The data analysis section is the one written least carefully and the one that weighs most when a reviewer wants to know whether your results can be trusted. It usually shrinks to one sentence (“data were analysed with SPSS 28”) and a list of tests, and with that the reader cannot tell which model you fitted, what you did with cases that did not fit or how you decided a result counted.
A good analysis section has a specific function: to give someone with the same tools the information to repeat what you did, and the reviewer the information to judge whether you did it well. It does not teach statistics or describe each test as a textbook would. It reports decisions.
This guide tells you which decisions to report, in what order, with a model paragraph and a list of what most often goes missing.
In many papers it sits inside the method, at the end, headed “Data analysis” or “Statistical analysis”. In others, especially theses, it is a chapter of its own. Its content is the same: how you went from the data collected to the results you report.
It helps to tell it apart from the results section. The analysis tells what you did and why; the results tell what you found. Mixing them is a frequent error, and it shows when the analysis section starts giving figures or when the results section has to justify decisions that belonged earlier.
An order that works follows the path of the data: from preparation to the final model. Each point can be a sentence or a paragraph, depending on how complex the study is.
“Analyses were run in R 4.3 (packages lme4 and emmeans). Participants who did not complete at least 80% of the anxiety scale were excluded (n = 6; criterion set before analysis). Remaining missing data (3.1% of responses, with no pattern related to group according to Little’s test) were handled with multiple imputation (20 datasets). To test Hypothesis 1 we fitted a linear model with the post-treatment score as dependent variable, group as predictor and the pre-treatment score as covariate. Normality of residuals and homogeneity of variance were checked, and robust standard errors are reported. The significance level was .05 two-tailed; effect size (Cohen’s d) and its 95% confidence interval are reported. Analyses by academic year were exploratory and not corrected for multiple comparisons.”
It is a single paragraph and it holds the information a reviewer needs: prior criteria, handling of missing data, the model specified, assumptions, decision criterion and what was exploratory. If any of those elements is missing, you are likely to get a question about it.
The counterpart of reporting decisions is not reporting what is not one. There is no need to explain what a Student t is or how an ANOVA is computed: the reader is assumed to know. Nor do you list every test that exists to justify the one you chose, or include results.
What does deserve a sentence is anything that departs from the usual: an uncommon test, a model that is not the standard one for that design, a transformation, a correction. There the reviewer will want to know why, and it is better that the answer is already in the text.
These are the gaps that most often provoke a reviewer comment:
A quick test: ask someone who knows the analysis, but not your study, to read only this section and tell you which model you fitted for each hypothesis. If they cannot, something is missing. Another: check that each analysis appearing in the results is announced here, and that each one announced here is reported there. An analysis that appears in one place and not the other is a certain comment.
For reporting the figures themselves, see our thesis data analysis guides in the blog and the guide to the statistical errors a reviewer spots first.
Paste your data analysis section, or the whole method, and you will see the first objection a reviewer would raise, with the exact sentence from your text and what it would take to fix it. It looks at what this guide describes: whether each analysis answers an aim, whether the sample size is justified, whether you say how missing data and assumptions were handled, whether multiple comparisons are corrected, and whether the procedure could be replicated from what you wrote.
It returns one objection, the one that weighs most, so you know what to fix first. It needs at least 350 characters and the section does not need to be finished. What you paste is not stored; if deciding something would require seeing the results, it says so separately.
It depends on the journal and the type of work. In papers it is usually the last part of the method; in theses, often a chapter of its own. Check the author instructions and the journal’s recent articles.
As long as it takes for the analysis to be reproducible and judged: a paragraph for a simple design, more for complex models. Do not try to fill space by explaining well-known tests; anything that departs from the usual does take space.
Standard ones, no. Less common ones, or those implemented in a specific package, are worth citing with their reference and the software version. It lets the reviewer check what you did.
Yes, always, along with the main packages if you use R or Python. The results of some models change between versions, and it is the minimum information for the analysis to be reproduced.
In this section, with an explicit sentence, and again in the results where they are reported. An exploratory analysis declared as such is legitimate; presented as confirmatory it is not.
Not always, but many journals ask for it or value it, and it is the best answer to any doubt about the analysis. If you cannot share the data, share at least the code or an exact description of the steps.