ACADEMIC WRITING GUIDE
How to Calculate and Interpret Effect Sizes in a Dissertation (With Worked Examples)
A dissertation can report a statistically significant difference without explaining whether that difference is substantial. Likewise, a result that does not reach a conventional significance threshold may still be compatible with a range of effects worth investigating. Effect sizes help students describe the magnitude and direction of results rather than relying on p-values alone.
The challenge is choosing an effect size that fits the research question and study design. A comparison of two independent groups requires different decisions from a correlation or an analysis of variance (ANOVA). The number itself must then be interpreted against its units, uncertainty, measurement quality, and research context.
This guide explains three common effect-size families through worked calculations, shows how to write them into a dissertation, and identifies the mistakes that often make results sound more certain than they are. Every dataset and numerical result below is hypothetical and created for teaching, not evidence from real participants.
What is an effect size?
An effect size is a quantitative expression of the size of a difference, relationship, or model-related effect. Some effect sizes retain meaningful units, such as a six-point difference on a test. Others standardize a difference or summarize a relationship, allowing certain comparisons across measures or studies.
For example, imagine two study groups whose average assessment scores differ by six points. Reporting the six-point difference describes the finding in the original measurement units. Reporting Cohen's d adds information about how large that difference is relative to the groups' observed variability. Neither automatically tells you whether the difference matters to a university, a lecturer, or a student.
The American Psychological Association's reporting guidance encourages communicating effect estimates and uncertainty rather than substituting statistical significance for substantive interpretation. Our existing guide to reporting statistical results explains the broader results-chapter context. This article focuses on selecting and calculating effect sizes.
Effect size is not the same as statistical significance
A p-value helps assess how incompatible the data are with a specified statistical model or null hypothesis, subject to its assumptions. It is not the probability that a research hypothesis is true, and it does not measure the practical importance of the estimate.
An effect estimate addresses a different question: How large is the observed difference or relationship? An accompanying confidence interval helps communicate estimation uncertainty. A large point estimate with a wide interval may provide weaker guidance than a smaller, precisely estimated result.
| Quantity | Question it helps answer | What it cannot establish alone |
|---|---|---|
| Mean difference | How far apart are the group averages, in original units? | Whether groups differ causally |
| Cohen's d | How large is a mean difference relative to a defined standard deviation? | Whether the difference matters in context |
| Pearson's r | What is the direction and strength of a linear association? | Causation or the size of a change in original units |
| Eta-squared, η² | What fraction of observed outcome variability is associated with a factor in a particular ANOVA model? | A causal share of variability |
| Confidence interval | How precise is an estimate under the stated method? | Certainty about the parameter or future outcomes |
Step 1: Choose the measure that matches the analysis
Do not pick an effect size simply because it appears in a thesis template. First identify your outcome, comparison, and model.
| Research situation | Useful estimate | Important design question |
|---|---|---|
| Two independent groups measured on a continuous outcome | Raw mean difference; standardized mean difference such as Cohen's d | Are samples independent? Is pooled standardization appropriate? |
| Same participants measured twice | Mean paired change; a clearly defined paired standardized effect | Which standard deviation is used for standardization? |
| Two continuous variables | Pearson's correlation r where linear association assumptions fit | Are outliers or nonlinearity affecting interpretation? |
| One-way ANOVA | η², or a justified alternative such as ω² | Is the factor definition and sums-of-squares basis clear? |
| Multivariable ANOVA | Partial η² where appropriate | Is this partial measure being confused with ordinary η²? |
| Binary outcome | Risk difference, risk ratio, or odds ratio according to design | Are odds being mistaken for probabilities or risks? |
There are multiple legitimate effect-size definitions within some families. This matters especially for repeated-measures research, unequal group variances, adjusted models, and partial eta-squared. Report the exact measure—not merely “effect size = 0.5.” For a broader choice of tests and outcomes, see our dissertation data analysis plan guide.
Worked example 1: Calculate Cohen's d for two independent groups
Hypothetical question: Do students in two independent tutorial formats differ in their average assessment scores?
Suppose the summary results are:
| Measure | Tutorial A | Tutorial B |
|---|---|---|
| Students (n) | 25 | 25 |
| Average score | 76 | 70 |
| Sample standard deviation | 10 | 8 |
Step A: Calculate the mean difference
Mean difference = 76 − 70 = 6 score points.
This is the most directly interpretable estimate: Tutorial A's observed sample mean is six points higher.
