The discussion section of a research paper explains what your results mean. It opens with the answer to your research question, interprets the main findings, compares them with earlier studies, weighs other explanations and the study's limitations, and ends with implications and next steps. It interprets the results; it does not repeat them or add new ones.
Below: the six moves in the order journals ask for them, a drafting method, a complete discussion annotated paragraph by paragraph, two shorter samples and the common mistakes. Every study, number and citation in the samples is invented to show the form. The closing section of a paper has its own guide: how to write a conclusion for a research paper.
What the discussion section does
In an IMRaD paper the results say what you found and the discussion says what it means. The introduction narrowed from the field to one question; the discussion starts from the answer and widens back out to the field. The quickest way to see the difference is one finding written for each section.
Results: Collisions at the four treated buildings fell from 112 in autumn 2024 to 41 in autumn 2025, while collisions at the four untreated buildings fell from 58 to 51 (Table 2).
Discussion: Marking the glass was followed by a drop in collisions five times larger than the drop at untreated buildings, which suggests that birds on this campus hit the windows because they could not see the glass as a barrier. Part of the difference may be regression to the mean: the university marked its four worst buildings, and unusually high counts tend to fall on their own.
The results sentence is past tense, numerical and silent about why. The discussion interprets: it makes a hedged claim, gives a reason and names the strongest reason to doubt it. APA Style's tense guidance follows the same split: past tense for results, present tense for implications, limitations and future directions.
| Results | Discussion | |
|---|---|---|
| Answers | ResultsWhat did you find? | DiscussionWhat does it mean, and how far does it reach? |
| Contains | ResultsNumbers, tests, tables and figures | DiscussionInterpretation, comparison with other studies, limitations, implications |
| Citations | ResultsFew or none | DiscussionThe studies you compare your findings with |
| Keep out | ResultsExplanations and “because” | DiscussionNew data and a second run through the numbers |
Some journals let the two sections merge. PLOS ONE allows Results, Discussion and Conclusions to be separate or combined and lists Conclusions as optional, so check the author guidelines before you plan the headings.
The six moves, in order
Medicine and psychology have published detailed guidance on the discussion, and the lists agree closely. The ICMJE recommendations suggest beginning with a brief summary of the main findings, then exploring possible explanations, comparing the results with other relevant evidence, stating the limitations and exploring the implications. The BMJ prescribes a five-part order, and APA's reporting standards for quantitative research (JARS–Quant) list the same elements. Combined, they give six moves:
| Move | What it does | Where guidelines ask for it |
|---|---|---|
| Move1. Principal finding | What it doesAnswers the research question in a sentence or two | Where guidelines ask for itICMJE: main findings first. BMJ: principal findings. JARS: support or nonsupport for each hypothesis |
| Move2. Interpretation | What it doesExplains the finding: the mechanism or reason | Where guidelines ask for itICMJE: mechanisms or explanations. BMJ: meaning of the study |
| Move3. Comparison | What it doesSets the finding against earlier studies and explains differences | Where guidelines ask for itICMJE: relevant evidence. BMJ: other studies. JARS: similarities and differences |
| Move4. Alternatives and surprises | What it doesWeighs other explanations and deals with unexpected results | Where guidelines ask for itJARS: sources of bias, threats to validity, imprecise measures |
| Move5. Strengths and limitations | What it doesSays how far the finding can be trusted and how far it reaches | Where guidelines ask for itICMJE: limitations. BMJ: strengths and weaknesses. JARS: generalizability |
| Move6. Implications and next steps | What it doesSays who should do what with the finding, and what study comes next | Where guidelines ask for itAll three: implications for practice, policy and research |
The order works in any empirical field, and two details are worth borrowing even outside medicine. The BMJ caps the discussion at five paragraphs, which forces a choice about what matters. The ICMJE adds cautions that read like a reviewer's checklist: distinguish clinical from statistical significance, avoid claiming priority, label new hypotheses clearly, and avoid conclusions the data do not support. If you are still planning the paper, the research paper outline template gives the discussion its own section to fill in before you draft.
How to write a discussion section
- Copy your question above the blank page
Paste the research question and hypotheses from your introduction at the top of the draft. The discussion answers that question, not a neighboring one. If it still feels vague, the research question examples show what a focused one looks like.
- Rank your findings
List three to five findings in order of importance. JARS–Quant separates primary, secondary and exploratory hypotheses; so should the discussion, with the primary answer first and exploratory results labeled as such.
- Write the answer, then the reason
State the main finding in plain words, without test statistics, and say whether it supports the hypothesis. Then explain it, and flag any explanation you did not measure as a possibility.
