How to write a methodology (and what gets it sent back)
Generic methodology advice gives you a list of components and pretends a methods section is the same document in every field. It is not. A social science methods section is judged on sampling and measurement, an engineering one on apparatus and uncertainty, and a computational one on data splits, baselines and whether anyone else could run your code. This page is organised the other way round from the usual guide: first the seven things that actually get a methods section sent back, then what your specific field expects, then how much detail is enough.
- The one thing the section has to do
- The seven things that get a methods section sent back
- Three fields, three different documents
- What each field checks, side by side
- How much detail is enough
- Qualitative work has its own bar
- The methods section and the methodology chapter
- Citing the instruments you built on
One rule sits above everything below: your department's handbook and your supervisor's template beat any general advice, including ours. Where a programme hands out a structure, the marking rubric is usually built from it line by line.
The one thing the section has to do
A methodology answers one question for the reader: why should I believe your results? Everything in it serves that, and the section fails when a reader finishes it still able to ask "but how does that follow?"
Two audiences read it and they want different things. A journal reviewer wants enough detail to repeat the work. A marker wants evidence that you chose a design deliberately and understand what it can and cannot show. The overlap is large but not total, which is why the advice everyone repeats, include enough detail to replicate, is right for submission and incomplete for coursework.
Our view, and it is a minority one: for coursework, "enough detail to replicate" is over-quoted and under-useful, because your marker is never going to replicate anything. What they will do is read your design against your research question and decide, in about a minute, whether one can answer the other. So write the fit sentence first, before the detail, in the shape "we used X because the question asks about Y, and X is what produces evidence about Y". Detail is necessary. Fit is what gets marked.
The seven things that get a methods section sent back
Roughly in order of how often each one is the actual problem.
The method cannot answer the question
A qualitative case study used to establish how widespread something is. A survey used to explain why people experience something the way they do. A model evaluated on the data it was fitted to. This is the failure that adding detail cannot fix, because the section is describing a study that answers a different question from the one the introduction asked. The repair is upstream: either change the question to the one your method answers, or state plainly what your design can establish and scope every later claim to it.
Nobody could repeat it
Naming the design but not the sample size, the instrument, the settings or the steps. The clearest statement of the bar comes from PLOS ONE's own submission rules: the methods "should provide enough detail to allow suitably skilled investigators to fully replicate your study" (PLOS ONE submission guidelines). The test is mechanical. Hand the section to someone working on a different topic and ask them to tell you what you did. Every place they guess is a gap, and the fixes are usually one sentence each.
The sample has no justification
Not a small sample, an unexplained one. APA's Journal Article Reporting Standards for quantitative work ask authors to report the intended sample size, the achieved size where it differs, and how the target was determined, along with inclusion and exclusion criteria and how participants were recruited (APA Style JARS, quantitative). A student sample of 42 with a sentence explaining why 42 reads as a decision. The same 42 with no sentence reads as whoever happened to answer.
The analysis does not fit the data
Rich interview material reported as percentages. A survey analysed with no named test. A threshold chosen after seeing the results. Say what you did to the raw data and why that operation suits both the data you collected and the question you asked, in that order.
There is no baseline
A number with nothing to compare it against is not a result. This is fatal in engineering and computational work and common in student projects everywhere: an accuracy figure with no baseline model, a measured value with no control condition, a system evaluated against nothing. If your methods section names no comparison, your results section has nothing to report.
It is written as a plan
The future tense gives it away: "participants will be recruited". That is a proposal, not a methodology. Report completed work in the past tense. The exception is a genuine proposal document, which has its own conventions, covered in how to write a research proposal.
The instruments cannot be traced
A methodology cites more heavily than any section except the literature review, because it names the scales, procedures and standards it stands on. Those citations have to resolve. This is where a section drafted with AI help quietly falls apart, and the last section of this page has the numbers.
Three fields, three different documents
The components everyone lists, design, participants, materials, procedure, analysis, ethics, limitations, are a skeleton. What actually gets checked differs so much by field that following generic advice faithfully is itself a way to lose marks.
Social science and psychology
The section lives or dies on sampling and measurement. Report who was studied and their relevant characteristics, how they were recruited, how the target sample size was determined, the inclusion and exclusion criteria, and every measure by name, citing the published instrument rather than describing it. Where you built your own, say how, and put it in an appendix. Reliability and validity evidence belongs here, not in the results. APA's JARS tables are the closest thing to a definitive list, and reading the one for your design takes ten minutes and tells you exactly what a reader in the field expects to find.
Ethics is a real subsection here, not a formality: the approving body and reference, informed consent, anonymity, and how the data is stored. Even a small undergraduate survey usually needs a line on consent and anonymity.
Engineering and experimental science
Here the section is about apparatus, calibration and uncertainty, and some journals set numerical requirements that no amount of good prose substitutes for. The ASME Journal of Fluids Engineering states in its editorial policy that it "will not accept for publication any paper reporting the numerical solution of a fluids engineering problem that fails to address the task of systematic truncation error testing and accuracy estimation", and goes on to require that methods be at least second order accurate in space and that grid independence be established (ASME editorial policy on the control of numerical accuracy). Whether or not you are writing for that journal, the shape of the demand generalises: name the instrument and its precision, say how it was calibrated, state the uncertainty on every quantity you report, and show that your result is not an artefact of the resolution you chose.
