Is my study big enough to find what I’m looking for?

Often the sample is not really yours to choose — it is your cohort, your clinic, your funded window. This tells you what that sample can realistically detect, and whether a non-significant result would actually mean anything.

Achieved statistical power

Free guide: choosing between qualitative, quantitative and mixed methods

Power only matters once the design is right. This guide walks through choosing the tradition your question actually belongs to. We’ll email it to you.

One email with your download, then occasional research tips. One-click unsubscribe, anytime. We never sell your data.

Reading a power result honestly

Power is the probability that your study finds a real effect of the size you specify. At 80% power, one study in five misses an effect that genuinely exists. Below about 60%, a non-significant result tells you almost nothing: you cannot distinguish “there is no effect” from “my study could never have seen it”.

That distinction is the whole reason this matters for a thesis. An underpowered null result is not a finding; it is an inconclusive study, and a good examiner will say so. Knowing your power before you collect data lets you change the design. Knowing it afterwards at least lets you write an honest limitations section.

Two-sided normal approximations suitable for planning. Clustered, repeated-measures, multi-factor and survival designs need a formal analysis in G*Power, R or a statistician's hands.

Frequently asked questions

My committee asked for a power analysis. Which kind do they mean?

Almost always an a priori one: done before data collection, using an effect size justified from prior literature or a pilot, to show that the planned sample can answer the question. Present it as the four inputs and their sources, not just a final number.

Can I run this after I have collected my data?

You can, but be careful which effect you enter. Computing power from the effect you observed — so-called observed or post-hoc power — is circular: it is just a restatement of your p-value and reviewers increasingly reject it. What is legitimate is entering the effect you would have considered meaningful and reporting what your sample could have detected. That belongs in your limitations section.

My power came out at 45%. What do I do now?

If you have not yet collected data: increase the sample, switch to a paired or within-subject design, improve the reliability of your measure, or narrow the question to one you can actually answer. If collection is finished, shift the framing — report estimates with confidence intervals rather than significance tests, present the work as exploratory or hypothesis-generating, and state the limitation plainly. Examiners respect a candidate who names the constraint before they do.

What significance level should I use?

Two-sided 0.05 is the default across most fields and needs no justification. Use 0.01 if you are testing many outcomes and have not otherwise corrected for multiplicity. A one-sided test at 0.05 needs a real argument that an effect in the opposite direction would be meaningless to you, and reviewers are sceptical of it.

Does this cover ANOVA or regression?

No. It handles two-group and paired mean comparisons, two proportions, and correlations. For factorial ANOVA, multiple regression, mixed models or survival analysis, use G*Power or an R package such as pwr or simr — and budget time for it, because these analyses take longer than candidates expect.

Is anything I enter stored or uploaded?

No. Everything runs locally in your browser. Nothing is transmitted or saved.

How many participants do I need? Convert between effect sizes All free tools