Most students pick their sample size the same way: they copy a number from a previous study, ask their supervisor for a round figure, or collect as many responses as they can before the deadline. Then, months later, they run their analysis and get a non-significant result — and they have no idea whether their hypothesis was wrong or their sample was simply too small to detect the effect.
A power analysis answers that question before you collect a single response. It tells you how many participants you need to give your study a fair chance of finding an effect that is really there. In this tutorial, you will learn what statistical power means, the four ingredients every power analysis needs, and how to run one in SPSS step by step.
What Is Statistical Power, in Plain English
Statistical power is the probability that your test will detect an effect when one genuinely exists. It is usually written as 1 minus beta, where beta is the probability of a Type II error — missing a real effect.
Think of it this way. Imagine a new teaching method really does improve exam scores, but only by a modest amount. If you test it on 12 students, random variation between students can easily swamp that modest improvement, and your test will probably come back non-significant. You would conclude, wrongly, that the method does not work. With 120 students, the same real improvement stands out clearly against the background noise.
Power is conventionally set at .80, meaning an 80% chance of detecting a real effect. That .80 figure is a widely used convention, not a law of nature, but it is the default in most fields and the value reviewers expect to see justified if you choose something different. A study with power of .50 is essentially a coin flip — and a study with very low power wastes participants’ time as well as your own.
Power matters for ethics as well as statistics. Collecting data from far more people than you need wastes their time; collecting far too few means their contribution was never likely to produce a trustworthy answer. A power analysis is how you justify your sample size to your supervisor, your ethics board, and your thesis examiners.
The Four Ingredients of a Power Analysis
Every power analysis balances four quantities. If you know any three, you can solve for the fourth:
1. Effect size
How big is the difference or relationship you expect? In SPSS power analysis, effect size is often expressed as Cohen’s d for t-tests (the standardized mean difference) or f for ANOVA. Small, medium, and large benchmarks exist — for Cohen’s d, roughly 0.2, 0.5, and 0.8 — but these are rough guides, not substitutes for thinking. The best source for an expected effect size is previous research on the same or a similar intervention: look at the means, standard deviations, and effect sizes reported in earlier studies, and be conservative. Overestimating the effect size is the most common way to end up with a sample that is too small, because larger assumed effects produce smaller required samples.
2. Significance level (alpha)
This is almost always .05 — the threshold you will use to call a result statistically significant. A stricter alpha, such as .01, requires a larger sample, because you are demanding stronger evidence before you declare an effect.
3. Desired power
Usually .80, as discussed above. Raising power to .90 gives you extra insurance but costs more participants — sometimes substantially more. The jump from .80 to .90 is worth considering when data collection is cheap or when missing a real effect would be costly.
4. Sample size
This is usually the unknown you are solving for. In a power analysis for planning a study, you enter the first three ingredients and SPSS calculates the number of participants you need.
The relationship between them is intuitive: bigger expected effects need fewer people, stricter significance levels need more people, and higher desired power needs more people. When any ingredient is uncertain, run the analysis under a few plausible values and see how sensitive the required sample is — that sensitivity check is often more informative than a single number.
A Priori vs Post Hoc Power Analysis
An a priori power analysis is done before data collection: it determines the sample size you need. This is the analysis you should do, and the one this tutorial covers.
A post hoc power analysis, computed after the study from the effect you actually observed, is widely discouraged. Once you know your result, post hoc power adds no new information — a non-significant result will always show low post hoc power, by arithmetic. If a reviewer asks for it, what they usually really want is a confidence interval for your effect size, or a sensitivity analysis showing what effect size your sample could have detected. Do the a priori analysis properly, and you will rarely need the post hoc kind.
How to Run a Power Analysis in SPSS (Step by Step)
SPSS Statistics version 27 and later includes built-in power analysis procedures. If you are using an older version, you can still follow the same logic in the free program G*Power, which asks for the same four ingredients; the steps below use the SPSS menus.
Before you start: decide on your test
Power analysis is test-specific. An independent-samples t-test, a paired-samples t-test, a one-way ANOVA, and a correlation each have their own procedure, because each has a different effect-size measure. Decide which test answers your research question first — if you are unsure, work that out before opening the power analysis dialog. The worked example below uses an independent-samples t-test, the most common case for two-group comparisons.
Step 1: Open the power analysis dialog
Click Analyze, then Power Analysis, then T Tests, then Independent-Samples T Test. A dialog opens with fields for the test parameters and the analysis you want to run.
Step 2: Choose “Estimate sample size”
In the dialog, set the analysis type to estimate the required sample size (rather than estimating power for a sample you already have). SPSS will then solve for N.
Step 3: Enter your values
- Test assumption: leave the default two-sided test unless you have a strong, pre-registered reason for a one-sided test. Two-sided tests are the standard expectation.
- Power value: enter 0.80.
- Significance level: enter 0.05.
- Effect size: enter your expected Cohen’s d. For this example, suppose previous studies of similar interventions found a medium effect, so enter 0.50.
- Group size ratio: leave this at 1 if you plan equal-sized groups, which is the most efficient design.
Step 4: Run the analysis and read the output
Click OK. SPSS produces a power analysis table and, optionally, a power-versus-sample-size plot. For a medium effect (d = 0.50), power of .80, and alpha of .05, the required total sample is around 128 participants — about 64 per group. Your exact figure will depend on the values you enter, but the structure of the output is the same: the table gives the required sample size per group and in total, and the plot shows how power rises steeply at first and then flattens out as you add participants.
