten studies, where tests of asymmetry are unreliable. When I² is 0%, the authors should explain why they kept a random-effects model and present a certainty rating such as GRADE alongside the results. For a true IPD analysis, the methods should also describe how trial datasets were obtained, checked and harmonised.
Pubrica's meta-analysis service covers both aggregate and patient-level syntheses. Model choice and regression work can be handled by the biostatistics and statistical programming team, and results chapters can be refined through scientific editing.
Questions about this sample
Which software and pooling model does the sample use?
All analyses were run in Review Manager (RevMan) 5.3 from the Nordic Cochrane Centre. Risk ratios were pooled with the DerSimonian and Laird random-effects model, with heterogeneity assessed by the Cochran Q statistic and I², and statistical significance set at a two-sided p-value of 0.05.
Which comparisons were not statistically significant?
Three comparisons had confidence intervals crossing one: M1 occlusions (RR 0.73, 95% CI 0.48 to 1.10) and two device subgroups for change in NIHSS score, Trevo (RR 3.00, 0.85 to 10.63) and Penumbra aspiration (RR 1.59, 0.79 to 3.20). All other pooled comparisons in the sample reached significance.
What does the mTICI scale measure?
The modified Thrombolysis in Cerebral Infarction scale grades how much blood flow is restored to the affected territory after endovascular treatment, from 0 (no reperfusion) through 2a and 2b to 3 (complete reperfusion). In the sample, grade 3 was the most common final result, at 36.92% of 130 records.
How does an IPD approach differ from what RevMan pools here?
RevMan pools summary statistics such as event counts and risk ratios from each study. An IPD analysis starts from each participant's record, so outcomes and subgroups can be defined the same way across trials and modelled in one or two stages. The sample's patient-level tables point towards that richer level of analysis.
Give yourself the academic edge today
Each order includes
- On-time delivery or your money back
- A fully qualified writer in your subject
- In-depth proofreading by our Quality Control Team
- 100% confidentiality, the work is never re-sold or published
- Standard 7-day amendment period
- A paper written to the standard ordered
- A detailed plagiarism report
- A comprehensive quality report
Enquire Now