Step B: Calculate the pooled standard deviation
For two independent groups under the pooled-variance standardization convention:
s_p = sqrt(((n_A − 1) × s_A² + (n_B − 1) × s_B²) / (n_A + n_B − 2))
Substitute the values:
s_p = sqrt(((24 × 100) + (24 × 64)) / 48)
s_p = sqrt(3936 / 48) = sqrt(82) ≈ 9.055
Step C: Calculate Cohen's d
d = (mean_A − mean_B) / s_p = 6 / 9.055 ≈ 0.663
Result: Cohen's d ≈ 0.66, with a raw mean difference of six points in this hypothetical sample.
A positive sign means the difference is in the direction of Tutorial A minus Tutorial B. Reversing the subtraction reverses the sign, not the underlying evidence.
Step D: Interpret responsibly
A value around 0.66 represents a difference of approximately two-thirds of the selected pooled standard deviation. Some readers recognize conventional “small/medium/large” benchmarks for d, but such thresholds are context-dependent heuristics. They cannot determine whether a six-point difference is educationally important or whether tutorial format caused it.
Ask instead: Is six points meaningful on this assessment? Is the scale reliable? Were participants assigned randomly or did groups differ before the tutorial? What interval and other evidence accompany the estimate?
Example dissertation paragraph:
In the hypothetical sample, Tutorial A had a higher mean assessment score (M = 76, SD = 10, n = 25) than Tutorial B (M = 70, SD = 8, n = 25). The observed mean difference was 6 points, corresponding to a pooled-standard-deviation Cohen's d of approximately 0.66. The standardized estimate describes the magnitude of the sample difference; because the study design has not established random assignment or control of confounding, it should not be interpreted as a causal effect of tutorial format.
Important: The paragraph does not invent a t statistic, p-value, or confidence interval. If you have performed those analyses, report their verified results separately. For small samples, a bias-corrected effect such as Hedges' g may be appropriate; identify the adjustment rather than silently relabeling d.
Worked example 2: Interpret a correlation effect size
Hypothetical question: Is perceived feedback usefulness related to student engagement scores?
Suppose an appropriate Pearson correlation analysis on 80 complete paired observations produces r = 0.40. This coefficient is stipulated for illustration; it has not been estimated from an underlying published dataset.
Pearson's r runs from −1 to +1. A value of +0.40 indicates a positive linear relationship in this hypothetical sample: students with higher feedback-usefulness scores tend also to have higher engagement scores. A value of −0.40 would indicate the opposite direction.
A simple descriptive calculation is:
r² = 0.40² = 0.16
That is 16%, but it needs careful wording. In a simple linear regression with an intercept and a single predictor applied to the same observations, r² equals the model's unadjusted R², describing the proportion of variation in the observed outcome accounted for by the fitted linear relationship. It does not mean “feedback causes 16% of engagement,” nor does it describe how much any individual student's engagement will change.
| Calculation | Hypothetical result | Meaning |
|---|---|---|
| Pearson's r | +0.40 | Positive sample linear association |
| r² | 0.16 | 16% unadjusted explained variation in the corresponding simple linear model |
| Sample size | 80 | Number of complete paired records, not separate item totals |
Example dissertation paragraph:
In this hypothetical sample of 80 complete cases, feedback usefulness and engagement showed a positive linear association (r = .40). For a corresponding single-predictor linear regression with an intercept, r² = .16, indicating that the fitted linear relationship accounted for 16% of observed engagement-score variation. This association does not establish whether perceived feedback usefulness increases engagement, whether engaged students evaluate feedback more positively, or whether other variables explain both measures.
Inspect a scatterplot before relying on a correlation: one influential observation can change r, while a curved relationship may be poorly described by a single linear coefficient. A confidence interval and appropriate model checks improve interpretation. Our quantitative data analysis guide explains how statistical findings relate to the questions being studied.
Worked example 3: Calculate eta-squared for a one-way ANOVA
Hypothetical question: How much of the variation in assessment scores is associated with tutorial-group membership across three groups?
Imagine a one-way ANOVA summary with the following internally consistent teaching values:
| Source of variation | Sum of squares (SS) | Degrees of freedom |
|---|---|---|
| Between groups | 120 | 2 |
| Within groups (error) | 480 | 57 |
| Total | 600 | 59 |
The degrees of freedom correspond to three groups and 60 total observations: between-groups df = 3 − 1 = 2; within-groups df = 60 − 3 = 57.
