Methods Sections That Make Your Work Reproducible
Oct 11, 2026
"The journal caps us at 4,000 words and the methods alone are eating 1,600. Can I trim them to the essentials and write 'as previously described' for the rest?"
That message came from an assistant professor in public health a few months ago. It was the third version of the question I'd seen that term. My answer was short, and I'll repeat it here because I think it's the most useful thing I can say about methods sections.
Cut your introduction first. Then cut your discussion. Touch the methods last, and when you do, move detail into a supplement instead of deleting it. Only write "as previously described" if the earlier description is complete, publicly accessible, and matches what you actually did this time. In my experience, that third condition fails more often than people expect.
She cut 400 words from the introduction, left the methods alone, and the paper went out for review without a single "insufficient detail" comment. That rarely happens.
Why the methods section protects you from rejection
Most researchers treat the methods as the boring part, the section you write fast so you can get to the results. Reviewers read it differently. For a careful reviewer, the methods section is where the paper is won or lost, because it's the only part that tells them whether your results can be believed.
When I review, I can't judge a finding I can't evaluate. If I don't know how participants were selected, how outcomes were defined, or which model produced your p-values, then I have two options. I can ask, which adds a revision round. Or I can assume the worst, which leads to a reject recommendation. Tired reviewers tend to assume the worst.
Most advice on this gets one thing wrong. People say reviewers reject papers for weak results. Sometimes they do. More often, I see papers rejected because the reviewer couldn't tell whether the results were weak. Vague methods turn a fixable question into a fatal doubt. Detailed methods turn a hostile reviewer into one who argues with your choices, and you can answer an argument. You can't answer suspicion.
The test I use: could a competent stranger redo this?
I don't ask whether the methods are "complete." That word means nothing in practice. I ask whether a competent researcher in your field, who has never met you and can't email you, could repeat the study from the paper and its supplements alone.
I learned to take that test seriously from a co-author on one of my own papers. Tomas ran the bench work for a study we published together. About three years later, a student in his lab tried to extend it and couldn't reproduce our baseline. They spent most of a semester troubleshooting. The cause was a centrifugation step we'd described as "brief." In the lab, everyone knew that meant two minutes at a particular speed. On paper, it meant nothing at all. Our own group couldn't reproduce our own work from our own paper.
Tomas sent me a one-line email afterwards: "We wrote that methods section for ourselves." He was right, and it's the most common failure I see. Authors write for people who were in the room. The paper's real audience wasn't.
What detail reviewers actually expect
The specifics vary by field, but the same categories come up in nearly every review I've written or read.
Who or what you studied, and how you got them
Give the setting, dates, eligibility criteria, and recruitment route. Explain how you arrived at the sample size. If it was a convenience sample, say so plainly. Reviewers forgive honest limitations much faster than ones they had to dig out. For lab work, give the strain, cell line source and authentication, sex and age of animals, and housing conditions.
Materials with identifiers, not just names
Use catalog numbers, lot numbers where they matter, clone IDs for antibodies, and exact software versions. Research Resource Identifiers (RRIDs) are worth adopting even when the journal doesn't require them. A body of work on RRIDs and resource reporting shows how often reagents in published papers can't be identified unambiguously.
Procedures in the order they happened
Describe the sequence, timings, and the quantities that someone would need to copy. Report deviations from your protocol or registration, and give the reason for each. A deviation you disclose is a footnote. One a reviewer finds by comparing your paper with your trial registration is a credibility problem.
Outcomes defined before you report them
Every outcome in the results should have a definition in the methods: how it was measured, when, by whom, and whether the assessors were blinded. If you have a primary outcome, name it once and clearly.
What you excluded, and why
Explain how you handled outliers, failed runs, dropouts and missing data. Unexplained shrinkage between "we enrolled 212" and "212 minus something" in Table 2 is one of the fastest ways to lose a reviewer's trust.
Here's what the difference looks like in practice:
Before: Samples were processed using standard procedures and analyzed by ELISA according to the manufacturer's instructions.
After: Plasma was separated within 30 minutes of collection by centrifugation at 1,500 × g for 10 minutes at 4°C, aliquoted, and stored at −80°C. No sample underwent more than one freeze-thaw cycle. IL-6 was measured in duplicate using a commercial ELISA kit (manufacturer, catalog number, lot number) following the manufacturer's protocol, with samples diluted 1:2. Duplicates with a coefficient of variation above 15% were re-run; two samples failed re-run and were excluded.
The second version is longer. It's also unarguable, and that's the point.
Describing statistics so a reviewer can check them
The statistics paragraph is where I see the most damage done in the fewest words. "Data were analyzed using appropriate statistical tests" is not a description. It's a request for the reviewer to trust you, and reviewers don't extend that kind of trust.
The clearest example I've seen came from Priya, an intensive care physician who was writing her first first-author paper while working full clinical weeks. Her draft said, "t-tests and chi-square tests were used where appropriate; p < 0.05 was considered significant." When I asked her to write out exactly what was compared with what, she realized that many of her observations came from the same patients measured on several days. She had been treating them as independent. Once she spelled out the analysis, the error was obvious. A reviewer would have caught it too, and the paper would likely have gone back with a reject-and-resubmit at best. She refit the data with a mixed-effects model. One of her headline results got weaker. The paper got much stronger.
Writing the statistics out in full isn't only for the reviewer. It's how you catch your own mistakes.
