For students with a draft discussion in an empirical psychology thesis, dissertation, or lab report. The page goes through the discussion one hypothesis at a time and covers the points psychology readers check: what a non-significant result allows you to say, which verbs your design supports, how far a student sample reaches, limitations that name their effect, and what replication research means for your comparisons.
No one sees their own blind spots, with or without AI help. Esmond points them out in your draft from within your field, along with factual errors and inconsistencies.
The discussion takes each result back to the hypothesis it tested and to the theory and studies that produced the hypothesis. It opens with a short summary of the main findings, works through what they mean, weighs other explanations, names the limits of the study, and ends with implications and a conclusion. The Psychological Methods group at Ruhr University Bochum adds the rule that runs through all of it: “It is also important not to draw conclusions that you cannot support with your data.”
This page assumes the statistics are already reported in full, as described for the psychology results section. The theory the discussion returns to was set out in the psychology dissertation introduction. For the general difference between the two chapters in any field, see results vs. discussion.
A hypothetical example of a commentYour discussion says loneliness leads to poorer sleep, but your study is a single cross-sectional survey. Which part of your data shows the direction of the effect?
A first comment on one part of your own draft, right away and without an account.
Get a free previewA psychology discussion is easiest to check when it follows the hypotheses in order. For each one, a reader looks for four steps: the result in one plain sentence (without repeating the statistics), whether it matches the prediction, how it compares with the studies cited in the introduction, and what it means for the theory behind the prediction.
Then compare your results with earlier studies. If your effect is smaller than in the study you built on, ask what differs: the sample, the measure, the manipulation, the setting. Bochum's information sheet asks students to consider alternative explanations, especially when results are unexpected, and to think about how those explanations could be tested in future studies. A result that matches the prediction needs that scrutiny too, since the same pattern may follow from a simpler account.
A non-significant test means the data didn't allow you to reject the null hypothesis at your alpha level. It doesn't show that the effect is zero. For your own text, that means no relabeling ("a trend", "marginally significant") and no claim that the effect is absent.
The confidence interval tells you which reading the data allow. For r = -.12, 95% CI [-.36, .14] with 60 participants, the data are compatible with no relationship, with a moderate negative one, and with a small positive one; the honest conclusion is that the study can't tell these apart. To claim that an effect is negligible, you need a narrow interval around zero, or an equivalence test against a smallest effect size of interest set in advance.
Power belongs in this paragraph when the sample turned out smaller than planned. To interpret p-values when the obtained sample is smaller than planned, the University of Guelph's Department of Psychology says “you should calculate power after you have finished data collection based on your obtained sample size.” Its guidance adds that this calculation uses the effect sizes specified before data collection, not the effect you observed. Power computed from the observed effect is a direct function of the p-value and adds no information.
Verbs in a discussion make claims about design. SWPS University's standards for master's theses in psychology treat the word "influence" as reserved for experimental research and ask that conclusions from correlational studies be worded precisely, to avoid unjustified cause-and-effect conclusions. The table shows the usual fit between design and wording.
| Design | What it can show | Wording that fits | Wording a reader will challenge |
|---|---|---|---|
| Experiment with random assignment | An effect of the manipulation, under these conditions | "caused", "led to", "reduced" (for the manipulated variable) | Claims about variables that weren't manipulated |
| Cross-sectional survey | An association at one point in time | "was associated with", "predicted" (statistically) | "increases", "leads to", "influences" |
| Longitudinal survey | Temporal order and change | "preceded", "predicted later" | Causal claims without ruling out third variables |
| Mediation model on cross-sectional data | A pattern consistent with a proposed pathway | "consistent with a mediating role" | "X works through Y" |
Many thesis studies recruit psychology undergraduates or online volunteers. That is a legitimate sample, and a discussion should say who it represents. A sample of 98 first-year psychology students, mostly women, taking part for course credit, supports conclusions about similar students. Statements about "people" or "adults" need a reason to expect the effect to hold in other ages, cultures, or settings, and the discussion should give that reason or narrow the claim.
