Psychology dissertation power analysis turns a vague target for participant numbers into a transparent argument about what a study needs to learn. It links the research question, statistical model, effect worth detecting, error rates, and feasible sample. Done well, it helps you design an informative project and explain its limits honestly. It does not guarantee a significant result.
This guide explains how to plan, check, and report power for common psychology dissertations. It covers experiments, correlations, regression, interactions, mediation, repeated measures, and designs constrained by time or access. The emphasis is on defensible decisions rather than copying a conventional sample size.
What a Psychology Dissertation Power Analysis means
Statistical power is the long-run probability that a statistical test will reject its null hypothesis when a specified alternative is true. In ordinary null-hypothesis testing, it depends on several connected choices: the assumed effect, sample size, alpha level, test direction, allocation between groups, and the analysis model.
Power is usually written as one minus beta. Beta is the probability of a false negative for the particular effect and model used in the calculation. Alpha controls the probability of a false positive when the null model is true. These probabilities describe repeated use of a procedure under stated assumptions. They do not tell you the probability that an individual hypothesis is true.
A power result is therefore conditional. A statement such as “the study has 80% power” is incomplete unless it identifies the effect size, statistical test, alpha level, sample size, and other design assumptions. If those inputs are unrealistic, the calculation can be precise but unhelpful.
| Input | Question it answers | Why it matters |
|---|---|---|
| Effect size | What population difference or association must the design detect? | Smaller target effects normally require more information. |
| Alpha | What long-run false-positive rate is accepted for the planned test? | A stricter threshold generally increases the required sample. |
| Power | What probability of detecting the target effect is required? | Higher desired power generally increases the required sample. |
| Design and test | How will observations, predictors, groups, and repeated measures be analysed? | The same participant count can provide different information in different designs. |
| Sample size | How many analysable observations are available or required? | This may be the output of an a priori analysis or an input to sensitivity analysis. |
Table of Contents
Begin with the inferential goal
Sample-size planning should follow the question, not precede it. First decide whether your main goal is to test a directional hypothesis, estimate a parameter precisely, compare plausible models, or describe a defined population. A conventional power calculation is most directly suited to testing a prespecified effect with a frequentist test.
If the central goal is estimation, confidence-interval width may be more informative than power alone. If access to a rare population fixes the maximum sample, a sensitivity analysis can show which effects the design can detect with chosen error rates. If the model includes indirect effects, clustered observations, unequal group sizes, or complex missingness, simulation may represent the design better than a simple closed-form calculator.
Lakens describes several legitimate ways to justify sample size, including power, desired accuracy, resource constraints, and near-complete population measurement. The best choice depends on the inferential goal. A dissertation should make that link explicit rather than treating one method as universally correct.
Specify the primary claim
Write one sentence identifying the primary outcome, focal predictor or contrast, population, and planned analysis. For example: “The primary test will compare mean post-intervention anxiety scores between two independently allocated groups using a two-sided test, adjusting for baseline anxiety.” This is much more useful than saying the project will “study whether an intervention works.”
Your psychology dissertation hypotheses should match this primary claim. If several outcomes or contrasts are equally primary, account for multiplicity in the error-rate plan. Calling every available test primary after data collection undermines the meaning of the original calculation.
Choose the right kind of analysis
A priori power analysis
An a priori analysis calculates the sample required for a specified effect, alpha, power, and design before data collection. It is useful when recruitment can still be changed and a hypothesis test is the main inferential tool. Record the input values and software settings so another researcher can reproduce the result.
Sensitivity analysis
A sensitivity analysis starts with a feasible sample and asks what effect sizes the planned test can detect at selected power levels. It is especially useful for dissertations limited by a cohort, placement, clinic, school, or recruitment period. It does not make a weak design strong, but it makes the limitation interpretable.
Precision-based planning
Precision planning chooses a sample to achieve a desired confidence-interval width for a mean, difference, correlation, or other parameter. This approach fits research questions focused on how large an association is rather than whether a p value crosses a threshold. The expected variability and proposed interval width must still be justified.
Simulation-based planning
Simulation repeatedly generates data under an explicit model, fits the intended analysis, and estimates how often the decision rule succeeds. It is valuable for multilevel models, mediation, moderation, unequal clusters, non-normal outcomes, and other designs for which standard calculators simplify important features. Save the code, seed, distributions, parameter values, convergence rules, and success criterion.