For this one-way ANOVA:
η² = SS_between / SS_total = 120 / 600 = 0.20
Result: η² = .20. In this particular unadjusted ANOVA model, the grouping factor accounts for 20% of the sample's total outcome sum of squares. This is a model-based descriptive proportion, not evidence that group membership caused 20% of students' outcomes.
You can also calculate the ANOVA F ratio to verify that the table is coherent:
MS_between = 120 / 2 = 60
MS_within = 480 / 57 ≈ 8.421
F = 60 / (480 / 57) = 7.125
The example intentionally does not assert a p-value. A real dissertation should report the value calculated from its own analysis, including model details and assumptions.
What about partial eta-squared?
Partial eta-squared, written η²p, uses the factor's sum of squares divided by the factor's sum of squares plus its associated error sum of squares. In this simple one-way design, the two definitions yield the same numerical value because the decomposition has only one factor and its residual error. In factorial or repeated-measures ANOVA, partial η² and ordinary η² generally differ and must not be swapped as if they were equivalent.
Example dissertation paragraph:
In the hypothetical one-way ANOVA of 60 observations across three tutorial groups, the group-related sum of squares was 120 and the total sum of squares was 600, giving η² = .20. This indicates that tutorial-group membership was associated with 20% of observed assessment-score variation under the stated model. The result describes association in the sample; the importance and causal interpretation of group differences would depend on the study design and additional evidence.
For additional guidance on presenting the model and avoiding raw software-output dumps, see our research tables and figures guide.
How to interpret effect sizes without overselling them
1. Start with meaningful original units
A six-point difference is easier to evaluate when readers know whether the assessment is scored out of 10, 100, or 1,000. A standardized effect helps comparisons but does not replace the raw-unit context.
2. Check the study design
Independent, randomized groups can support different claims from self-selected, cross-sectional groups. An effect size is not a causal identification strategy. If the study is observational, use language such as “associated with” or “differed between.”
3. Consider precision and sample size
A point estimate from a small sample may be unstable. Report an appropriately calculated confidence interval where possible and avoid making decisions solely from whether the interval excludes zero. Do not construct a guessed interval from a rounded effect size or another study's standard error.
4. Use thresholds cautiously
Cohen-style benchmark conventions may help orient beginners, but meaning depends on the outcome, its scale, population, prior evidence, measurement reliability, and potential costs or benefits. A modest effect can matter in a large population; a seemingly large standardized difference may be inconsequential on a poorly chosen outcome.
5. Match the effect definition to the actual model
In paired studies, different standardized differences use different standardizers. In ANOVA, eta-squared, partial eta-squared and omega-squared are related but not interchangeable. Report enough information that another researcher could reproduce the measure you chose.
6. Separate evidence from recommendation
Your dissertation can reasonably discuss whether the estimated difference warrants further investigation. It should not recommend an intervention as proven effective merely because an estimate or p-value appears impressive. Our guide to dissertation recommendations based on findings shows how to keep practical proposals proportional to the evidence.
A comparison table for choosing and reporting an effect size
| Method or question | Estimate to report | Essential accompanying context | Avoid saying |
|---|---|---|---|
| Independent group comparison | Raw mean difference and defined d or g | n, means, SDs, interval if computed, group assignment | “d proves the teaching method worked” |
| Pearson correlation | r, with interval if computed | n, linearity, outliers, measurement quality | “r = .40 means 40% explained” |
| One-way ANOVA | η² or justified alternative | Model, SS basis, df, estimates and intervals where feasible | “η² is a causal percentage” |
| Factorial ANOVA | Precisely labeled partial η² or other measure | Factor, interaction, associated error term | “Partial η² equals ordinary η²” |
| Logistic regression | Odds ratio and interval | Outcome definition, reference category, covariates | “An odds ratio is a risk ratio” |
Common mistakes in dissertation effect-size reporting
Treating p-values as effect sizes. A significant p-value does not describe the magnitude of a result. Report a relevant effect estimate and uncertainty when the analysis supports it.
Using the wrong denominator. For Cohen's d, a pooled standard deviation differs from the standard error. Dividing the mean difference by the standard error does not give Cohen's d.
Confusing r with r². A correlation of .40 has a squared value of .16, not .40. Even .16 must be interpreted within the corresponding model rather than as a causal percentage.
