A reviewer who cannot tell how you produced your results will not trust your results. That judgment often lands in the methods section, not because the analysis was wrong, but because the description was incomplete enough that rerunning the study would require guessing.
Methods writing is not a stylistic exercise. It is a contract: these are the steps, these are the materials, this is how decisions were made, and this is what would change if someone else followed the same instructions. The sections below are the ones reviewers map to that contract first.
What is the methods section actually for?
It exists so an informed reader can evaluate whether your design supports your claims and, in principle, reproduce the analysis.
That second word matters. Full reproduction is rare in many fields, but partial reproduction is routine: reanalyze your data with your code, rerun a simulation with your seeds, repeat an intervention with your protocol. Reviewers ask whether those partial reruns are possible from the text alone.
A methods section that only narrates what you did in the past tense (“we collected data,” “we performed analysis”) without naming instruments, versions, inclusion rules, and stopping criteria forces the reviewer to infer. Inference is where trust erodes.
What should every methods section include?
There is no universal template across disciplines, but the same structural moves appear in empirical papers whether the unit is a cell line, a survey panel, or a language model benchmark.
Cover these moves somewhere in Methods (or in linked supplements if the journal allows):
- Study design and setting. What was observed, manipulated, or compared, and over what period or in what environment.
- Participants, units, or corpus. Who or what was included and excluded, with counts at each filtering step.
- Materials and instruments. Questionnaires, assays, devices, datasets, prompts, and software, with versions and citations to established protocols where they exist.
- Procedure. The order of operations a lab member would follow on day one of the study.
- Outcomes and measures. What was recorded, how it was operationalized, and any validation you rely on.
- Analysis plan. Statistical models, correction rules, exclusion criteria applied after data collection, and how uncertainty is reported.
If any of those moves is missing, a reviewer will ask for it. If two are vague, the paper often gets a methods comment even when the headline result looks strong.
How do you test whether a stranger could rerun your study?
Print the methods section and highlight every noun that names something concrete: a kit, a survey item set, a random seed, a URL, a hyperparameter, an IRB protocol number. Circle every verb that describes an action someone would take: “centrifuged,” “imputed,” “fine-tuned,” “double-coded.”
For each circled verb, ask: could a competent colleague in my subfield perform that action tomorrow with only what appears in the manuscript and its supplements?
Run three spot checks that catch most gaps:
The reagent check. Pick one material central to the result (antibody lot, model checkpoint, survey wave, corpus snapshot). Search the PDF for its identifier. If you find a brand name but not a version, a date, or a DOI, add one.
The fork check. Find every place you made a discretionary choice: bin width, window length, train/validation split, outlier rule, prompt template. Either justify it with a citation to a standard practice or report that you tried reasonable alternatives and the conclusion held. Unlabeled forks read as fishing.
The code check. If analysis code exists, state where it lives (repository, archive DOI, “available on request” only if the journal policy allows it and you will honor it). Name the language, key packages, and versions that affect numeric output. Reviewers increasingly expect that a table can be regenerated, not merely described.
How should you structure the methods section?
Order follows the reader’s mental model of the study, not the order you happened to do the work.
In experimental psychology and many biomedical journals, the CONSORT flow diagram has trained reviewers to expect participant counts at enrollment, allocation, follow-up, and analysis. You do not need the diagram in every field, but you do need the numbers it would contain if someone drew it. A sentence like “412 enrolled, 387 completed follow-up, 12 excluded for protocol violation before analysis” answers the first question a statistical reviewer asks.
In computational papers, mirror that transparency with a data pipeline sentence: raw corpus size, filters applied, final n used in training and evaluation, and whether splits are document-level or instance-level. Reviewers in machine learning have learned to distrust benchmarks where train and test leakage is possible from how the split is described.
A structure that survives review in most empirical fields:
- Design overview (one short paragraph: observational, experimental, simulation, mixed).
- Sampling and recruitment (or data source acquisition).
- Variables and measurement (instruments before procedure if measurement defines the protocol).
- Procedure (chronological steps).
- Analysis (models, assumptions, software).
Qualitative and computational papers often swap 3 and 4: define the corpus or interview guide before you describe how it was applied. The test is whether each subsection can be read alone without forward references like “as described above” that point to undefined terms.
