ACADEMIC WRITING GUIDE
How to Conduct a Power Analysis for a Dissertation (With Worked Examples)
A dissertation can have a well-written research question and a carefully chosen statistical test yet still produce imprecise findings if the study is too small. A power analysis helps you think about sample size before collecting data. It asks a practical question: If an effect of a specified, meaningful size is real, how likely is my planned study to detect it under stated assumptions?
This guide shows how to plan an a priori power analysis, calculate an illustrative two-group sample size, explore alternative assumptions, and write an honest methodology paragraph. It also explains when a power calculation is not the right way to justify a sample. Every numerical scenario below is hypothetical and is not a recommendation that all dissertations use the same sample size.
What is statistical power?
Statistical power is the probability that a specified statistical test rejects its null hypothesis when a particular alternative is true, under the assumptions of the analysis. Power is often written as 1 minus beta, where beta is the probability of a Type II error for that particular alternative. For example, 80% planned power means that, in repeated studies conducted under the assumed effect and design, about 80% would reject the null at the chosen significance threshold. It does not mean that there is an 80% probability that your hypothesis is true.
Four inputs usually matter most: the effect size you want to detect, the significance threshold (alpha), the desired power, and the sample size. Other features matter too, including outcome variability, one- versus two-sided testing, group allocation, correlation between repeated measurements, clustering, missing data, and model structure. Penn State's explanation of power analysis provides the statistical foundation for these relationships.
| Input | Meaning | Question to answer |
|---|---|---|
| Effect of interest | The difference or association the study should detect | What change is substantively meaningful? |
| Alpha | Chosen Type I error probability for the specified test | Why use this significance threshold? |
| Target power | Detection probability at the assumed effect | What chance of detecting that effect is acceptable? |
| Sample size | Number of usable observations or participants | Can the study obtain enough complete data? |
An 80% power target and 5% alpha are common teaching conventions, not mandatory rules. Different contexts justify different choices.
Power analysis is not the same as choosing a sampling method
Power analysis addresses how many usable observations a specified statistical analysis may require. It does not determine which people should be recruited, whether the sampling frame is complete, or whether results generalise to a population. A large convenience sample may still contain systematic selection bias.
If you have not established your population, recruitment plan and eligibility criteria, first review how to choose a sampling method and sample size. Power calculations belong alongside those decisions, not in place of them.
Likewise, a qualitative interview study ordinarily does not justify its sample through hypothesis-test power. Qualitative sample adequacy follows from its design, research question, analytic approach and depth of evidence.
Step 1: Identify the primary research question and planned analysis
Begin with the main comparison or parameter, not a convenient sample-size formula. Consider this hypothetical question:
Among first-year university students, does a structured study-skills workshop lead to a different average end-of-term study-confidence score than the usual orientation programme?
Suppose students would be assigned independently to two groups and the prespecified primary outcome is a continuous score. An independent two-group mean comparison could be a possible planning model. But if students are assigned by classroom, the observations may be clustered; if the same individuals are measured repeatedly, the correlation structure changes; and if baseline scores will be adjusted in a regression model, the appropriate power procedure may differ.
Specify your actual analysis before choosing a calculator. Our guide to writing a dissertation data analysis plan helps connect research questions, variables and statistical methods.
Step 2: Choose a defensible effect size
The effect size in a power analysis is not a forecast of what you hope the results will show. It is the effect magnitude against which you want the study to be informative.
For the hypothetical workshop, imagine that a six-point difference on the confidence scale would be educationally meaningful. Assume a common outcome standard deviation of 12 points. The standardised difference for planning is then six divided by twelve, or 0.50. This ratio describes the planning assumption, not a measured result.
A sound justification may draw on comparable prior studies, pilot estimates with suitable uncertainty, scale interpretation, or a minimum effect that would change a practical decision. UCLA's introduction to power analysis discusses the tradeoffs involved in choosing assumptions and why simply borrowing a convenient effect-size benchmark can be misleading.
For more detail on interpreting standardised differences, consult our effect-size calculation guide.
Step 3: Set alpha, power and test direction before collecting data
For our worked illustration, assume two-sided alpha = 0.05 and target power = 0.80. A two-sided test permits a difference in either direction. We will use normal-approximation critical values of 1.96 for the two-sided 5% threshold and 0.84 for 80% power.
These values help explain the arithmetic but do not replace the appropriate finite-sample or exact software calculation. In particular, a two-sample t test with an unknown standard deviation uses a noncentral t distribution for accurate prospective power planning, and the final requirement may differ from the approximation below.
Worked example 1: Approximate sample size for two independent groups
Our hypothetical planning inputs are two equal-sized independent groups, a mean difference of 6 points, assumed standard deviation of 12 points, two-sided alpha of 0.05 and desired power of 80%.
