Human-verified effect-size data preparation for meta-analysis.
The Paper 6 meta-analysis pools 236 studies of internationalisation and firm performance. Here is where they were conducted, when they appeared, and how strong the effect they reported was. Every number is read from the locked M-AIDA dataset.
Fifty years of evidence, assembled and locked.
From 1978 to 2026, the corpus grows from a handful of Anglo-American studies to 236 studies across 35 economies. Every record passed PRISMA screening, human verification and an immutable lock before analysis.
Explore the atlas →A small positive effect, honestly reported.
The pooled correlation is r = .074 with substantial heterogeneity (I² = 62%). Correcting for publication bias lowers it to .035. The dissertation treats these corrections as part of the finding, not a footnote.
See the forest plot →Random-effects r̄ = .074, 95% CI [.060, .088]; heterogeneity I² = 62%. Trim-and-fill adjusted r̄ = .035. Three-level model estimated in R (metafor). Search window 1977-2026; analytic set k = 236 studies, K = 286 effect sizes after PRISMA 2020 deduplication.
Pooling evidence is an old idea. Doing it well is a modern discipline.
From collecting case notes in antiquity to formal estimators and, in 1989, meta-regression in economics, M-AIDA sits at the end of this line, as the extraction step that keeps the chain auditable.
The Hippocratic tradition compared and pooled case records, the intuition behind synthesis, long before its mathematics.
Averaging correlations across typhoid-inoculation studies, widely read as the first true quantitative synthesis.
Statistical Methods for Research Workers gives synthesis its first repeatable procedure for combining results.
His AERA presidential address christens the field and reframes the literature review as data in its own right.
The earliest internationalisation-performance study inside this project's 1977-2026 search window.
Statistical Methods for Meta-Analysis turns a practice into a rigorous, teachable methodology.
Meta-regression analysis carries synthesis into economics and, with it, international-business research, where effects vary by context, not just by chance.
MAER-Net guidance and PRISMA 2020 make transparency and reproducibility the price of admission.
An AI-assisted extraction step with a human at the gate: the model proposes, the investigator verifies, the record is locked. Here it synthesised 236 studies for the P6 meta-analysis of internationalisation and firm performance worldwide: 35 economies plus cross-border samples, 1978 to 2026.
Note
The literature, on the couch: every study a marker, the pooled truth a diamond. M-AIDA's job is only to read the numbers off the page, faithfully, and on the record.
Figure 1. A timeline of quantitative research synthesis, from the collection of case records in antiquity to meta-regression in economics (Stanley & Jarrell, 1989) and AI-assisted, human-verified extraction today. Square markers denote individual milestones; the diamond denotes the pooled synthesis. Source: the authors, after O’Rourke (2007).
From a PDF to a locked, analysis-ready record.
Extract
The full text is read, and a language model proposes the focal statistic, r, or a t / β converted to r by published rules, with a data-year and a sample size.
Verify
The investigator inspects every field, overrides what is wrong, and assigns the moderators the model is never allowed to guess, institutional regime, digital phase, context.
Lock
Approved records are sealed with a timestamp and cannot be edited. Only locked records are exported to R (metafor) for estimation, the audit trail is the point.
Huong AI sets the rules; the formulas do the rest.
Reported directly: kept as published.
From a t statistic; df = n − 2 when unreported (Cohen, 1988).
From a standardized beta, always flagged (Peterson & Brown, 2005). Anything below 0.70 waits for human review.
Built so a committee can check it.
The division of labour is deliberate and disclosed: the model reads statistics, the investigator decides everything that matters, and the provider is swappable so the method never depends on one vendor.
Faster than hand-coding, safer than full automation.
We built M-AIDA to turn hundreds of hours of manual coding into a few, without giving up the rigour a doctoral thesis demands.
School of Economics, Can Tho University, Vietnam
ORCID 0000-0002-7711-2487 · 0000-0003-0667-3137 · Correspondence: thuyhuongctu@gmail.com
Six capabilities, one record-level boundary.
M-AIDA prepares data; it does not run the meta-analysis. Each function below serves a single stage of that preparation, and every analysed record leaves the tool only after it is verified and locked.
A preparation stage that fits the evidence chain.
M-AIDA does not replace screening tools or the estimator. It occupies the narrow stage between full-text statistics and a verified dataset, and hands that dataset to established methods without re-inventing them.
One specialized stage, not another all-in-one platform.