Use subheadings that name content, not process labels. “Statistical analysis” is fine. “Data” is not fine when it hides three different datasets. Reviewers skim headings; make them informative.
What do reviewers flag first in a methods section?
Desk reviewers and statistical reviewers converge on a short list.
Under-specified exclusion. “Participants with incomplete data were excluded” without stating how many, which fields triggered exclusion, and whether exclusion was decided before or after seeing outcomes. Pre-registration or a dated analysis script closes this.
Pseudo-replication. Treating technical replicates, repeated measures from the same subject, or multiple tweets from one account as independent observations without a model that accounts for nesting. The fix is explicit clustering, mixed models, or a clear unit of analysis.
HARKing signatures. Outcomes or subgroups that appear in Results without a prior definition in Methods or registration. If you added an exploratory analysis, label it exploratory in Methods, not only in the Discussion.
Instrument drift. A scale, assay, or benchmark version changed mid-study without a stratified analysis or sensitivity check. Name versions and dates.
Irreproducible AI workflows. “We used GPT-4” without model snapshot, temperature, prompt text, and whether outputs were filtered or manually edited. For model-assisted studies, treat prompts and seeds like laboratory protocols.
Power and sample size without a plan. Reporting a post hoc power calculation after a null result, or stating that “the sample was determined by availability” when the primary endpoint is binary and rare. Reviewers in clinical and social science fields expect either a prespecified target or an explicit precision argument. If you could not recruit to target, say so and discuss how wide the confidence interval is.
None of these require more pages if you front-load identifiers and move bulky detail to supplements with inline pointers (“full survey text, Supplementary Methods S2”).
How long should the methods section be?
Length follows complexity, not importance. A well-powered clinical trial with prespecified endpoints needs more space than a secondary analysis of a public dataset with a published codebook.
A useful ceiling for many journal formats: if Methods exceeds roughly 40% of the main text, move instrument manuals, ethics approvals, and extended protocols to supplements and keep the main section to decisions that affect interpretation.
Reviewers rarely complain that Methods is too long when every paragraph answers a reproducibility question. They complain when length is padding: restating the introduction’s rationale or listing software without saying what each package did.
When two methods papers in your target journal both cite the same core protocol, cite that protocol and describe only your deviations. “We followed Smith et al. (2019) with the following changes:” is both shorter and more auditable than rewriting a standard pipeline in your own words.
What belongs in supplements instead of the main methods?
Move material that a reviewer needs once but should not wade through on first read:
- Full survey or interview guides
- Detailed inclusion/exclusion flow diagrams with counts
- Extended parameter tables for models or simulations
- Ethics approvals and consent language (summarize in main text; attach full forms if required)
- Additional robustness checks that do not change the primary conclusion
Keep in the main Methods anything that changes how a reader weights the primary result: stopping rules, primary endpoint definition, handling of missing data, and any deviation from a preregistered plan.
How do you revise methods after peer review?
Methods revisions are usually additive. Reviewers ask for clarity, not a new study.
When a comment says “please clarify,” respond with a specific insertion: a sentence with a number, a version string, or a citation. When a comment asks for a new analysis, add the analysis to Results and add the decision rule that defines it to Methods, even if the rule was implicit before.
Avoid silent fixes. If you correct an error in the protocol description, say so in the response letter. Editors compare versions; undisclosed changes erode trust faster than the original omission.
Some dedicated academic tools, Fukuro among them, will flag inconsistencies between your methods claims and your results section (sample sizes that do not match, outcomes mentioned without definition). That catches internal contradictions; it does not replace a colleague reading for reproducibility gaps.
What should you do before you submit?
Run this ordered pass on the final PDF:
- Read Methods without reading Results. Write down the primary outcome and design in your own words. If you cannot, a reviewer cannot either.
- Run the reagent, fork, and code checks on three items each.
- Match every figure and table caption that cites a method (error bars, test name, n) to a sentence in Methods.
- Confirm ethics, data availability, and code availability statements match what Methods promises.
- Ask a co-author who did not write Methods to perform one partial rerun or to list the first three questions they would need answered to try.
The methods section is the part of the paper most likely to generate revision requests that are cheaper to fix before submission than after three months in review. Treat it as the specification, not the appendix.