Define an effect worth detecting
The most consequential input is often the target effect. Avoid selecting “small,” “medium,” or “large” solely from generic conventions. Those labels are not substitutes for subject knowledge, measurement scale, prior evidence, or the consequences of an effect.
A smallest effect size of interest is the smallest population effect that would matter for the theoretical or practical question. Its justification might use a clinically meaningful score difference, an educational decision threshold, a theoretically important change, or an effect below which two explanations are practically indistinguishable. Express it on the outcome scale where possible, then convert it to the standardised form required by the software.
Expected effects from published studies need cautious adjustment. Small studies often produce unstable estimates, and selective reporting can exaggerate the visible literature. A meta-analysis may help, but only if its populations, measures, interventions, and designs are sufficiently comparable. Consider the lower end of a plausible range rather than automatically using the pooled point estimate.
A psychology dissertation pilot study is usually better for testing recruitment, timing, randomisation, questionnaire completion, and data quality than for estimating a precise effect. If the pilot is small, its effect estimate can vary widely. Treat it as one uncertain input, not as unquestionable evidence.
Document the effect-size source
For every numerical input, record its source and transformation. If you convert a raw mean difference to Cohen’s d, document the standard deviation used. For a correlation, explain why the chosen value is meaningful in context. For regression, distinguish the incremental effect of a focal predictor from the total variance explained by the complete model.
It is good practice to calculate results across a plausible effect range. A curve or compact table often communicates uncertainty better than a single answer. If reasonable inputs imply sample sizes far beyond your resources, revise the question or design. Do not choose an inflated effect merely to make the desired sample feasible.
Match the calculation to the planned design
The statistical test in the calculation should match the primary analysis closely. A two-group calculation does not justify a multi-predictor model merely because both use the same participants. Likewise, power for a main effect says little about power for an interaction, which is often the harder target.
| Design | Focal quantity | Planning issues |
|---|---|---|
| Independent groups | Mean difference or standardised difference | Group allocation, variance equality, one-sided or two-sided test, covariate adjustment |
| Paired or repeated measures | Mean change or within-person contrast | Correlation between occasions, number and spacing of measurements, attrition |
| Correlation | Population correlation | Restricted range, measurement reliability, non-linearity, influential observations |
| Multiple regression | Incremental variance or focal coefficient | Number and correlation of predictors, covariate reliability, model specification |
| ANOVA | Omnibus effect or planned contrast | Number of groups, unequal allocation, contrast priority, multiplicity |
| Moderation | Interaction coefficient | Predictor distributions, reliability, scale, centring, plausible interaction size |
| Mediation | Indirect effect | Sizes of component paths, estimator, interval method, temporal ordering |
| Multilevel design | Fixed effect or variance component | Number of clusters, cluster sizes, intraclass correlation, random-effects structure |
Experiments and group comparisons
For a two-group experiment, specify whether allocation is equal, whether the test is one-sided or two-sided, and whether baseline scores will be adjusted. Unequal allocation generally reduces efficiency for a fixed total sample, though it can be sensible when one condition is expensive or when the eligible population is imbalanced. Plan the exact primary contrast rather than powering only an omnibus effect if that contrast answers the research question.
Our guide to psychology dissertation experimental design explains randomisation, manipulation checks, and validity decisions. For more than two groups, the psychology dissertation ANOVA guide covers omnibus tests and follow-up comparisons.
Correlation and regression
Correlation power depends on the population association you want to distinguish from the null, not on a correlation observed in the final sample. Consider restricted score ranges and imperfect reliability because both can attenuate observed relationships.
For multiple regression, determine whether the primary target is the overall model, a single coefficient, or the additional contribution of a block of variables. These are different tests. Power may be reduced when predictors are highly correlated because the unique contribution of one predictor becomes harder to estimate. See the dedicated psychology dissertation regression analysis guide for model checks and reporting.
Interactions, mediation, and complex models
Moderation tests whether an association changes across values of another variable. Interaction effects are commonly smaller and noisier than main effects, so a calculation based on the main effect can be seriously misleading. Specify predictor distributions and measurement reliability. The moderation analysis guide provides a focused workflow.
