How to Write the Discussion Section of a Medical Manuscript: A 2026 Guide

Ask any editor at a major medical journal where manuscripts lose momentum, and you will hear the same answer: the discussion. Not the methods, which authors usually prepare carefully. Not the results, which present themselves. The discussion is where careful scientists often stop being careful writers, either repeating the results in different words, overclaiming what the data can support, or burying every conclusion under so many hedges that the reader finishes without knowing what to believe.

The discipline of writing a good discussion is distinct from the discipline of doing good research. You can have a perfectly designed trial and a muddy, circuitous discussion that leaves reviewers unsatisfied. Conversely, a modest but well-framed study with a tightly reasoned discussion can succeed at journals that seem out of reach. This guide focuses on the mechanics: what the discussion section is actually supposed to accomplish, how to organize it for a medical readership, where most authors go wrong, and what the current editorial environment in 2026 expects.

What the Discussion Section Is Actually For

The discussion is not a second results section. It is not a literature review. It is not an opportunity to list everything you wished the study had included. Its purpose is interpretation: taking what you found and explaining what it means in the context of everything else that is known. That sounds simple, but it requires a shift in posture that many researchers find uncomfortable, because interpretation involves judgment, and judgment can be wrong.

In the biomedical literature, the discussion section performs three specific functions. It situates your findings within the existing body of evidence, identifying where your results align with prior work, where they conflict, and what might explain the conflicts. It evaluates the strength of the evidence you have produced, including the study's limitations and what they mean for interpreting the results. And it draws conclusions about what the findings imply for practice, policy, or future research.

Notice that none of these functions require the discussion to be long. The JAMA editors' guide, updated in 2025 and now part of Oxford Academic's authorship resources, recommends five to eight focused paragraphs for most original research. MDPI's author guidance, updated in July 2026, makes the same point: the discussion should occupy roughly a third of the total manuscript length, with emphasis on depth rather than breadth. Length alone does not demonstrate intellectual engagement. A four-paragraph discussion that answers the main scientific question is more useful than an eight-paragraph discussion that answers nothing.

Opening with What You Found, Not What Others Found

The single most common structural error in medical discussions is beginning with background instead of findings. Authors often open the discussion by restating what is already known, as if repeating the introduction will remind the reader why the study matters. It does not. The reader has already read the introduction and the results. They want to know what you think about what you found.

The first paragraph of the discussion should state the principal finding, plainly and specifically, without redundant statistics. Not the finding as it appeared in the results table, but the finding as you interpret it. There is a difference between "Patients in the intervention arm had a 23% lower rate of readmission at 30 days" (a results statement) and "These findings suggest that the bundled discharge protocol reduces short-term readmission at a magnitude consistent with the effect size observed in two smaller prior trials" (an interpretive opening). The second version is what a discussion is for.

If your study has more than one main finding, the opening paragraph still needs to anchor on the primary outcome. Secondary findings can be introduced in subsequent paragraphs, but they should follow the logic set by the primary analysis rather than appearing as a list of miscellaneous observations.

Contextualizing Findings Without Losing the Thread

After you have introduced your principal finding, the second function of the discussion takes over: placing the finding in context. This means comparing your results to prior studies, and doing so in a way that actually helps the reader understand the cumulative state of evidence, not just performing the act of citation.

The mistake here is treating literature comparison as a checklist exercise. Authors will name five or six prior studies, note that their results are "consistent with" or "differ from" each one, and move on. That approach does not help a reader who wants to know why the field believes what it believes and how your data changes that picture. A more useful pattern is to group the prior evidence thematically: studies that support your finding and why their populations or designs are comparable; studies that appear to contradict your finding and what methodological differences might explain the discrepancy; and, where relevant, meta-analytic estimates that provide a reference point for your effect size.

One practical discipline is to never cite more studies than you can say something specific about. If you are citing a paper only to show it exists, that citation is probably not serving the discussion. Reviewers who know the literature well will notice when comparisons are superficial, and editors at journals like NEJM, JAMA, or The Lancet routinely comment on discussions that cite prolifically but reason poorly. The journal submission environment in 2026, with far more manuscripts competing for review attention, rewards concision and sharpness over volume.

Handling Contradictory Evidence Honestly

The most credible discussions engage directly with evidence that does not support their conclusions. Reviewers are trained to identify papers that present a selective literature review, and the temptation to do this is real. If three prior studies found the opposite of what you found, downplaying them or glossing over them is a judgment error, not just an ethical one. It makes your conclusions less reliable, because you are asking readers to trust an interpretation built on incomplete information.

