The division of labor between text, tables, and figures
Tables and figures exist to carry data the prose cannot hold. A table gives exact values a reader might want to compare or reuse; a figure makes a relationship visible at a glance. The text does neither. Its job is to tell the reader which pattern in that display matters and where to look for it.
The failure mode is duplication. When a paragraph recites every cell of a table, the reader has to process the same information twice and the section doubles in length for no gain. Weak: “Group A scored 4.2, Group B scored 3.8, and Group C scored 3.1 (see Table 1).” Stronger: “Scores declined steadily across the three conditions (Table 1), with the largest drop between B and C.”
A useful rule is that any number in the running text should be one a reader needs while reading that sentence. Two or three numbers per paragraph is usually the ceiling. Everything else lives in the display, where it can be scanned rather than read.
Ordering results by question, not by analysis
Many drafts present results in the order the analyses were run, which typically means preliminary checks, then whatever was tried first, then the additions made after the first pass. That order encodes the history of the project. A reader is trying to follow the argument, and needs the findings arranged to match the questions the introduction promised to answer.
Subheadings that echo your research questions solve this almost by themselves. They make the coverage auditable — a reader can check that every question has a corresponding result — and they force you to notice when an analysis has no question behind it. Analyses that were exploratory should be labeled as such rather than presented alongside confirmatory ones.
Within a section, move from the general to the specific: descriptive statistics and sample characteristics, then the main test, then any subgroup or sensitivity analysis. Keep manipulation checks and assumption tests short. They establish that the analysis was legitimate; they are not findings in their own right.
Reporting numbers so a reader can check them
A complete report includes the descriptive picture — means or medians with a measure of spread, or counts and percentages — before any inferential test. A p-value without the underlying numbers is close to uninterpretable, because it conflates the size of an effect with the size of the sample. Give the estimate, its precision, and the test, in that order.
Effect sizes and confidence intervals are now expected in most fields, and they are what make a result usable in later synthesis. Report them for the main comparisons at minimum. Where an analysis produced many coefficients, the full model goes in a table and the text names the two or three that carry the argument.
Style guides are specific about formatting here, and reviewers do notice. Statistical symbols are usually italicized, exact p-values are preferred to threshold statements below a certain point, and decimal conventions differ by measure. Set these once at the start of drafting rather than fixing them under deadline.
- Descriptives before inferential tests — the reader needs the raw picture first.
- Estimate, interval, then test statistic and exact p-value.
- An effect size for every primary comparison.
- Exact n for every analysis, including where missing data reduced it.
Reporting the results that did not cooperate
Null and unexpected findings belong in the results, reported with the same completeness as the ones that worked. Omitting them makes the paper look tidier and makes the literature worse, and a reader who compares your results with your stated analysis plan will spot the missing comparison. Preregistration makes this visible immediately.
Report a null result as an estimate with an interval rather than as an absence: a wide interval around zero and a tight one mean very different things, and “no significant difference was found” hides which you have. The language stays neutral here — the question of whether the effect is truly absent or merely undetected is a discussion question. Report the same statistics you would have reported if the comparison had come out the other way.
Unexpected findings need the same restraint. Report the result plainly, mark clearly whether the analysis was planned or exploratory, and resist adding the explanation that has occurred to you. Save it; it will be one of the better paragraphs in your discussion.
Where the results end and the discussion begins
The boundary is not about tone but about verifiability. A statement belongs in the results if someone with your dataset could confirm it. As soon as a sentence requires a judgment about why something happened or what it means for the field, it has crossed into the discussion, even if it sounds modest.
Some journals and many qualitative traditions combine the two sections, and in mixed-methods or case-study work the combination is often clearer. Even then the distinction is worth maintaining inside each paragraph: evidence first, then interpretation, so a reader can always tell which is which. Combining sections is a formatting decision, not permission to blur the two.
Qualitative results have their own version of the same discipline. Quotations are the data, so give them enough context to be interpretable — participant identifier, role, and the question they were answering — and let the surrounding text state the pattern across cases rather than paraphrasing each quote. A quotation that is immediately restated in your own words is doing no work.