Mediation power depends on the joint behaviour of the component paths and the method used to estimate the indirect effect. A generic regression calculation cannot fully represent that distribution. Simulation is often the clearer approach. Our mediation analysis guide explains temporal assumptions and indirect-effect reporting.
Plan for analysable data, not only recruitment
A calculated sample normally refers to observations included in the primary analysis. Recruitment targets must therefore allow for plausible non-consent, withdrawal, incomplete data, failed attention checks, technical loss, and prespecified exclusions. Distinguish these stages rather than adding one unexplained percentage.
Suppose the primary analysis requires 180 complete cases and prior records suggest that about one in ten recruited participants will not provide analysable primary-outcome data. Dividing 180 by 0.90 gives a recruitment target of 200. This arithmetic is only an illustration. Your assumed loss rate should come from comparable studies, pilot logistics, platform records, or a justified conservative range.
Avoid treating deletion as the automatic response to missing values. Your missing-data strategy can change the information available and should be aligned with the planned model. Prevention is often more effective: reduce questionnaire burden, test reminders, make key measures easy to complete, and separate essential from exploratory measures.
For clustered data, participant count alone is not enough. Information depends strongly on the number of independent clusters and the similarity of observations within each cluster. Adding many pupils within only a few schools may contribute less information than recruiting additional schools. Use a calculation or simulation that represents the hierarchy.
Use power software transparently
G*Power is a free program for many common tests. The official Heinrich Heine University page provides the software and documentation. R packages and specialist applications can support precision calculations and simulation. The choice of tool matters less than whether the statistical model and inputs match the dissertation.
Do not report only a screenshot. State the program and version, analysis type, statistical test, effect-size measure and value, alpha, desired power, test direction, allocation ratio, predictors or groups, and resulting sample. For simulation, provide code and all generating assumptions. A future reader should be able to reproduce the answer without guessing which menus you selected.
A practical software workflow
- Choose the exact primary statistical test.
- Select a priori, sensitivity, precision, or simulation planning.
- Enter a justified target effect or a plausible range.
- Set alpha, desired power, test direction, and design parameters.
- Check that the output refers to total sample rather than sample per group.
- Adjust the recruitment target for expected loss without changing the analysable target.
- Repeat the analysis across plausible assumptions.
- Save the output, software version, and a written decision record.
Worked psychology examples
Example 1: a two-group intervention
A student plans to compare a brief sleep intervention with an information-only condition. The primary outcome is a validated sleep-quality score after the intervention, adjusted for baseline. The student defines a raw-score improvement that would be practically worthwhile, uses evidence from comparable populations to estimate residual variability, and converts this to the form required for the planned model.
The student calculates the analysable sample across a range of plausible effects, then increases the recruitment target for documented loss in similar online studies. The proposal identifies the primary contrast, two-sided alpha, desired power, allocation ratio, assumed baseline correlation, software, and version. It also states that if recruitment reaches only the feasible lower bound, interpretation will emphasise confidence intervals and the corresponding sensitivity analysis.
Example 2: a resource-limited correlation study
A placement allows access to at most 140 consenting staff members. Rather than claiming that 140 is “enough,” the student performs a sensitivity analysis for a two-sided correlation test at the prespecified alpha and power levels. They compare the detectable range with effects that would matter for the theory and with estimates from closely matched studies.
The result shows what the project can and cannot detect. The dissertation reports the finite access limit, expected response rate, measurement reliability, and planned handling of covariates. If the smallest meaningful correlation is below the design’s sensitivity, the student may simplify the question, seek another site, improve measurement, or present the work as estimation with appropriately cautious conclusions.
Common power-analysis mistakes
| Mistake | Why it fails | Better response |
|---|---|---|
| Using a generic medium effect | The label may not reflect the construct, scale, or decision. | Define a meaningful raw or standardised effect and justify it. |
| Powering a main effect when testing an interaction | The focal effect and sampling distribution differ. | Plan directly for the interaction, often using simulation. |
| Using an observed post hoc effect | Observed power largely restates the p value and adds little information. | Report the estimate, interval, assumptions, and prospective sensitivity. |
| Ignoring exclusions and attrition | The final analysable sample can fall below the planned target. | Separate analysable and recruitment targets with evidence-based loss assumptions. |
| Changing inputs to fit available resources | This conceals the design’s actual sensitivity. | Report the constraint and revise the question or interpretation. |
| Powering every exploratory test | The design becomes incoherent and multiplicity is hidden. | Identify one or a small number of primary tests and label exploration clearly. |
Post hoc “observed power” calculated from the effect estimate in the same dataset is particularly unhelpful. It is mathematically tied to the observed test result and does not rescue an inconclusive study. Confidence intervals, effect estimates, and a sensitivity analysis based on the achieved sample provide clearer information.