Switching eta-squared definitions. Explicitly distinguish ordinary and partial eta-squared. This is especially important in multi-factor designs.
Interpreting a benchmark as a universal verdict. A conventional label such as “medium” is less informative than an explanation of the observed difference and why it might matter.
Inventing precision. Never report an interval, p-value, or standardized estimate that was not actually calculated and checked for the correct design.
Ignoring missingness and exclusions. State the analytic n and any material exclusions. Our missing survey data guide explains how missing records can change the evidence base.
Overclaiming causality. Effect sizes summarize results; causal claims require suitable research design, assumptions and evidence.
Dissertation-ready reporting templates
Independent-groups template:
In [study groups], the observed mean difference in [outcome and units] was [difference], with [group means, SDs and sample sizes]. Using [defined standardizer], the standardized mean difference was [d or g]. The [calculated interval, if available] indicates [bounded precision statement]. Given [study-design limitation], this is interpreted as [difference or association, not unsupported causation].
Correlation template:
Among [n] valid paired observations, [variable X] and [variable Y] showed a [positive/negative] linear association (r = [value], [interval if calculated]). The finding is interpreted alongside [scatterplot/measurement considerations] and does not by itself establish a causal relationship.
ANOVA template:
In the [specified ANOVA] model, [factor] was associated with an effect-size estimate of [precisely labeled η², partial η² or other measure]. The reported estimate was computed from [identified sum-of-squares basis] and is interpreted as a model-specific description of observed variation rather than proof of causal influence.
Adapt these templates to actual analyses and your institution's statistical-reporting conventions. Do not fill a template with numbers from a different dissertation.
Frequently asked questions
Is Cohen's d the same as the mean difference?
No. A raw mean difference is expressed in the outcome's original units. Cohen's d expresses the difference relative to a specified standard deviation. Both can be informative, but they answer different questions.
Can I calculate Cohen's d from a t-test?
Sometimes, if you know the exact t-test design, sample sizes and other relevant assumptions. Different designs and standardization conventions require different conversions. Do not use an independent-groups formula for paired observations without checking its definition.
Does a large effect size mean a statistically significant result?
Not necessarily. Effect magnitude, estimation precision and test results are related but distinct. A large estimate based on imprecise data may have considerable uncertainty.
Should I report confidence intervals for effect sizes?
Where methodologically appropriate and feasible, yes. Use an interval procedure suited to the chosen estimator and design, and explain limitations. Never fabricate a confidence interval.
Is r squared always the percentage of variance explained?
For simple ordinary least-squares linear regression with an intercept and one predictor, computed on the same observations, Pearson's r² equals the unadjusted R². This identity does not license causal interpretation or automatically carry over to every regression model.
Is partial eta-squared better than eta-squared?
Neither is universally better. They use different denominators and communicate different aspects of an ANOVA model. Follow the model and disciplinary reporting requirements, label the estimate accurately, and do not compare non-equivalent values casually.
Final effect-size checklist
- Have I identified my actual outcome, research question, and statistical design?
- Is the effect-size definition appropriate for independent, paired, correlational or model-based data?
- Can I reproduce every input and calculation from the analysis output?
- Have I clearly labeled the sign, original units, standardized metric, and group/reference order?
- Have I included relevant sample sizes, uncertainty and limitations without fabricating results?
- Have I avoided treating significance thresholds as universal measures of importance?
- Are tables, narrative, metadata and conclusions numerically consistent?
- Does my interpretation distinguish association from causation?
Final thoughts
An effect size can make a dissertation's quantitative findings more informative, but it is not a magic label that turns numbers into practical conclusions. Choose a measure that fits the analysis, explain the formula or software definition, connect it to meaningful original units, and report uncertainty and design limitations honestly.
If you would like structured feedback on the clarity, organization, and presentation of your own dissertation results, Acadelyra's academic-support services can help you strengthen explanations and reporting while keeping your analyses, research decisions and academic responsibilities your own.
Sources and further reading
- Lakens, D. (2013). Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and ANOVAs. Frontiers in Psychology, 4, 863. Full open-access paper. DOI: 10.3389/fpsyg.2013.00863.
- American Psychological Association. APA Style Journal Article Reporting Standards. Reporting guidance for empirical studies.
- American Psychological Association. Guidelines for Reviewing Manuscripts for the Journal of Counseling Psychology. Guidance includes reporting suitable effect sizes with precision estimates.