How to write a discussion section

The inverted funnel

The introduction narrows from the field to your question; the discussion widens back out from your findings to the field. The standard sequence: restate the key findings in plain language (one paragraph, no statistics re-run), interpret them against your hypotheses, compare with prior work — agreement and conflict both — then limitations, implications, and future directions, closing on the contribution. Reviewers navigate by this shape; honoring it makes the paper feel finished.

The opening paragraph is a genre move worth practicing: answer the research question directly in the first two sentences. "This study asked whether X; the results indicate it does, but only under Y" orients every paragraph that follows. Discussions that open with throat-clearing ("Much research has examined…") squander the reader’s sharpest attention.

Interpret without overclaiming

The discussion’s core skill is calibration: claims sized exactly to the evidence. Concretely — match causal language to design (correlational studies "are associated with", only experiments "cause"); generalize only to populations your sample can speak for; and distinguish statistical from practical significance, ideally via effect sizes ("statistically significant but small, d = 0.15, likely negligible in practice"). Overclaiming is among the most cited reasons for major revisions; underclaiming merely wastes your contribution.

Hedged language is not weakness — it is precision. "These results suggest" and "one interpretation is" signal an author who knows where the evidence ends. The skill is hedging the interpretation while stating the finding itself plainly.

• Effect size benchmarks (what counts as small/large)

Put your results in conversation with the field

Every key finding should be located relative to prior work: does it replicate, extend, contradict, or qualify what others found? Agreement is quick to write; disagreement is where discussions earn their keep. When your results conflict with published work, offer a concrete account of why — different population, measure, dosage, era, analysis — rather than a shrug. That account is often the most-cited paragraph of the paper, because it is where the field learns something structural.

Unexpected findings deserve honest treatment, clearly labeled as post-hoc: describe what you found, offer candidate explanations, and mark them as hypotheses for future work rather than quietly promoting them to headline results — the move known as HARKing, which preregistration exists to prevent.

Limitations that build credibility

A good limitations paragraph does three things per limitation: names it specifically, states its plausible effect on the results (direction and rough size, not just existence), and where possible explains why the conclusions survive it. "The sample was drawn from one hospital, which may limit generalizability to primary-care settings; however, the effect replicated across all three wards" is a credibility machine. Generic boilerplate ("small sample size, more research needed") is the opposite — it signals the section was written to be skipped.

Choose the limitations that actually threaten your claims, usually two to four. Listing ten trivial ones buries the real conversation and invites reviewers to add their own.

The closing paragraph and the future-work trap

Future directions should be specific enough to be someone’s actual next study — "testing whether the effect survives a two-week delay would distinguish the memory account from the attention account" — not the ritual "more research is needed". Then close with the contribution: one paragraph, no new evidence, stating what the field can now do or believe that it could not before. Many readers read the abstract, the first discussion paragraph, and this one; make all three consistent.

Where Kahubi fits

Kahubi drafts the discussion from what the project actually contains: the statistics flow supplies the real effect sizes and their benchmarks, library chat surfaces where your findings agree or conflict with the papers you have collected — with citations back to them — and the agent writes the comparison in your voice, marking [CITATION NEEDED] where a claim would need a source you do not yet have.