Skip to content

Writing · Beginner

Write a discussion that explains the finding without overstating it

Explain what the findings mean and where the interpretation stops.

11 min read · Oct 26, 2026

Key points

  • Compare studies through methods.
  • Connect limitations to conclusions.
  • Propose a specific next test.
Contents

Explain what the findings mean and where the interpretation stops.

The Discussion explains what the findings mean, how they relate to existing knowledge and what remains uncertain. It should help readers judge the contribution, not persuade them to ignore limitations. Interpretation is strongest when it connects each claim to a design, measurement and result that can support it.

A good Discussion can make an inconclusive result useful. It may show that a common measurement is ambiguous, that an apparent effect changes across conditions, or that current precision is insufficient for a practical decision. Its value does not depend on turning every observation into a success story.

Begin with the answer to the original question

Open with the principal finding in plain language. Include the magnitude when that helps interpretation, but do not repeat the entire Results section. State the population, system or setting to which the answer directly applies.

For the supplied teaching dataset, an appropriate opening is: “Controller B had lower positioning error than controller A in all twelve constructed cases, with a mean paired reduction of 3.58 mm. This demonstrates a reporting method for paired data; it does not establish the performance of an actual navigation system.”

The second sentence is essential because the observations are synthetic. A real experiment could support an empirical claim, but only within its actual design and evaluation scope. Avoid copying the language of a strong causal study into an observational report merely because the result looks similar.

ICMJE identifies interpretation, relevant evidence and limitations as central Discussion functions in medical manuscripts. Those functions are also useful across disciplines, although their arrangement and terminology vary. S01

Separate observations, interpretations and implications

An observation reports what was measured or found. An interpretation explains what the observation suggests. An implication concerns what follows for understanding, further research or a decision. Each step usually requires additional reasoning.

Level Hypothetical example What would strengthen it
Observation Mean completion time was lower with interface B in the tested tasks Transparent measurement and analysis
Interpretation B may reduce the effort needed to locate commands Evidence about task behaviour or usability mechanisms
Implication B may be worth evaluating in a wider user group External evaluation under relevant conditions

The observation does not automatically establish the interpretation. Completion time could change because the interfaces differ in familiarity, task order or error handling. The implication should therefore identify the next evaluation rather than immediately recommend universal adoption.

A mechanism is a process through which an effect occurs. Confounding arises when other variables create or distort an apparent exposure and outcome relationship. A plausible mechanism is not established simply because it fits the observed direction of change.

Compare with earlier work through design and measurement

Do more than write that your finding “agrees with previous studies”. Specify the relevant agreement: similar magnitude, direction, condition dependence or interpretation. Compare populations, timing, instruments, outcomes, controls and analysis before deciding whether two studies address the same question.

A study of research confidence and a study of demonstrated laboratory skill may both concern mentoring, yet their outcomes are not interchangeable. A materials test at one temperature may not directly contradict another conducted at a different temperature. An engineering benchmark may favour different methods when failure costs change.

Do not infer that studies disagree merely because one has p < 0.05 and the other does not. The relevant comparison is between their effect estimates, with appropriate uncertainty. Gelman and Stern explain why different significance labels do not establish a difference between effects. S204

When explaining disagreement, distinguish a tested explanation from a hypothesis. “Different sampling times may contribute to the contrast” is a proposal if timing was not directly evaluated. It should not become “The contrast was caused by sampling time” without supporting evidence.

Use theory to interpret, not to decorate

A theoretical framework should help explain the relationship between the question, design and findings. If a social science paper uses a theory of help seeking, show which aspects of the material support, complicate or extend its expectations. Do not cite the theory in the Introduction and then ignore it in the Discussion.

A natural science paper may compare a mechanistic explanation with alternatives such as measurement artefact or compensatory processes. An engineering paper may explain an accuracy and efficiency tradeoff using the system's design. If the study did not measure a proposed mechanism, make that gap explicit.

An alternative explanation is a credible account of the findings that differs from your preferred interpretation. Good writing treats alternatives according to their plausibility and evidence. It does not list every imaginable possibility, and it does not dismiss a serious alternative in a single sentence without analysis.

Write limitations as a chain of consequences

A limitation paragraph should identify the issue, explain what it could change, describe any steps taken to address it, and state what remains unresolved. “The sample was small” is only the beginning.

For example: “Only a small number of independent growth containers was studied. Consequently, the interval for the mean difference is wide and does not distinguish a negligible change from a change relevant to the classroom comparison. Repeated measurements of seedlings improved description within containers but did not replace independent containers.”

The paragraph explains the consequence of sample size and prevents the reader from mistaking repeated readings for additional replication. Its conclusion should match the actual interval. Do not assert that uncertainty is large without inspecting the analysis.

Limitation Question to answer in the Discussion
Selection Who or what could enter the study, and who is missing?
Measurement Could the instrument or scoring rule distort the construct of interest?
Confounding Which alternative causes remain plausible?
Missing data Could the probability of missingness relate to the outcome?
Analysis assumptions Which conclusions depend on an uncertain model choice?
Scope Which new settings or populations have not been evaluated?

Avoid claiming that every limitation makes your estimate conservative. The direction of bias may be unknown. If so, say that the direction is uncertain and explain why. Honest uncertainty is more informative than a reassuring claim without evidence.

A claim is supported by evidence, bounded by design and uncertainty, and connected to a specific next study.

