What Is Statistical Power? (Research Specs)
Statistical power is the probability that a study will detect an effect of a specified size when that effect is real under the planned design and analysis. It equals 1−β. Researchers use it to plan sample size and judge whether a study can detect a meaningful effect. Power depends on assumptions, not a guarantee.
As each research cycle brings proposal deadlines, recruitment plans, and data collection into view, one question matters early: can the study detect the effect it aims to measure? Statistical power helps you answer that question before you spend time and funds collecting data. It is a planning tool, not a promise of a significant result.
Research Question, Study Design, and Power-Analysis Overview
Statistical power is conditional: it describes the chance of rejecting the null hypothesis if a specified alternative is true. Its value depends on the effect you expect, the study design, the chosen test, and the significance level. A power analysis is useful only when those assumptions match the research question.
Define the effect you need to detect
The effect size is a measure of how large a difference or relationship is. For planning, use the smallest effect that would matter in practice or theory, not simply the largest estimate found in a small earlier study. An optimistic assumption can make a study look adequately sized when it is not.
Choose an effect-size measure suited to the outcome and analysis. Examples include a mean difference in the outcome’s original units, a standardized mean difference, a correlation, or an odds ratio. State why that size is meaningful and where the estimate came from, such as prior research, a pilot study, or a field-based threshold.
A useful planning question is: “What result would change what researchers or practitioners do?” For example, a small change in a symptom score might matter in one setting but not another. The answer should guide the effect-size assumption before you calculate sample size.
Match the calculation to the study design
An a priori power analysis estimates the sample size needed before data collection begins. Specify the primary outcome, hypothesis, target population, study design, planned statistical test, alpha level, target power, and effect-size assumption. These details define the calculation.
Distinguish confirmatory analyses, which test planned hypotheses, from exploratory analyses, which look for patterns that can guide later research. If a study has several primary tests, plan how to control the risk of false positives across them. A power calculation for one test may not fit a set of tests.
Common planning targets are 80% power and, where missing an effect would be especially costly, 90% or higher. These are conventions, not universal rules. Likewise, α=.05 is common for a single primary test, but the right choice depends on the field, consequences of errors, and any multiple-testing plan.
| Planning choice | What to specify | Why it matters |
|---|---|---|
| Effect size | Smallest meaningful effect and its rationale | Sets the effect the study aims to detect |
| Alpha (α) | Type I error rate for planned tests | Sets the false-positive threshold |
| Target power | Often 80% or 90%, with justification | Sets the desired chance of detection |
| Test and design | For example, two-group comparison or regression | Determines how assumptions translate into sample size |
| Planned sample | Required total and, if relevant, group sizes | Shows whether the design is feasible |
Statistical Assumptions, Sample Size, and Analysis Requirements
Sample size is the number of participants, observations, or units the study needs for its planned analysis. There is no single number that ensures adequate power across all studies. The estimate depends on the outcome, effect size, variability, design, test, and error-rate choices.
Account for design features that change the required sample
A calculation for a simple independent-groups test may not fit a clustered or repeated-measures study. In a clustered design, people within the same school, clinic, or worksite may be more alike than people in different clusters. That similarity reduces the amount of independent information in the data.
One common measure is the intraclass correlation coefficient (ICC), which describes how similar observations are within a cluster. For equal cluster sizes, a basic design effect is 1 + (m−1)ICC, where m is the average number of observations per cluster. This can help illustrate why a clustered study often needs more participants than a simple random-sample calculation suggests. Use a method suited to the actual design for the final analysis.
Repeated measures, unequal group sizes, covariates, and planned subgroup analyses can also change power. So can missing data and participant dropout. For attrition, a basic recruitment adjustment is: required completed sample ÷ (1 − expected attrition rate). This adjustment accounts for losses, but it does not replace a design-specific power analysis.
Use an appropriate method and check feasibility
Use a validated tool or statistical method that matches the planned model and test. Options include dedicated power-analysis software and statistical packages. Check that the tool handles your design features, effect-size metric, and test direction. If you are unsure, ask a statistician or methods adviser to review the inputs.
The output should include a sample size, not just a power percentage. Record whether the number refers to the total sample or each group, and whether it means people recruited or participants with usable data. If recruitment limits make the required sample unrealistic, revisit the design transparently. Options may include improving measurement precision, using a more efficient design, or narrowing the research question. Do not quietly lower the target or assume a larger effect to make the calculation fit.