For a simplified two-group comparison with a common standard deviation, a normal-approximation planning equation is:
n per group ≈ 2 × SD² × (z-alpha/2 + z-power)² / difference²
Substitute the teaching values:
n ≈ 2 × 12² × (1.96 + 0.84)² / 6²
Because 12² = 144, (1.96 + 0.84)² = 7.84 and 6² = 36:
n ≈ (2 × 144 × 7.84) / 36
The approximate required sample is 62.72 participants per group. Always round a required participant count up, giving 63 usable participants per group, or 126 usable participants total for this simple approximation.
| Parameter | Assumption | Role |
|---|---|---|
| Difference worth detecting | 6 points | Numerator of practical rationale; squared in denominator of formula |
| Common standard deviation | 12 points | Captures anticipated outcome variability |
| Two-sided alpha | 0.05 | Corresponds to normal critical value 1.96 |
| Desired power | 0.80 | Corresponds to approximate normal critical value 0.84 |
| Required usable sample | 63 per group | Approximation rounded up, not exact t-test output |
| Total usable sample | 126 | Equal allocation across two groups |
Interpretation: Under these assumptions, a sample near this size is a starting point for planning the stated comparison. It does not ensure a statistically significant result, prove the workshop works, or establish that students were recruited without bias. A software calculation for the exact design is the appropriate final check.
Worked example 2: Allow for nonresponse or attrition
The planned analysis requires 63 usable participants in each group. Suppose, for illustration, that relevant recruitment experience supports anticipating a 20% loss before the primary outcome is available. The expected usable fraction is 0.80.
Divide the usable target by the expected usable fraction:
63 / 0.80 = 78.75, so the recruitment target becomes 79 per group. That gives 158 participants in total to recruit or enrol under the assumed loss rate.
| Measure | Per group | Across both groups |
|---|---|---|
| Approximate usable target | 63 | 126 |
| Assumed outcome-data loss | 20% | 20% |
| Recruitment target after rounding | 79 | 158 |
This is an allowance, not a guarantee. The eventual completed sample could differ. The 20% assumption must come from relevant evidence or be labelled as a hypothetical planning scenario. If missingness is related to participant outcomes, increasing invitations alone does not solve the resulting bias. Our guide to handling missing survey data explains this distinction.
Worked example 3: Sensitivity analysis when the effect is uncertain
The six-point target may be plausible, but perhaps a smaller four-point difference would also matter, or only an eight-point difference would justify implementing the workshop. Holding the simplified assumptions and variability constant shows how sample size changes.
For a four-point difference, the approximation gives 141.12 participants per group, rounded up to 142. For an eight-point difference, the approximation gives 35.28 participants per group, rounded up to 36.
| Meaningful difference | Approximate usable sample per group | Approximate usable total |
|---|---|---|
| 4 points | 142 | 284 |
| 6 points | 63 | 126 |
| 8 points | 36 | 72 |
This table exposes an important risk: choosing a conveniently large effect can make an under-resourced study appear adequate. If the academically meaningful effect is four points, planning exclusively around eight points would leave the study less able to detect the smaller effect.
Sensitivity analysis should therefore be part of the justification, especially when earlier evidence is limited or imprecise.
Worked example 4: Precision-based planning is a different question
Not every dissertation is primarily testing a difference or association. Suppose the objective is to estimate the proportion of students using a university support service with a desired 95% confidence interval margin of error of approximately five percentage points.
For an illustrative simple random sample from a large population, with an unknown proportion conservatively set to 0.50, the elementary normal-approximation formula is:
n ≈ z² × p × (1 − p) / margin²
Using z = 1.96, p = 0.50 and margin = 0.05:
The preliminary precision-based sample size is 384.16, rounded up to 385.
This is a preliminary large-population calculation under simple-random-sampling assumptions; nonresponse, finite-population correction where justified, clustering, weighting, subgroup needs and better interval methods may alter the actual design. Importantly, this is not an 80%-power calculation. It answers a precision question rather than a hypothesis-detection question. See the University of Colorado's sample-size and power resource for the distinction between planning estimators and statistical tests.
How to conduct a power analysis using software
Once you've defined the model, use software that matches the primary planned analysis. For example, G*Power supports various test families, while specialised packages may be necessary for clustered, longitudinal, survival, multilevel or other complex designs. A published review of sample-size planning with G*Power explains its key inputs and common test families.
A defensible software workflow is:
Specify the actual test family and statistical test; do not substitute a similar-sounding procedure.
Choose an a priori sample-size calculation rather than a retrospective observed-power calculation if your question concerns recruitment planning.
Enter the effect-size assumption with its justification, alpha, target power, direction and allocation ratio.
Set any additional design parameters, such as repeated-measures correlations or number of groups, only when they reflect your design.
Record the software, version, options, calculated usable sample size and any rounding.
If necessary, adjust a recruitment target separately for anticipated missing outcomes, attrition or ineligibility.
Vary the uncertain assumptions to show whether the proposed study remains feasible.
Do not report a software-derived exact sample size unless you actually ran that procedure. Our simple numerical example above should never be relabelled as a verified G*Power result.
How to write the power analysis in your dissertation
A useful methodology paragraph should communicate the question, planned test, assumptions, source of assumptions, usable sample, and recruitment adjustment. It should distinguish the calculated minimum from what you actually recruited and analysed.