The University of Hamburg's guidelines list the appropriateness of generalizing the results among the criteria for the discussion, next to embedding the results in the current state of research and discussing strengths and weaknesses of the study.
A useful limitation in a psychology thesis says what the problem is, which result it affects, and in which direction it could bias that result. "The sample was small" says little. "With 60 participants, power was below .80 for correlations smaller than about r = .35, so the null result for Hypothesis 2 is inconclusive" tells the reader how to read a finding. The same goes for self-report measures (social desirability may have reduced reported alcohol use), common-method variance in a single questionnaire, or an online sample with attention-check failures.
Bochum's information sheet asks for constructive limitations, with specific suggestions on how future studies could address them. SWPS University's standards suggest describing, briefly, a study that would be better than the one reported. For placement and the difference between scope and limitations in general, see the master's thesis limitations section.
Since the replication crisis, a psychology reader checks what an effect is compared with. Large replication projects in psychology have found many effects to be smaller on repetition than in the original studies. A single original study is therefore a weak benchmark. If your effect is smaller than that study's, the gap may say as much about the original as about your sample; meta-analytic or replication estimates are a fairer comparison.
Applied implications should fit the whole confidence interval. An intervention effect of d = 0.60 with a 95% interval from 0.10 to 1.10 is compatible with a very small benefit, so a recommendation to roll the intervention out widely claims more than the data show. Theoretical implications work the same way: one study can support or weaken a prediction, and the discussion says which, and how strongly.
The University of Warsaw's Faculty of Psychology lists its substantive criteria for the discussion of results in a master's thesis: “the correctness of the interpretation of the results; the logic of the references to the hypotheses formulated and the relevant theoretical strands; the adequacy of the limitations of the study indicated”. Those three criteria match the three sections above: interpretation per hypothesis, the link back to theory, and limitations that fit the study.
Bochum's group allows AI tools only as assistants and requires every use to be documented in the thesis. Hamburg's lab requires use of generative AI to be transparent and documented as well. If you study under rules like these, ask your supervisor before you use Esmond on your discussion, and document it as required.
These fragments are made up. None of them comes from a real thesis.
Esmond evaluates every assignment before you submit. It never writes it.
Esmond is an AI tool that doesn't write for you. If your course doesn't allow any AI tools, don't use Esmond for that work. If you're unsure, ask your instructor or advisor.
In a discussion, Esmond can flag claims in your text that go further than your results or design: a non-significant result described as no effect, causal wording for correlational data, conclusions about a population your sample doesn't represent, a limitation that isn't tied to the result it affects, or a hypothesis the discussion never returns to. See the sample report for the format of the comments, and how Esmond works for the steps. The other chapter pages start from thesis chapter feedback.
Esmond reads what you upload. If the file holds the whole draft, it can set this chapter against the others; if it holds one chapter, the review stays within that chapter and may raise a point that another chapter already answers.
Esmond doesn't rerun your analysis, doesn't search the literature for studies you missed, and doesn't decide which interpretation is right. You weigh each comment and decide what to change.
A short summary of the main findings, an interpretation per hypothesis in relation to the theory and studies from the introduction, alternative explanations, limitations with their effect on the findings, implications for theory and practice, and a conclusion. A dissertation with several studies often adds a general discussion across them.
Report what the confidence interval allows: which effect sizes the data rule out and which they don't. Avoid both "no effect" and "a trend". If the sample was smaller than planned, report power for the effect size you planned for, not for the effect you observed.
The logic is the same: interpret each result against the hypothesis, compare with earlier findings, name limitations, and stay within the data. A lab report discussion is shorter and usually covers one study and fewer hypotheses.
Not in full. Refer to each result in words and give a number only where the argument needs it, such as an effect size you compare with earlier studies or the bounds of an interval that limit a claim.
Only as far as the design allows. An experiment with random assignment supports causal claims about the manipulated variable; a correlational or cross-sectional study supports associations. SWPS University's standards reserve the word influence for experimental research.
Read on November 8, 2026.
Get your paper evaluated before you submit