What replaces "participants" here is the specimen, the test rig or the simulation domain, described in the same detail: dimensions, materials, boundary conditions, operating range.
Computational work and machine learning
The unit of trust is the artifact, not the prose. ACM's artifact policy separates three things students routinely conflate: repeatability, the same team repeating its own measurement; reproducibility, a different team getting the same result using the original artifacts; and replicability, a different team getting it with artifacts they built independently (ACM Artifact Review and Badging). Its badges reward making artifacts available, functional and reusable, which tells you what this field considers a complete methods section: one that points at code and data.
Conference practice has hardened that into a form. NeurIPS requires every submission to complete a paper checklist covering, among other things, whether the code, data and instructions needed to reproduce the main experimental results are provided, with a section reference for each answer (NeurIPS paper checklist guidelines). Write your methods section so it could answer that checklist and you will not leave out what markers in this field notice: the data source and how it was split, the baselines you compared against, the hyperparameters and how they were chosen, the random seeds, the hardware and the runtime.
What each field checks, side by side
| Generic component | Social science | Engineering or experimental | Computational or machine learning |
|---|---|---|---|
| What you studied | Participants, recruitment, criteria | Specimen, rig, simulation domain | Dataset, provenance, licence |
| How much of it | Sample size and how it was set | Number of runs and repeats | Train, validation and test split sizes |
| Instruments | Validated scales, cited | Apparatus, precision, calibration | Model, library, version |
| Settings | Conditions, order, timing | Boundary conditions, operating range | Hyperparameters, seeds, hardware |
| Comparison | Control or comparison group | Reference case or analytical solution | Baselines and ablations |
| Error | Reliability, validity, missing data | Measurement and discretisation uncertainty | Variance across seeds, intervals |
| Openness | Ethics approval, consent, storage | Data availability, standards followed | Code and data availability |
Read down your own column. Any row that is empty in your draft, and should not be, is the sentence to write next.
How much detail is enough
Two rules settle almost every "should I include this?" question.
Include what a reader needs to repeat the work or to judge it. Sample size, exact measure, analysis run, number of coders, model version: all needed. The chronology of your fieldwork, the software that worked fine, the steps anyone would take anyway: not needed.
Cite instead of re-describing. When you use a published instrument, a standard procedure or a documented dataset, name it and cite it. This is the biggest single source of unnecessary length in student methodologies, and citing is also what lets a reader reach the detail you are leaving out.
On length, one measured number. In 58 published papers we counted for a companion piece on paper structure and section proportions, methods was the longest section, a median of 26 percent of the body text, and it ran longer than the results section in 33 of them. If your methods section is the shortest thing in your empirical paper, that is worth a second look. The caveat is that those papers appeared where no word limit applies; a journal capping you at 3,000 words compresses methods first, usually into supplementary material.
Qualitative work has its own bar
A qualitative methodology is not a quantitative one with the numbers removed, and it is marked on different things. It has to say how participants or cases were selected and why those, since the sample is chosen for what it can reveal rather than for representativeness. It has to describe the analysis concretely: which approach, who coded, how codes were developed, how disagreements were resolved, and what made you stop. And it has to address the researcher's own position, because in qualitative work the researcher is part of the instrument and the reader needs to know who was in the room.
Mixed-methods work carries one extra job that is easy to skip: saying how the two strands connect. Whether they ran in parallel or in sequence, and which one informed the other. Two methods reported side by side with no integration is two studies in one document, and readers say so.
The methods section and the methodology chapter
Same content, different burden of proof. A methods section in a paper reports. A methodology chapter in a thesis, usually chapter three, reports and defends: why this design and not the obvious alternative, why this sample size, what the approach cannot establish. The defending is the part students underwrite, and it is most of the difference in length. A good test for a chapter: for every major choice, is there a sentence naming the alternative you did not take and why? If not, you have written a long methods section.
Both are past tense, and both sit between the literature and the results. Where the whole document sits, and what changes at masters level, is in our thesis guide. If the study is still being designed, how to write a research question is the step before this one, and it is the step that decides whether the first failure above can happen to you at all.
Citing the instruments you built on
Every scale, procedure and standard your methodology stands on has to resolve to something a reader can open. This is where drafting with a general AI tool does specific damage: models produce references that look flawless, with believable authors and a correctly formatted DOI, for papers that were never written. A peer-reviewed audit of ChatGPT's bibliographies found 55 percent of GPT-3.5 citations and 18 percent of GPT-4 citations were fabricated, and that among the citations pointing at real papers, 43 percent and 24 percent respectively still carried substantive errors (Walters and Wilder, Scientific Reports, 2023, doi.org/10.1038/s41598-023-41032-5). A fabricated citation is worse here than anywhere else in the paper, because it sits inside your claim of rigour.
CiteOwl is built for that seam. It searches real literature, reads what it finds, and shows the quote behind each citation, so a scale or procedure your methods section leans on is a paper you can open. It will draft and structure the section from your account of what you did, with every change arriving as a diff you accept or reject, and it will not invent the method: the design, the sample and the analysis are yours to choose and defend. To test references already sitting in a draft, our free citation checker resolves them against CrossRef with no account.