The right approach is to name contradictory studies explicitly and offer a genuine account of what might explain the difference. Population differences, intervention timing, comparator arms, outcome definitions, follow-up length, and setting are all legitimate sources of heterogeneity that can produce conflicting findings from methodologically sound studies. If you cannot offer a plausible explanation for the discrepancy, say so. An honest admission that the evidence is inconsistent and that further work is needed to resolve it is more credible than a forced synthesis that papers over real uncertainty.

This matters especially for meta-analyses and systematic reviews, where the discussion should address heterogeneity directly rather than treating pooled estimates as more definitive than they are. Reviewers who read clinical trials and systematic reviews at major journals will check whether your discussion takes heterogeneity seriously. A discussion that presents a pooled relative risk without engaging with the I-squared value or the distribution of individual study effects is producing misleading reassurance.

How to Write the Limitations Section

Limitations deserve their own section in most medical journals, though some journals integrate them into the discussion body. Either way, the limitations passage is one of the most carefully read parts of the manuscript, and it is consistently one of the most poorly written. Two failure modes dominate.

The first is under-declaration: listing only the limitations that feel least damaging and ignoring the ones that should actually affect how readers interpret the results. A retrospective chart review that presents adjusted associations as if they were close to causal is a different paper from one that openly discusses residual confounding, unmeasured variables, and the implications for interpretation. Reviewers notice which version they are reading. At COPE-affiliated journals, where editorial standards on research integrity have tightened since 2025, under-declared limitations create a substantive credibility problem, not just a stylistic one.

The second failure mode is over-apologizing to the point of self-destruction. Some authors list every imaginable limitation, each presented with language that seems to suggest the findings should not be trusted. That is not intellectual honesty; it is overcorrection that undermines the paper's value. The goal of the limitations section is not to disclaim the findings but to give the reader enough information to decide how much weight to put on them. Name the real limitations. Explain their direction and magnitude of impact. Note, where genuine, what the study did to minimize them.

One increasingly common expectation at clinical journals is that the limitations section address not just study design weaknesses but also transparency limitations: whether data are available to independent researchers, whether the analysis code is documented, whether the protocol was pre-registered. The CONSORT 2025 update and the updated ICMJE recommendations from January 2026 both push in this direction. Authors who work from pre-registered protocols and publicly deposited datasets have a genuine advantage here, because they can note these features in response to limitations that might otherwise be harder to address.

Clinical Significance Versus Statistical Significance

This distinction has become one of the most watched aspects of medical manuscript evaluation in recent years, and the discussion section is where it has to be addressed. Statistical significance is a property of the analysis; clinical significance is a judgment about whether the observed effect matters to patients, practitioners, or health systems. A p-value below 0.05 does not, by itself, establish that a result is meaningful. Many medical journals now ask authors to address clinical significance explicitly, and some, including BMJ and JAMA, routinely return manuscripts for revision when the discussion conflates the two.

The practical way to handle this is to identify the minimum clinically important difference (MCID) or a contextual threshold for your outcome before you write the discussion, and then explicitly compare your observed effect to that threshold. If your intervention reduced systolic blood pressure by 2 mmHg with a p-value below 0.001 in a large trial, the discussion should address whether a 2 mmHg reduction is meaningful in the target population, what the existing MCID literature says, and whether the statistical significance reflects real-world significance or a large sample size overwhelming a small effect.

Confidence intervals carry more information than p-values for this purpose, and most current reporting guidelines prefer them. The SAMPL guidelines, which many statistical reviewers at medical journals consult, specifically recommend presenting effect sizes with confidence intervals and interpreting the clinical range rather than the p-value boundary. If your results include a tight confidence interval around a small effect, say so clearly and let the reader judge whether that precision is reassuring or whether it just confirms a modest result.

Implications, Future Directions, and Where Most Authors Overreach

The closing section of a discussion usually addresses what the findings mean for practice or policy and what research should follow. This is the section most prone to overclaiming, because it is where authors feel pressure to make the paper sound important.

The test for a clinical implication is whether it is genuinely supported by your evidence at the level of confidence your study warrants. A single-center retrospective study does not establish a clinical practice change. It may suggest a hypothesis worth testing in a prospective design, or provide new evidence for the feasibility of an approach that has been theoretically attractive. That is a real contribution, and you can say so without pretending the evidence is stronger than it is. Reviewers at major journals are sensitive to this, particularly at journals like the New England Journal of Medicine or The Lancet, which have raised their standards for the language around clinical recommendations since the pandemic-era preprint wave.

Future directions should be specific. "Future research should explore this topic further" is not a useful sentence and editors notice it. Better future direction statements name the specific design that would address the remaining uncertainty, the population it should target, and the outcome it should prioritize. If your study was a pilot or feasibility trial, the natural future direction is a powered randomized controlled trial, and you can describe the design parameters that your pilot suggests are feasible.

