Discussion & Limitations

In the Discussion you finally say what your numbers mean, and it is the section where papers most often go wrong. After pages of careful results, it is tempting to finish with one big confident claim. Don't. A good Discussion interprets as far as the evidence reaches and no further. Its limitations are specific, not boilerplate, and its future directions follow from what you found. The skill is calibration: matching the size of your claim to the size of your evidence.

The Discussion recipe

A clean discussion moves through five steps, roughly in this order:

1. Restate the main finding in plain words. Leave out the statistics and say what happened: "students who slept more tended to report higher wellbeing."
2. Relate it to prior work. Does it agree with, extend, or contradict what others found? Here you return to the literature from your Introduction.
3. Interpret. Now you may say what it means, why it might have happened, and what mechanism could explain it. This is the one section where interpretation belongs.
4. Limitations that matter: not a ritual "the sample was small," but the specific weaknesses that could actually change the conclusion, and how.
5. Future directions that follow: the next study this one implies, not a generic "more research is needed."

Useful limitations

Almost every thesis lists "a limitation was the small sample size," and it tells the grader nothing. A useful limitation says what it could have changed. Compare:

• ❌ "A limitation is that the sample was small."
• ✅ "With 38 participants the study was underpowered to detect small effects, so the non-significant interaction should not be read as evidence of no interaction."

The second names the limitation, its direction, and its consequence for a specific claim. Good limitations often point at the exact sentence in your Discussion they undercut. Graders give credit for limitations written this way.

One questionnaire, two variables

One specific limitation is almost never named in a student thesis, although it applies to the most common student design: both variables measured by self-report, on the same form, at the same sitting. Anything shared by the two measures other than the constructs themselves becomes part of their correlation. Mood on the day, how the items were worded, a habit of ticking the middle box, wanting to look consistent to the researcher: these push both answers in the same direction at once. That shared push is called common-method variance.

The arithmetic shows how large the problem can be. Suppose a method factor accounts for a share m of each measure's variance, and the two traits themselves correlate ρ. The correlation you observe is (1 − m)ρ + m. Set ρ to zero and the observed correlation is just m. A shared method accounting for a tenth of each measure's variance produces r = .10 between two unrelated variables, and a quarter produces r = .25.

Apply that to the study in the panel below, whose observed r = .28 came from one questionnaire. If a shared method accounts for 10% of the variance in each measure, the implied true correlation is .20. If it accounts for 25%, the implied true correlation is .04, and the finding almost disappears. Nothing in the data can tell you which. A limitation written this way names an alternative explanation, gives its direction (always upward), and puts a number on how much of the conclusion is at stake.

This limitation also has a fix, so raise it in your future directions too: measure the two variables by different methods or at different times, and the shared push mostly goes away. A sleep diary kept nightly and a wellbeing scale completed weeks later cannot share a mood or a response style the way two blocks of one survey can. Survey & Questionnaire Design covers the item-level half of this, and Bias & Blinding the participant-level half.

Calibrate your verbs to your design

The fastest way to overclaim is a single verb. Correlational data cannot license "causes," "leads to," "boosts," or "improves." Those are causal verbs, and only a design that supports causal inference justifies them (see Causal DAGs & Confounding). From an observational study you may write "was associated with," "predicted," "tended to," "was linked to." Two more calibration traps:

• Strength words. A correlation of r = .28 is not "a strong effect." Match the adjective to the number.
• Universal quantifiers. A group-level result doesn't mean "every" participant, and "all students" or "will feel better" turns an average into a guarantee.
• The level you measured at. If your rows were countries, schools or wards rather than people, every sentence has to stay at that level. Sliding down to individuals is the ecological fallacy, and it is a serious error. A relationship measured between group averages can be the reverse of the one inside those groups (Observational Designs works a case where +.77 between regions is −.10 between people). Climbing the other way is just as unsafe. Finding an effect in your 40 participants does not license a claim about the classes, clinics or countries they came from.

Underclaiming is a mistake too: calling a real association "no meaningful relationship" throws away information (see Writing About Non-Significant Results). Whatever you settle on here has to survive being compressed into 200 words, so check it again when you write the abstract, which is all most readers will ever read.

Catch the overclaims

Below is the result card from one study. Read it, then judge each discussion sentence: is it calibrated to this evidence, does it overclaim (says more than the data support), or does it underclaim (says less)? After you answer, the word that breaks calibration is highlighted.