What a clear statistics paragraph includes
The unit of analysis (patient, sample, cage, hospital) and how clustering or repeated measures were handled
The model or test for each outcome, named specifically (for example, "linear mixed-effects model with a random intercept for patient," not "regression")
Covariates, and whether they were chosen in advance or selected from the data
How assumptions were checked, and what you did when they weren't met
Missing data: how much there was, and how you handled it
How you handled multiple comparisons, or a clear statement that you didn't adjust and why
Which analyses were pre-specified and which were exploratory
Software and version, including packages for anything non-standard
What you report: effect sizes with confidence intervals, not p-values alone
A sample sentence I'd be happy to see in any manuscript:
Because each patient contributed measurements on up to five ICU days, we analyzed lactate trajectories using a linear mixed-effects model with fixed effects for day, treatment group and their interaction, and a random intercept for patient. Age and baseline SOFA score were included as covariates, as specified in the statistical analysis plan before data unblinding. We report estimated differences with 95% confidence intervals. Analyses of secondary outcomes were not adjusted for multiplicity and should be interpreted as exploratory. Analyses were performed in R version 4.x using the lme4 package.
(In your own paper, give the exact version number. I've left it general here only because yours will differ.)
If you can't write that paragraph for your study, that tells you something, and it's better to learn it now than from Reviewer 2. Concerns about how statistical methods are reported in published papers have been documented for decades. Editors know this literature, and many of us read statistics paragraphs with it in mind.
Reporting guidelines: use them while drafting, not after
The EQUATOR Network maintains the reporting guidelines most journals point to. These include CONSORT for randomized trials, STROBE for observational studies, ARRIVE for animal research, STARD for diagnostic accuracy, and PRISMA for systematic reviews and meta-analyses. If your study type has a guideline, the journal will very likely ask for the completed checklist at submission.
My honest opinion is that most authors use these checklists badly. They write the paper, then fill in the checklist the night before submission, citing page numbers for items they never really addressed. Editors can tell. A checklist item that points to a page where the item isn't covered is worse than leaving it blank.
Use the checklist as an outline instead. Open it before you write a word of the methods and work through it item by item. It will feel mechanical, and I'll admit the guidelines have a real downside: they can push writing toward box-ticking, and they don't fit every hybrid design neatly. Even so, the alternative is relying on memory, and memory is how Tomas and I ended up with "brief centrifugation."
For clinical trials, make sure your methods match your registration. The ICMJE recommendations require prospective registration for trials in member journals, and reviewers increasingly open the ClinicalTrials.gov record next to your manuscript. Any differences in outcomes, sample size or analysis plan need to be acknowledged in the paper, not left for someone else to find.
Word limits, supplements and "as previously described"
Word limits are real, and I won't pretend otherwise. The trade-off I recommend is to keep every decision in the main text and move the operational detail to the supplement.
Decisions are things like eligibility criteria, the primary outcome, the analysis model and how you handled missing data. These shape how a reader interprets your results, so they belong in the paper. Operational detail covers buffer recipes, full instrument settings, complete code and every sensitivity analysis. That can go in a supplement or a public repository, provided you point to it from the main text.
"As previously described" is acceptable when three things are true: the cited paper is accessible, it actually contains the detail, and you did exactly what it says. If you changed anything, even slightly, write "as previously described (ref), with the following modifications:" and list them. I've followed citation chains three papers deep looking for a protocol that, it turned out, nobody ever wrote down in full.
The mistakes I see most often
Writing for insiders: lab shorthand, unexplained abbreviations, "standard conditions"
Outcomes that appear in the results but were never defined in the methods
Sample sizes that change between the methods, tables and figures with no explanation
"Where appropriate" doing the work that a named statistical test should do
Ignoring clustering, so repeated measures or multiple samples per animal are treated as independent
Presenting exploratory analyses as if they were planned
Software named without a version, and code described as "available on request"
A reporting checklist completed after the fact that points to the wrong pages
A pre-submission check you can run in an afternoon
Give your methods section, and only that section, to a colleague outside your group. Ask them to write down every question they'd need answered to repeat the study. Each question is a gap.
Open the relevant EQUATOR checklist and confirm that every item is genuinely addressed on the page you cite.
Trace every number in your results back to a sentence in your methods that explains where it came from.
Compare your methods with your protocol or registration, and disclose every difference along with the reason for it.
Read the statistics paragraph and check that it names the unit of analysis, the model, the covariates, how missing data were handled, and the software version.
Check that the data and code statements say where the material actually lives, not just that it exists.
Write the methods while you still remember
The best methods sections I've handled weren't written at the end. They were assembled from records kept along the way: protocol notes, version histories, the paper that justified a particular cutoff, and the email where someone explained why sample 14 was excluded. The worst ones were reconstructed from memory months later.
This is one place where your tools matter. I've watched researchers keep their literature, notes and drafts in an AI-native research workspace like PACR, so the source for a methodological choice sits alongside the draft paragraph that describes it. When a reviewer asks why you used a particular threshold, you can find the paper in a few seconds instead of an afternoon. If you're starting a new project, it's worth setting up that workspace before the first experiment rather than the week before submission.
Whatever tools you use, do one thing this week. Pick the methods section of your current draft, hand it to someone who wasn't involved in the study, and ask them whether they could do what you did. Their answer will tell you more than any reviewer's.