Interpret uncertain results without declaring failure

A result that does not cross a statistical threshold can still be informative. Ask which effect sizes remain compatible with the analysis and whether that range resolves the scientific question. A narrowly estimated difference near zero and a very imprecise difference near zero are not the same finding.

A confidence interval does not repair bias, unrepresentative sampling or an inappropriate model. Its interpretation depends on how data and analyses were generated. Greenland and colleagues discuss these limits and the common mistake of treating statistical nonsignificance as support for no effect. S205

Do not use “no effect” as a synonym for p > 0.05. If equivalence is the aim, explain the prespecified range of differences considered practically unimportant and use a suitable method. If the study was exploratory, frame the estimate as a basis for further investigation rather than a definitive acceptance or rejection of a theory.

Worked Discussion: natural science

The following is a hypothetical writing example. No plant study or numerical result is being reported.

The observed height difference would support a conclusion about the measured height outcome under the tested schedules and classroom conditions. It would not by itself show that the taller seedlings were healthier or that the same schedule would improve later biomass or survival. Those outcomes involve different measurements and may not change in the same direction.

One possible explanation is that the schedules alter growth allocation, but the proposed study does not directly measure the process responsible. Differences in temperature or other environmental conditions could also contribute if they were not adequately separated from the light schedule. The Discussion should describe the actual control and monitoring procedures before arguing for a particular mechanism.

The next useful study would follow independently assigned growth units while measuring both height and an additional justified growth outcome. That design would test whether the initial height comparison corresponds to a broader developmental change. The present project would remain useful as a bounded comparison and as a demonstration of a reproducible measurement approach.

The passage moves from the measured outcome to alternatives and then to a targeted next study. It does not inflate “height” into a global judgement of plant quality. In an actual Discussion, add relevant sources and replace conditional wording with the results that were observed.

Worked Discussion: social science

This fictional example concerns a voluntary mentoring survey.

An association between attendance and reported research confidence would be consistent with mentoring being relevant to how students evaluate their abilities. However, students who choose to attend may differ in prior experience, motivation or available time. Even an adjusted model can leave unmeasured differences unresolved. The result therefore concerns the observed relationship rather than the effect of assigning mentoring sessions.

The distinction between confidence and demonstrated competence is also important. A student may become more aware of task complexity and report lower confidence while developing stronger practical skills. Conversely, confidence may increase without a corresponding improvement on a task. Because the hypothetical survey measures confidence only, it cannot distinguish these possibilities.

A useful follow up would combine a clearly defined confidence measure with a relevant performance assessment and a design that better addresses selection into mentoring. The present association could inform the choice of measures and comparisons, but it would not alone justify a claim that expanding attendance will improve achievement.

This passage shows why a limitation is part of interpretation rather than an apology at the end. A qualitative version could instead examine how students describe the meaning of confidence and how those accounts challenge an existing conceptual framework.

Worked Discussion: engineering

For the synthetic controller example, a full interpretive paragraph should retain the provenance of the data:

The constructed comparison illustrates how pairing can clarify an algorithm evaluation. Because each case contains both controller values, the difference is defined within a shared case rather than across unrelated environments. Reporting the individual differences also shows that the average reduction is not produced by one isolated case. These properties help teach the relationship between design and analysis, but they do not validate a controller. The values were deliberately constructed, and no physical or simulated evaluation generated them.

A real evaluation would need to establish whether an accuracy improvement persists across relevant noise levels, layouts and hardware conditions. It would also need to report failures and resource costs. Lower average error may not make a controller preferable if it fails more often or exceeds a real time computation constraint. The appropriate next comparison depends on the intended application and its requirements.

This is an example of external validity, the relevance of findings beyond the studied conditions. It is separate from whether a calculation was performed correctly. A perfectly reproduced analysis of a narrow benchmark can still leave deployment performance uncertain.

Adapt the Discussion for a review

A review Discussion should explain what the body of evidence collectively supports and how the review changes understanding. Discuss the limitations of the included studies separately from limitations of the review process. Weak outcome measurement in the literature is different from your review searching only one database or excluding studies in languages you could not assess.

Explain inconsistent definitions, variation in effects, risk of bias and gaps in evidence. Avoid implying that a large number of citations creates certainty. A narrative review should acknowledge how selection and interpretation may shape its account. A systematic review should connect conclusions to the appraisal and synthesis actually performed. A scoping review should not promise a causal effectiveness conclusion if its methods were designed to map evidence.

End with a bounded conclusion

A conclusion should answer the original aim and preserve the most important qualification. It should not introduce a new outcome, mechanism or policy recommendation. A useful pattern is: the study found this within this scope; the evidence supports this interpretation; the main unresolved issue is this.

Practice: Revise “Our study proves that controller B should replace existing systems everywhere.” Model answer for the teaching dataset: “The synthetic example demonstrates how to report a paired error comparison with an effect estimate and uncertainty. It provides no evidence about deployment performance.” Model answer for a real but limited evaluation: State the actual benchmark result and conditions, then identify the untested deployment requirement. Use Template 7 to link each discussion claim to its supporting result and limitation.

Sources

S01: ICMJE. Preparing a manuscript for submission to a medical journal

S204: Gelman A, Stern H. 2006. The Difference Between “Significant” and “Not Significant” is not Itself Statistically Significant. The American Statistician 60(4):328 to 331.

S205: Greenland S, Senn SJ, Rothman KJ, Carlin JB, Poole C, Goodman SN, Altman DG. 2016. Statistical tests, P values, confidence intervals, and power: a guide to misinterpretations. European Journal of Epidemiology 31:337 to 350.