Report the decision in your proposal and dissertation
The method section should explain why the planned sample is informative for the primary goal. APA’s quantitative reporting standards provide a useful framework for transparent reporting. You can also register the decision before collection; the psychology dissertation preregistration guide explains how to separate confirmatory and exploratory choices.
A concise report might state: the primary test; target effect and its rationale; alpha; desired power; sidedness; design parameters; software and version; required analysable sample; expected loss and resulting recruitment target; and any sensitivity checks. If constraints set the sample, state the constraint first and report what effects the resulting design can detect.
After data collection, report the achieved sample and explain deviations from the plan. Do not silently replace the original calculation with one based on the observed effect. The psychology dissertation results-section guide can help you present estimates, intervals, and tests without overstating certainty.
Quality checklist before approval
- The primary outcome and statistical test are named.
- The target effect has a theoretical, practical, or evidence-based rationale.
- Alpha, desired power, sidedness, and multiplicity decisions are explicit.
- Group allocation, predictor count, repeated measures, or clustering are represented correctly.
- The analysable sample is distinguished from the recruitment target.
- Attrition and exclusions use a defensible estimate or range.
- Results are checked across plausible assumptions.
- The software, version, settings, and output are reproducible.
- Resource limits and design weaknesses are stated honestly.
- The interpretation does not treat power as proof of validity or significance.
Frequently asked questions
Is 80% power always enough for a psychology dissertation?
No. Eighty percent is a convention, not a universal scientific rule. Choose desired power by considering the costs of false negatives, available resources, the importance of the primary claim, and the precision you need. Report why your chosen level is appropriate.
Can I use a published effect size?
Yes, if the study is sufficiently comparable, but do not copy its estimate automatically. Check population, measures, design, analysis, uncertainty, and publication bias. A conservative plausible effect or a range is often more defensible than a single optimistic estimate.
What if I cannot recruit the calculated sample?
Report the constraint and conduct a sensitivity analysis. Then consider narrowing the primary question, improving measurement, simplifying the design, adding recruitment sites, or shifting emphasis toward estimation. Do not alter the assumed effect simply to obtain a convenient answer.
Should I calculate power after collecting data?
Do not use observed power based on the sample effect to interpret the same result. Report effect estimates and confidence intervals. If useful, state the design’s prospective sensitivity for meaningful effects using the achieved sample and prespecified test.
Does a larger sample fix poor research design?
No. More observations can improve precision and power, but they cannot repair invalid measurement, confounding, biased recruitment, careless manipulation, or an analysis that does not answer the question. Power planning belongs inside a broader design review.
Do qualitative dissertations need power analysis?
Usually not. Qualitative sample adequacy is judged using the methodology, information needs, diversity of relevant perspectives, and analytic depth. Do not import a hypothesis-testing calculation into a design with a different logic.
Conclusion
Psychology dissertation power analysis is strongest when it begins with a precise inferential goal and ends with an auditable decision. Define the primary test, choose a meaningful effect, represent the actual design, plan for analysable data, and show how conclusions change across plausible assumptions. If resources impose the sample, report the constraint and the study’s sensitivity rather than disguising it.
Use power as one part of ethical research planning. Recruiting too few participants can waste effort and expose people to burdens without answering the question, while recruiting more than needed also carries costs. If you want independent support, our psychology dissertation service can help you review your design and explain your own calculations transparently. It will not fabricate data, guarantee significance, or replace your supervisor’s requirements.
Authoritative references
- Lakens: Sample Size Justification
- Brysbaert: How Many Participants Do We Have to Include in Properly Powered Experiments?
- Perugini, Gallucci, and Costantini: A Practical Primer to Power Analysis
- Giner-Sorolla and colleagues: Power to Detect What?
- EQUATOR Network: DELTA2 target-difference guidance
- Heinrich Heine University: G*Power
- APA Style: Quantitative research reporting standards