Hypothetical proposal paragraph
The proposed study will compare average end-of-term study-confidence scores between two independently assigned groups. For an illustrative preliminary calculation, a six-point difference and common standard deviation of 12 points were assumed. Using a two-sided significance threshold of 0.05 and target power of 0.80, the normal approximation indicated approximately 63 usable participants per group (126 total). Because this is not an exact finite-sample power calculation, the final recruitment plan will be confirmed using software appropriate to the prespecified analysis. A separately justified allowance for anticipated outcome-data loss will be incorporated before recruitment begins.
Hypothetical completed-study reporting paragraph
The proposal-stage sample-size justification targeted 63 usable participants per group under its stated preliminary assumptions. The methods section reports the final software-based calculation and recruitment target. The results section separately reports the number invited, enrolled, retained and included in each analysis, together with relevant missing-data handling and departures from the original plan.
The second passage is deliberately a reporting framework rather than invented results. Replace it with the actual study figures, dates and approved procedures.
Common mistakes to avoid
Using an unrelated formula. A simple two-group equation does not automatically apply to repeated measures, clustered recruitment, logistic regression, interactions or multilevel models.
Picking an effect because it produces a small sample. Start with what would be meaningful and defensible, not what fits your budget.
Confusing an approximate normal calculation with exact software power. Show the assumptions and use a model-specific calculation for the final design.
Treating nonsignificance as proof of no effect. A study may be insufficiently informative about a clinically or educationally relevant difference. Discuss estimated effects and uncertainty, not just whether a p-value crosses a threshold.
Calling a sample representative because it is large. Power does not repair selection bias, missing coverage or self-selection.
Ignoring multiple primary tests. A design with multiple primary outcomes or comparisons may require a justified multiplicity strategy; an unadjusted calculation may overstate the effective power for the intended claims.
Running observed power using the obtained p-value and calling it a planning justification. A priori decisions belong before analysis, and post-hoc observed-power figures derived from estimated effects often add little information beyond the estimates and intervals themselves.
Claiming that 80% power guarantees success. Even when the assumptions are correct, results vary from sample to sample.
For complementary guidance on hypothesis formulation, see how to write a clear and testable research hypothesis.
Power-analysis checklist
Is the research question linked to one primary planned analysis or estimation target?
Is the test direction and significance threshold stated and justified?
Does the effect-size assumption represent a meaningful difference or association?
Is the variability estimate supported by relevant evidence?
Have dependence, clustering, unequal allocation or repeated measurements been considered?
Is the selected software procedure appropriate for the design?
Are usable sample size and recruitment target clearly distinguished?
Is any attrition or nonresponse assumption supported and transparently reported?
Have uncertain inputs been examined through sensitivity analysis?
Are planned figures kept separate from actual recruitment and analysis counts?
Frequently asked questions
Is 80% power mandatory for a dissertation?
No. It is a common benchmark, not a universal requirement. Choose and justify power in light of the discipline, consequences of missed effects, resources and institutional expectations.
Does a larger sample automatically improve validity?
No. It can improve precision or power under the assumed design, but it cannot independently eliminate selection bias, poor measures or confounding.
Can I calculate power after collecting data?
You can evaluate what effect sizes your realised design could detect under specified assumptions, or discuss precision and confidence intervals. However, using the observed effect and p-value to calculate retrospective observed power does not substitute for an honest a priori design rationale.
Can I use this two-group formula for my regression dissertation?
Not without checking the model. Regression power depends on the parameter of interest, predictor distribution, correlations, covariates and other assumptions. Use a procedure designed for your planned model.
What if I cannot recruit the calculated sample?
Report the limitation transparently. Reconsider the design, seek appropriate methodological guidance and assess what estimates or effects the feasible sample could meaningfully address. Do not quietly change your original assumptions just to make the numbers fit.
Should qualitative researchers use G*Power?
Ordinarily not for a qualitative interview study whose primary goal is interpretive understanding rather than statistical hypothesis testing. Justify qualitative sample adequacy in a manner consistent with the chosen methodology.
Final thoughts
Power analysis helps researchers make their sample-size assumptions visible before collecting data. Its value is not the appearance of a precise number but the reasoning behind that number: the planned question, meaningful effect, variability, test, uncertainty and practical constraints.
If you need help reviewing the clarity of your dissertation methodology, sample-size justification or statistical reporting, Acadelyra's academic support services can help you identify areas to improve while keeping your research decisions, analyses and academic responsibilities your own.
Sources and further reading
Penn State — Power Analysis. Explains statistical power and simplified sample-size calculations.
UCLA Statistical Methods — Introduction to Power Analysis. Discusses assumptions, effect sizes and planning tradeoffs.
University of Colorado Boulder — Calculating Sample Sizes and Power for Research. Outlines prospective sample-size planning.
Sample size determination and power analysis using G*Power software. Peer-reviewed discussion of power calculations for different statistical test families.