What Editors and Reviewers Actually Notice in 2026

The editorial environment has shifted in ways that directly affect what a strong discussion looks like. With more submissions than ever competing for review slots (a 33% increase in submissions on major platforms in the first quarter of 2026 compared with the same period in 2025), editors are spending less time on manuscripts that require major structural work in the discussion. Papers that require a reviewer to do the interpretive work that the discussion should have already done are at higher risk of rejection at the review stage.

Reviewers increasingly flag two specific problems in discussions. The first is what some call "discussion-as-introduction": opening with background instead of findings, filling space with known literature, and deferring the interpretive work until the final paragraph. The second is inconsistency between the results and the discussion, where the results present one set of findings and the discussion interprets a somewhat different or more favorable version of them. Statistical reviewers in particular are trained to cross-reference the results and the discussion, and inconsistencies cause disproportionate suspicion even when they are unintentional.

AI writing tools have added another dimension. Discussions that show signs of AI-assisted writing, with its characteristic over-hedged and over-structured prose, are now familiar enough that reviewers at several major journals will comment on them. This does not mean AI tools cannot be used to improve language or check for missing elements. It means that the intellectual work of building the discussion's argument should be yours, and if you use a language tool to refine phrasing, you should apply enough editorial judgment that the final text reflects your reasoning rather than the model's default rhetorical patterns.

A Pre-Submission Discussion Checklist

Before you finalize the discussion, read it as a reviewer who is skeptical of the main conclusion. Ask whether each paragraph does its job or whether it is filler. This is easier to do after a short break from the manuscript, which is one of several reasons why leaving the discussion revision to the last minute is a practical risk as well as an intellectual one.

If the answer to the last question is no, the discussion is not finished. A reader who closes a paper without a clear sense of what they should now believe has not been served by the discussion. That is the ultimate test: not whether the section is long enough or covers all the required elements, but whether it actually answers the question the study was designed to ask.

The Relationship Between the Discussion and Your Reporting Checklist

Most major reporting standards include specific discussion requirements that authors often overlook when they are focused on completing the main text. CONSORT 2025, updated from the earlier version, includes explicit checklist items for the discussion of generalizability, clinical relevance, and the interpretation of trial results in light of other evidence. STROBE, the observational study reporting guideline, asks specifically for a discussion of causal versus associational language. PRISMA 2020 requires a discussion of certainty of evidence and a discussion of applicability.

The practical implication is that your discussion and your reporting checklist should be written together, or at least cross-checked before submission. If your CONSORT checklist cites a page number for a discussion item, read that passage to confirm it actually addresses what CONSORT expects. Many manuscripts submit reporting checklists that cite passages where the relevant content is thin or absent, and reviewers who know these standards will notice when there is a mismatch between the checklist and the text.

The same applies to pre-registration. If your study was pre-registered, the discussion should address the relationship between your pre-specified hypotheses and outcomes and your actual results. If secondary or exploratory analyses produced the most interesting findings, say so clearly and explain how that affects the inferential weight those results should carry. Journals that have adopted open science practices, including the BMJ, PLOS Medicine, and several specialty journals, increasingly expect this level of transparency in the discussion rather than treating all results as equally confirmatory.

Writing for Your Actual Audience

The discussion is usually addressed to specialists, but the calibration matters. A discussion written for a surgical subspecialty journal can assume deep familiarity with the technical context. A discussion written for a general medical journal like BMJ or JAMA should be interpretable by internists, general practitioners, and health policy readers who are smart and medically literate but not specialists in your field.

This distinction shapes vocabulary, the depth of methodological explanation, and how much context you need to provide for the comparator studies you cite. When in doubt, ask whether a senior clinician in a related but not identical specialty would follow your reasoning. If they would get lost around the third paragraph, the discussion is probably too technical or too specialized for a general venue.

One practical test is to read only the opening sentence of each paragraph and check whether those sentences, taken together, form a coherent interpretive argument. If they do, the paragraph structure is carrying the reader through the logic correctly. If the opening sentences are all background claims or transitional filler ("It is also worth noting that..."), the discussion is organized around completion of tasks rather than construction of an argument.

Getting the discussion right is largely a matter of practice and feedback. The authors whose discussions read most cleanly have usually written many drafts, read many well-edited papers in their field, and paid close attention to how reviewers have challenged their conclusions in the past. If you receive a review that says the discussion is "overclaiming" or "not consistent with the study design," treat that as specific feedback about your interpretive habits, not just about that one manuscript. Those habits transfer, in both directions.

Further Reading