Illustrative study-planning scenarios
The examples below are hypothetical. They show how the same power target can lead to different sample requirements when the research question and design change. No sample-size number is given because each case needs specific effect-size and variability assumptions.
| Scenario | Key power-analysis issue | Planning response |
|---|---|---|
| Two groups receive different interventions | Meaningful group difference and allocation ratio | Define the smallest meaningful difference, then calculate for the planned test |
| Several clinics recruit participants | Similarity within each clinic | Include clustering and the expected number and size of clinics |
| Participants are measured more than once | Within-person change and correlation across measurements | Use a repeated-measures method that matches the planned analysis |
| Several primary outcomes are tested | Greater chance of at least one false positive | Set a multiplicity plan and account for it in power calculations |
Research Execution, Data Quality, and Interpretation
A power analysis supports a study plan; it does not control how data are collected or analyzed. Recruitment, measurement quality, and adherence to the planned analysis still matter. Report the assumptions and actual sample so readers can assess how closely the completed study followed the design.
Follow the plan and document changes
Before recruitment, write down the primary outcome, planned sample, effect-size rationale, alpha, target power, analysis, and handling of missing data. Preregistering these choices, when appropriate for the project, can make it easier to distinguish planned tests from later exploratory work.
Monitor recruitment and data quality using a prespecified plan. If a change becomes necessary, document what changed and why. For example, if recruitment falls short, explain the final sample and any implications for precision. Avoid changing the primary outcome or analysis without a clear record.
A common practical mistake is to treat every enrolled participant as usable data. Set out how you will handle incomplete records, exclusions, or missed follow-ups, and make the power calculation reflect the sample needed for analysis. These choices should be consistent with the research design and ethical requirements.
Interpret results without overclaiming
A nonsignificant p-value does not prove that the null hypothesis is true. If the study had low power for a meaningful effect, the result may reflect limited information rather than the absence of an effect. Even with good planning, estimates can be uncertain.
For completed research, emphasize the effect estimate and its confidence interval. The estimate shows the direction and size of the observed result; the interval communicates its uncertainty. Do not use observed, or “post hoc,” power calculated from the study’s observed effect and p-value as proof that a null result is conclusive. It adds little beyond those results.
In reporting, include the power-analysis method, assumptions, planned and achieved sample sizes, and the primary analysis. If the study differs from the plan, say so. Clear reporting helps readers judge what the findings support and what remains uncertain.
Evidence-Based Research Practice and Next Steps
Power analysis is part of responsible research practice, not a box to tick. It connects the research question to the resources and design needed to answer it. A clear rationale also helps advisers, reviewers, and research partners understand why the chosen sample is suitable.
I find it useful to treat the calculation as a chain of decisions: define the meaningful effect, choose the design and test, set error rates, then estimate the sample. If one link changes, revisit the others. A different primary outcome, for example, may require a different effect-size measure and calculation.
Use this short checklist before finalizing a study plan:
- Identify the primary outcome and hypothesis.
- Define the smallest effect that would be meaningful.
- Choose the test and account for design features.
- Justify alpha and target power for the research context.
- Estimate the required sample, including expected data loss.
- Check whether recruitment and resources can support that sample.
- Record the assumptions and planned analysis.
If the required sample is not feasible, seek methods advice before data collection. A revised design may still answer a focused question, but the limits should be clear. The goal is not to guarantee significance. It is to give the study a defensible chance to detect an effect worth detecting.
Frequently asked questions
What does 80% statistical power mean?
It means that, if the specified effect is real and the analysis assumptions hold, the planned study has an 80% probability of rejecting the null hypothesis. It does not mean there is an 80% chance that the hypothesis is true or that the study will produce a significant result.
Is 80% power required for every study?
No. It is a common planning target, not a universal rule. Some fields or decisions may call for a higher target, while feasibility and ethical limits may shape other designs. Justify the target using the consequences of missed effects and relevant disciplinary guidance.
What is the difference between power and significance level?
Power is the chance of detecting a specified effect when it is real. The significance level, alpha, is the planned probability of a Type I error: rejecting the null hypothesis when it is true. Both are planning choices, and they address different risks.
How do I choose an effect size for a power analysis?
Choose the smallest effect that would matter for the research question. Explain the basis, such as prior evidence, a meaningful threshold, or a justified pilot estimate. Avoid relying only on a large effect from a small or selectively reported study.
Does a larger sample always increase power?
For a given design and analysis, adding usable observations generally improves the chance of detecting a specified effect and can improve precision. But more participants cannot fix every issue, such as poor measurement or a mismatched statistical model.
How do I account for participant dropout?
Estimate the number of usable cases required, then increase recruitment to allow for expected losses. A basic adjustment is required completed sample divided by one minus the expected attrition rate. For complex designs, also check whether dropout could bias results.
Can a nonsignificant result show that there is no effect?
No. A nonsignificant result alone does not establish that no meaningful effect exists. Consider the effect estimate, confidence interval, sample size, design, and uncertainty. A study may be unable to distinguish a small effect from no effect.
Should I report observed or post hoc power?
Do not use observed power to claim that a completed study had adequate power or that a null result is definitive. Report the planned power analysis and actual sample instead, then interpret effect estimates and confidence intervals in light of the study’s design.
(This article was written by one of our staff writers, Nathan Sterling. Visit our Meet the Team page.)