- Compare with the studies you reviewed
Return to the two or three closest studies from your literature review and say where you agree, where you differ and why: sample, setting, measure or design. The literature review template keeps those studies in a synthesis matrix, which makes this step quick.
- Face the unexpected and the alternatives
Address the result that went against you before a reviewer does. Then weigh the other explanations for your main finding, such as chance, a change in measurement or regression to the mean, using your own data.
- Give each limitation a consequence
For each limitation, say which way it could push the result and whether it threatens the conclusion. A limitation that biases results toward no effect makes a negative finding less convincing and leaves a positive one, if anything, conservative; readers want to know which you have.
- End on implications sized to the evidence
Name the reader who can use the finding and what they could do, at the scope one study supports, then describe the study that would remove your main limitation. If the paper has a separate conclusion, leave the final recommendation to it.
Annotated discussion section example
The paper: a university put dot-pattern markers on the glass of its four buildings with the most bird collisions and compared them with four unmarked buildings, using daily searches by the same two observers in the autumn migrations before and after. Collisions at the marked facades fell from 112 to 41, and at the unmarked buildings from 58 to 51. North-facing marked glass showed no drop.
Marking the glass of four campus buildings was followed by a 63% drop in recorded bird collisions at those facades, against a 12% drop at unmarked buildings over the same migration season. The difference suggests that most collisions at the marked buildings happened because birds could not see the glass as a barrier, and that a visible pattern was enough to change their flight. The reduction was largest on facades that reflect the campus woodland, where the glass had shown birds an image of trees and sky.
What it does. The first sentence answers the question with the one comparison that matters, marked against unmarked, and leaves the counts to Table 2. The second interprets, hedged with "suggests." The third adds a detail that supports the interpretation.
The size of the effect falls between earlier estimates. Ruiz and Halvorsen (2019) reported a 70% reduction after markers were applied to a single museum facade, and Aboagye et al. (2022) a 41% reduction across 12 office buildings. Two differences may explain the spread. Ruiz and Halvorsen marked only the most dangerous facade, where there was the most to prevent, while Aboagye et al. spaced their markers twice as far apart as ours. If spacing explains the spread, closer markers should prevent more collisions, a hypothesis no study has yet tested directly.
What it does. It places the result between two others and explains the spread with features of each design, which says more than "consistent with previous research." The idea about spacing is labeled a hypothesis, as the ICMJE asks.
Collisions on north-facing glass did not fall, although it received the same markers. These windows look onto a lit parking lot rather than trees, and the collisions there clustered in the hour before dawn, which points to birds drawn by the lights rather than misled by reflections; markers would not prevent that. Two other explanations for the overall drop are unlikely. Fewer birds may have passed through in 2025, but collisions at the unmarked buildings barely changed. Search effort could have changed, but the same observers walked the same routes on the same schedule in both years.
What it does. The surprise gets an explanation offered as likely, not proven. Then the two objections a skeptic would raise first are answered with evidence from the study itself.
The main threat to this conclusion is that the buildings were not randomly assigned. The university marked the four with the most collisions, so part of the drop may be regression to the mean, the tendency of unusually high counts to fall on their own. Carcass counts also underestimate collisions, because scavengers remove some birds and injured birds may fly off before they are found. If removal was similar in both years, it lowers the counts without changing the comparison, but a change in scavenger activity could bias it either way. Finally, the data cover one autumn on one campus and cannot show whether the effect lasts, or holds in spring, when different species pass through.
What it does. It opens with the strongest reason to doubt the result instead of a stock line about limitations, and each limitation comes with its consequence: the first could inflate the effect, the second probably leaves the comparison intact, the third limits how far it reaches.
For campus facilities staff, the findings support marking the facades with the highest collision counts first. They do not support marking north-facing glass near lit areas, where shielding or switching off lights before dawn is the more likely remedy. A randomized comparison of marker spacings across several campuses, with daily searches in both migration seasons and a trial of carcass removal rates, would show how much of the drop the markers caused and which spacing works best.
What it does. The advice names its reader and stays within what one autumn on one campus can support, including what not to do. The proposed study answers all three limitations from the previous paragraph, so the discussion ends on a design rather than on "more research is needed."
Two shorter discussion examples
A qualitative study
The translators we interviewed treated untranslated words not as gaps in their skill but as decisions about the reader: they kept a word when the nearest English one would bring the wrong associations, as with kinship terms and dishes. This complicates the view that leaving a word untranslated is a last resort (Ferrand, 2014). Because our 22 participants translated literary fiction, the pattern may not transfer to technical translation, where readers need precision more than atmosphere.
Why it works. A qualitative discussion answers with a pattern and asks where it transfers, not how far it generalizes statistically. The limit named is who the participants were, the one that matters for an interview study.
