The second read a finished manuscript gets before it goes out — what it checks, the two moments it earns its place, what it costs, and what automation does and does not change.
A statistical review before journal submission is a methodological and statistical read of a finished manuscript, run for the author rather than the journal. It asks, ahead of time, the two questions every editor and methodological reviewer will ask later: does the study design support the claim being made, and do the reported numbers — and the interpretation built on them — hold together. Methodological and statistical problems are among the most common reasons clinical manuscripts are turned away, and many are visible to an editor in minutes; we catalog the patterns in why manuscripts get rejected for statistics and why editors desk-reject before peer review. Catching the same problems a week before submission, instead of in a decision letter months after, is the entire point.
Two boundaries keep the term honest. A pre-submission review is not peer review — it does not replace the journal's process, it prepares a manuscript for it. And it is not design-stage statistical input: the highest-leverage statistical help happens before data collection, when endpoints, sample size, and the analysis plan can still change. No review of a finished draft, human or automated, recovers a design decision that was settled before the data existed. This article is about the read a finished draft gets.
Whoever performs it — a departmental biostatistician, a commissioned reviewer, or software — a competent pre-submission statistical review keeps returning to six questions.
1. Design–claim fit. Can this design carry that claim? A retrospective cohort written in causal language, a single-arm series read as comparative efficacy, a secondary endpoint quietly promoted to the headline — the mismatch between what was done and what is concluded is the most consequential thing a review can catch, because no amount of correct arithmetic repairs it.
2. Statistical appropriateness. Was the analysis right for the data and the question: paired measurements treated as independent, clustered patients analyzed as if they were not, a normal summary forced onto skewed data, multiple comparisons made without a plan for them.
3. Numerical consistency. Do the numbers reconcile — percentages against their denominators, subgroup counts against the analyzed total, reported means against what whole-number data can produce, each p-value against its test statistic and degrees of freedom. These checks are arithmetic: they can be recomputed from the manuscript alone, and the base rate justifies running them — when statcheck was run across eight psychology journals, about half the papers had at least one inconsistent p-value ↗. We explain the two best-known checks in the GRIM test and when a p-value doesn't match its test statistic.
4. Reporting-guideline adherence. CONSORT for randomized trials, STROBE for observational studies, TRIPOD for prediction models, PRISMA for systematic reviews — item by item. Many journals require the completed checklist at submission, and a missing item is a concrete, citable reason to send the paper back. Our walkthroughs take each one in turn: CONSORT, STROBE, TRIPOD, PRISMA.
5. Results–conclusion calibration. Is the strength of the conclusion matched to the effect size and precision actually shown? A wide interval crossing the null read as demonstrated equivalence, or an abstract claiming more than the primary analysis shows, draws reviewer fire even when every number in the tables is right.
6. Reference integrity. Do cited DOIs and PMIDs resolve, and do they resolve to the articles they claim to be? Citation errors propagate silently, and this is the most mechanical check of the six — we expose our reference layer as a free reference-integrity checker you can run today.
The concrete version of this list is public. The full catalog of deterministic checks RigorMD runs — what each check needs from the text, and the severity it can reach — is published at every check we run →, including a worked recomputation you can follow line by line.
Before first submission. The unit of cost in publishing is the rejection cycle: a round trip through formatting, submission, review, and a decision letter is measured in months, and a statistical problem a reviewer catches costs a full cycle to fix. A pre-submission review moves that discovery to the week before submission, while the fix costs an afternoon. It earns the most where statistical support was thinnest — the unfunded retrospective cohort drafted without a statistician is exactly the manuscript whose avoidable errors are still all in it.
Before resubmission, after statistical reviewer comments. A decision letter with statistical critiques defines precisely what the revision must answer — and a revision can introduce new inconsistencies of its own while fixing the old ones. Reviewing the revised draft before it goes back checks both directions. How to answer the hard comments point by point is covered in responding to Reviewer 2's statistics comments; if you are at that moment now, answer the reviewers with a checked revision describes how a re-review fits.
And when it is not the tool. If the study is still being designed, a manuscript review is the wrong instrument entirely — find design-stage statistical input first, because that is where endpoints, power, and the analysis plan are still negotiable. A review of a finished draft is the second-best time to catch a problem; the best time is before the problem is committed to data.
A commissioned methods and statistics review of a finished manuscript by a contracted human expert typically runs $150–$400 at published list prices, more for fast turnaround or a comprehensive methodological critique, with turnaround measured in days. A departmental biostatistician, where you have one, is usually institutionally or grant-funded — the cost is the queue, not a price, and unfunded projects often wait longest. Single-purpose automated checkers are free and narrow: excellent at the one thing each verifies, silent on everything else. RigorMD's automated pre-submission review is $30, returned in hours.
Those are different instruments, not different prices for the same thing. Thirty dollars does not buy what a commissioned expert's judgment buys; it buys breadth, recomputed arithmetic, and availability on demand. Which instrument fits which manuscript — including when the right answer is a human statistician — is the subject of your options for statistical review before submission, with a side-by-side table. Current pricing is always at rigormd.com/pricing.
What the human gives you has no automated equivalent. A statistician brings design-stage input, field context, dialogue, and accountability — a person who can shape the analysis before the data exist and defend it to reviewers afterward. The constraint is access: academic statistical units are capacity-strained ↗, and limited time concentrates on funded work. When design-stage involvement is available to you, take it — that category is worth protecting for the questions that genuinely need a human expert.
What automation changes is independence, determinism, and availability. RigorMD's review reads a manuscript with two independent engines, blind to each other, and reconciles them — where they disagree, the disagreement is surfaced in the report rather than averaged away. Beneath them, a deterministic forensic layer recomputes what the reported numbers allow — the same checks, the same way, on every manuscript — and shows its work, so a flagged number comes with arithmetic you can re-derive yourself. And it is available tonight, in hours, without a queue. No individual reviewer, however good, is two independent readers plus a recomputation engine; no automated system, however broad, is a colleague with judgment and skin in the game.
The difference is easiest to see in the open. The pipeline — classification, blind two-engine appraisal, deterministic forensics, reconciliation, adversarial verification of serious findings — is described in how RigorMD reads a manuscript. And for a public demonstration, we ran the engine against a landmark trial later retracted for randomization irregularities: it raised the documented problem independently, and — the harder half — sized its severity against what the republished paper had disclosed. That walkthrough is catching what the record already knew.
And the limits, stated plainly. An automated review flags; it never certifies. It reads the submitted draft, not your raw dataset. It cannot attend the next design meeting, defend a modeling choice to a skeptical reviewer, or take responsibility for an analysis. It is a complement to peer review and to statistical expertise — not a substitute for either.
The fastest way to calibrate is to read one. A de-identified full sample report → shows what the findings, severities, and quoted grounding look like on a real manuscript. The check catalog is at /checks, the methodology at /methods, and if you are between a decision letter and a resubmission, start at /respond-to-reviewers. The pre-submission review is $30. The goal, whichever route you take, is narrow and honest: fewer avoidable rejection cycles.