M-AIDA is a human-verified research software system for extracting, transforming, governing and releasing effect-size records from primary-study documents for a later meta-analysis. It does not screen the literature and it does not estimate the model; it prepares the records that sit between those two jobs.
Shared with existing platforms
- A language model reads full-text PDFs and proposes structured fields.
- A human inspects, overrides and approves what the model proposes.
- Each extracted field links back to the passage it came from.
- Actions leave an audit trail. Rayyan, Elicit, Covidence and DistillerSR all do these.
Where M-AIDA specializes
- Effect-size conversion with the formula, inputs and result stored on every record.
- An explicit lifecycle: machine-proposed → verified → locked → released.
- A record-level release boundary: only records that pass the release conditions are exported.
- Auditable open source with a version DOI, tied to one dissertation's method.
The locked dataset becomes a forest plot.
Once the records are verified and locked, they feed a three-level random-effects meta-analysis in R. Each square below is one study's effect size r; the amber diamond is the pooled estimate across the corpus. The point of M-AIDA is the effect sizes themselves, prepared cleanly enough to trust.
How the evidence base grew, and where.
A Western, English-language field
The first studies come from the United States and the United Kingdom: large multinationals, measured against accounting returns.
The field globalises
Europe and Latin America join, and the first East Asian evidence appears. The question stops being a purely Anglo-American one.
Asia-Pacific takes the lead
China, Taiwan, India and Korea become the centre of gravity. This is the evidence the dissertation is built to explain.
The corpus, locked
236 studies across 35 economies, plus cross-border samples, all verified and locked in M-AIDA before the three-level model was run in R.
Bubbles are placed at country centroids from the locked P6 dataset (35 located economies, 194 studies). A further 42 studies use cross-border or multi-region samples and are counted in the totals but not placed on the map. Colour encodes the country mean of the extracted Pearson r.
From a full-text PDF to a locked dataset.
1 · Upload a full-text PDF
2 · Verify each conversion, then lock
| Study | Source statistic | Formula | Effect size r | Confidence | Status | Action |
|---|
3 · Release only the locked records
Open for reading, citation, and verification.
- Full source, published openly
- FastAPI backend · React frontend
- Academic Source-Available Licence
- Protocol registered before extraction (PRISMA 2020, Item 24a)
- osf.io/z37kn · DOI 10.17605/OSF.IO/Z37KN
- Registered 18 May 2026 · public
- 10.5281/zenodo.21282516 (all versions)
- 10.5281/zenodo.21282517 v7.1.1
- Citable, versioned, permanent
- Copyright registration in progress
- Can Tho University & the two authors
- Non-commercial academic use with citation
The method has a paper trail.
Works underpinning the history above and the effect-size conventions M-AIDA applies (APA 7).
- Buckley, P. J., Dunning, J. H., & Pearce, R. D. (1978). The influence of firm size, industry, nationality, and degree of multinationality on the profitability of the world’s largest firms. Weltwirtschaftliches Archiv, 114(2), 243-257.
- Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum.
- DerSimonian, R., & Laird, N. (1986). Meta-analysis in clinical trials. Controlled Clinical Trials, 7(3), 177-188.
- Fisher, R. A. (1932). Statistical methods for research workers (4th ed.). Oliver & Boyd.
- Glass, G. V. (1976). Primary, secondary, and meta-analysis of research. Educational Researcher, 5(10), 3-8.
- Havránek, T., Stanley, T. D., Doucouliagos, H., Bom, P., Geyer-Klingeberg, J., Iwasaki, I., Reed, W. R., Rost, K., & van Aert, R. C. M. (2020). Reporting guidelines for meta-analysis in economics. Journal of Economic Surveys, 34(3), 469-475.
- Hedges, L. V., & Olkin, I. (1985). Statistical methods for meta-analysis. Academic Press.
- O’Rourke, K. (2007). An historical perspective on meta-analysis: Dealing quantitatively with varying study results. Journal of the Royal Society of Medicine, 100(12), 579-582.
- Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., … Moher, D. (2021). The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ, 372, n71.
- Pearson, K. (1904). Report on certain enteric fever inoculation statistics. British Medical Journal, 2(2288), 1243-1246.
- Peterson, R. A., & Brown, S. P. (2005). On the use of beta coefficients in meta-analysis. Journal of Applied Psychology, 90(1), 175-181.
- Stanley, T. D., & Jarrell, S. B. (1989). Meta-regression analysis: A quantitative method of literature surveys. Journal of Economic Surveys, 3(2), 161-170.
