The dataset (.csv or .xlsx, per arm and outcome) and either a results table (.csv) or the manuscript, from which the reported estimates are extracted with a verbatim quote.

Fixed-effect results?

Not re-run yet; they come back as NOT COMPARED, not as failures.

How close is close enough?

The engine returns two decimals, so the tolerance floor is 0.005 on the estimate and on the confidence limits.

Statistical audit Auditors Premium

Every pooled estimate, recomputed from your own data

The forest plot was made in March, the dataset was corrected in May, and the manuscript still reports the March numbers. The statistical reviewer will find it.

Run this tool See all tools

You give

The per-arm dataset (.xlsx or .csv: study, outcome, n/mean/SD or events/total per arm) and the table of reported results (.xlsx/.csv: outcome, measure, effect, CI, I², model) or the manuscript (.docx/.md).

You get

A spreadsheet with the per-outcome comparison (reported vs recomputed, delta, tolerance, flags) and the per-study data used, plus the .md and .docx report with the verdict REPRODUCED / PARTIALLY / NOT REPRODUCED on the first line.

How it works

# Statistical audit
$ scopetools run audit-stats
Model-assisted, with deterministic checks.
Each outcome is re-run in our R engine (meta/metafor, the same endpoints as the Analysis): random effects by REML with Hartung-Knapp when the paper says so, MD/SMD or RR/OR following the reported measure. It compares effect, CI, I² and τ² within tolerance (effect ±1% or the last decimal, CI ±2%, I² ±2 points) and, when they disagree, tries the alternative measure and SD×√n. If you send only the manuscript, one model step extracts the reported numbers, and each row is kept only with a verbatim quote verified in the text.

Deterministic matching of the reported outcomes to the dataset outcomes; then each one is re-run in the MetaScope engine (the same R behind PlotScope) with the model the manuscript declares. A miss is retried with the alternative measure (RR↔OR, MD↔SMD) and SD from SE — a hit on the retry is PARTIAL with a flag, never REPRODUCED. You get the comparison table with the recomputed values.

What it costs

30 credits per manuscript (US$1.50), for subscribers, whatever the number of outcomes.

Subscribers only — charged in credits per run, like every other tool.

Charged only when it delivers — a failed run costs nothing.

Credits are shared across every MetaScope tool — buy once, spend anywhere.

Limits

Reproduces random-effects models (the engine does not re-run fixed-effect). Each re-run renders a forest on the server: dozens of outcomes take a few minutes. It does not judge whether the analysis was right, only whether the reported number comes out of the data.

Frequently asked questions

What do I send?: The dataset (.csv or .xlsx, per arm and outcome) and either a results table (.csv) or the manuscript, from which the reported estimates are extracted with a verbatim quote.
Fixed-effect results?: Not re-run yet; they come back as NOT COMPARED, not as failures.
How close is close enough?: The engine returns two decimals, so the tolerance floor is 0.005 on the estimate and on the confidence limits.

Numbers reproduce? Get the full reviewer's statement with the Manuscript audit.