A null result
Contrary to our hypothesis, students who received worked examples with their weekly problem sets did not score higher on the final exam than students who received the problem sets alone (mean difference 1.2 points out of 100, 95% CI −2.1 to 4.5). The interval rules out a benefit larger than about four and a half points, smaller than the effects earlier classroom studies reported (Lindqvist & Oduya, 2018; Prakash, 2021). One reason may be uptake: fewer than half the students opened the examples in a typical week, so the study tested access to examples more than their use. Building the examples into the problem sets would test the effect of studying them.
Why it works. The null result comes first, not buried, and the confidence interval says what the study rules out, which tells readers more than "no significant difference." The explanation uses the study's own data. For why a non-significant result isn't evidence that there is no effect, see these null hypothesis examples.
Common discussion section mistakes
Here is the kind of draft the bird-collision discussion replaced. Each sentence shows a habit reviewers flag.
In this study, collisions decreased from 112 to 41 at treated buildings and from 58 to 51 at untreated buildings, and north-facing windows had 19 collisions in 2024 and 18 in 2025. This proves that window markers are highly effective and could save millions of birds if used everywhere. This is the first study ever to test markers on a campus. However, there were some limitations. More research is needed.
| Sentence | Problem | Fix |
|---|---|---|
| Sentence“collisions decreased from 112 to 41…” | ProblemRepeats the results; the ICMJE advises against repeating data in detail | FixState the finding once, in words, and point to the table |
| Sentence“This proves…” | ProblemOverclaims from one campus, one season and buildings chosen, not randomized | FixUse “suggests” and say what the design allows |
| Sentence“…could save millions of birds” | ProblemAn implication far beyond the data | FixKeep implications at the scope of the evidence |
| Sentence“the first study ever…” | ProblemA priority claim, which the ICMJE advises avoiding | FixSay what is new about the design instead |
| Sentence“some limitations” | ProblemNames nothing | FixName each limitation and its consequence |
| Sentence“More research is needed.” | ProblemTrue of every paper | FixDescribe the study that would settle the question |
Two more problems are harder to spot in your own draft:
- New results. An analysis that first appears in the discussion belongs in the results section, where readers expect to find it.
- Ignoring the studies that disagree. Comparing only with work that agrees with you reads as selective, and a reviewer who knows the contrary study will cite it back.
To catch new results and overclaims in a long draft, you can attach the results section to a chat with the Strokes Assistant and ask it to list every number in the discussion that doesn't appear in the results, and every sentence that claims a cause where the design shows only an association.
Stock phrases and direct claims
Stock openers announce a move without making it. The sample above makes each claim directly, and the contrast is the easiest edit to borrow:
| Move | Stock phrase | Direct version, from the sample |
|---|---|---|
| MovePrincipal finding | Stock phrase“The main finding of this study is that…” | Direct version, from the sample“Marking the glass of four campus buildings was followed by a 63% drop in recorded bird collisions at those facades…” |
| MoveComparison | Stock phrase“These results are consistent with previous research.” | Direct version, from the sample“The size of the effect falls between earlier estimates.” |
| MoveAlternatives and surprises | Stock phrase“One result was unexpected.” | Direct version, from the sample“Collisions on north-facing glass did not fall, although it received the same markers.” |
| MoveLimitations | Stock phrase“This study has several limitations.” | Direct version, from the sample“The main threat to this conclusion is that the buildings were not randomly assigned.” |
| MoveImplications | Stock phrase“More research is needed.” | Direct version, from the sample“A randomized comparison of marker spacings across several campuses… would show how much of the drop the markers caused.” |
Frequently asked questions
How long should a discussion section be?
Long enough to interpret the main findings and no longer. The BMJ asks for no more than five paragraphs, a useful discipline elsewhere too, while theses often give the discussion a full chapter. Check your journal's author guidelines or your assignment, and cut any paragraph that only repeats the results.
Do you cite sources in the discussion section?
Yes. Comparing your findings with earlier work is one of the discussion's main jobs, so cite the studies you compare with, usually the closest ones from your literature review, and explain why your results agree or differ. Avoid bringing in a body of literature the introduction never mentioned.
What tense is the discussion section written in?
Mostly the present tense. APA Style recommends the present tense for implications, conclusions, limitations and future directions, and the past tense for what you or other researchers found: participants scored higher (past), which suggests that the method works (present). Keep the same pattern in every paragraph.
Can the results and discussion be combined?
In some journals, yes. PLOS ONE, for example, lets authors present Results, Discussion and Conclusions as separate sections or combined. In a combined section, give each result first and its interpretation right after. If your journal or assignment says nothing, keep them separate, as in the standard IMRaD layout.