Read the plot as well as the table. It shows you the trade-off: the first extra participants buy a lot of power, but beyond a certain point each additional participant buys very little. That curve is useful when you need to negotiate a realistic target with a supervisor — it shows exactly what you lose by cutting the sample, and what little you gain by inflating it.
Step 5: Do a sensitivity check
Repeat the analysis with a smaller effect size than you hope for — for example, d = 0.30 instead of 0.50. You will see the required sample grow sharply. If your study is only feasible at the optimistic effect size, say so openly in your proposal and consider whether the design can be strengthened, for instance by reducing measurement error or using a within-subjects design, which typically needs fewer participants for the same power.
A Worked Example
Suppose you are comparing exam anxiety between students who attend a workshop and a control group, using an independent-samples t-test. Previous research on similar workshops reports effect sizes around d = 0.45. You set alpha at .05 and power at .80, with equal group sizes.
Running the SPSS power analysis with those values gives a required sample of roughly 78 participants per group, or about 156 in total. If you can only recruit 100 students, the analysis tells you honestly what that costs: your power for a d = 0.45 effect drops to roughly .60, meaning a 40% chance of missing a real effect. At that point you have three honest options: recruit more participants, strengthen the intervention or measurement to justify a larger expected effect, or switch to a design with more power, such as a pre-test/post-test design analyzed with ANCOVA. What you should not do is run the study at 100 participants and pretend the shortfall does not matter.
What If You Cannot Reach the Required Sample Size?
This is common, especially in student research with hard deadlines. A power analysis does not magically produce participants, but it changes the conversation from guesswork to evidence:
- Report the shortfall openly. State the required sample, the sample you achieved, and the power that sample provides for your expected effect. Examiners respect that transparency far more than an unexplained round number.
- Reduce noise. Reliable measures, clear instructions, and careful data screening all increase power without adding participants, because they shrink the error variance your effect must stand out against.
- Consider a more powerful design. Within-subjects designs, covariates in ANCOVA, and pre-registered directional hypotheses (used sparingly and justified in advance) can all raise power.
- Do not keep collecting data until the result becomes significant. That practice, sometimes called optional stopping, inflates the false-positive rate and undermines the very logic of the significance test.
Common Mistakes in Power Analysis
- Basing the effect size on a single small pilot study. Small studies produce noisy effect estimates, and the largest ones get published, so pilot-based effect sizes tend to be inflated. Use meta-analyses or several previous studies where possible, and run a sensitivity analysis.
- Using a one-sided test to shrink the required sample without justification. Reviewers will notice, and it is not a legitimate shortcut.
- Confusing statistical significance with practical importance. A huge sample can make a trivial difference significant; a power analysis tells you about detection, not about whether the effect matters. Always interpret the size of the effect alongside its p-value.
- Forgetting attrition. If you expect 15% of participants to drop out, inflate your target: divide the required sample by 0.85 so the final analyzed sample still meets the target.
- Running the analysis after the data are in and calling it a priori. Do it before recruitment, and record the date and the values you used.
Reporting in APA Style (7th ed.)
Report the power analysis in your Method section, under Participants or Design, so readers can judge whether your sample was adequate. Include the test, the assumed effect size and its source, alpha, desired power, and the resulting target sample — and whether you met it. For example:
“An a priori power analysis (SPSS Statistics, Power Analysis procedure) for an independent-samples t-test indicated that 128 participants (64 per group) were required to detect a medium effect (d = 0.50) with power of .80 at alpha = .05 (two-tailed). A total of 134 participants were recruited, of whom 130 provided complete data.”
If you fell short, report that too, with the achieved power for the expected effect size. Transparency about power is a strength of a study, not an admission of failure — it shows the reader exactly how much weight the results can bear.
Key Takeaways
- Power is the probability of detecting a real effect; .80 is the conventional target.
- Every power analysis balances effect size, alpha, desired power, and sample size — know three to solve for the fourth.
- In SPSS (version 27 and later), use Analyze, then Power Analysis, then the procedure matching your planned test.
- Base your effect size on previous research, be conservative, and always run a sensitivity check with a smaller effect.
- Inflate your target for expected dropout, and report the analysis — and any shortfall — openly in your Method section.
FAQs
Do I need a power analysis for every study?
For any study that tests a hypothesis with inferential statistics, yes — it is the standard justification for your sample size, and reviewers and examiners increasingly expect it. For purely descriptive or exploratory pilot work, a precision-based justification (for example, the width of the confidence interval you can achieve) may be more appropriate; state which approach you used and why.
What if I have no previous research to estimate an effect size from?
Use the conventional benchmarks as a starting point — a medium effect (d = 0.50) is a common default — but be explicit that it is a convention, not an empirical estimate. Run the analysis at small, medium, and large effects, and consider a pilot study to get a better estimate before committing to a full sample.
Can I do a power analysis in older versions of SPSS?
The built-in Power Analysis menu arrived in SPSS Statistics version 27. In earlier versions, the free program G*Power performs the same calculations and asks for the same four ingredients, so the logic in this tutorial transfers directly.
Is a bigger sample always better?
Up to a point, yes — but with diminishing returns. The power curve flattens as sample size grows, so beyond the required sample each extra participant adds little power while adding cost and time. An excessively large sample can also make trivially small effects statistically significant, which is why you should interpret effect sizes, not just p-values.
What should I do if my supervisor says 30 participants per group is enough?
Thirty per group is a rule of thumb sometimes quoted for the central limit theorem, not a substitute for a power analysis. Run the analysis with your expected effect size and show your supervisor the numbers: if 30 per group gives power of only .50 for a realistic effect, the analysis makes the case for a larger sample better than any argument can.
Need this done on your data? Get a free quote → /services/