Power Analysis Tool

Imagine throwing a net to catch fish. The bigger the fish (effect size), the easier it is to detect with a small net. But if the fish are small and you want to be sure you catch them, you need a much larger net — and even then, some will slip through. Power analysis is how you figure out how big a net you need before you start fishing.

Statistical power balances three competing pressures: how big an effect you expect, how willing you are to falsely declare an effect (alpha), and how willing you are to miss a real one (beta). Tighten your tolerance on any one of these and the sample size climbs. Relax one and costs fall, but the study becomes less trustworthy in exactly that dimension. There is no free parameter to adjust.

The calculation this tool runs is the standard formula for a two-sided independent-samples t-test. It is the most commonly required form in academic ethics applications, grant proposals, and clinical trial registration. Understanding the inputs is more useful than memorizing the formula, because the judgment calls — what effect size is plausible, what false-negative risk is acceptable for this decision — are yours, not the calculator's.

Use this tool when you are designing a study and need a defensible minimum enrollment target before seeking ethics approval, grant funding, or clinical trial registration. It is also useful when reviewing a completed study to assess whether a non-significant result reflects a true absence of effect or simply insufficient power to detect a real one.

This calculation is appropriate for continuous outcomes analyzed with a t-test — weight, blood pressure, test scores, reaction times. It does not directly apply to binary outcomes (proportions, event rates), survival analyses, or repeated-measures designs. For those designs, the sample size formula changes, and you would need a specialized calculation. If your outcome is dichotomous, the required sample can be substantially different from what this tool returns.

Do not use this result as your final answer when designing a multi-site clinical trial or a study where the regulatory submission standard is FDA or EMA-level documentation. Those contexts require a statistician to sign off on the power calculation and the assumptions behind it. This tool gives you the planning number and the right inputs to bring to that conversation — not a substitute for it.

Using the observed effect from a small pilot study as the planning effect. Pilot studies are run with small n precisely because the true effect is unknown. Small samples produce wildly variable estimates of effect size — a pilot of ten people might return d of 1.2 when the true population effect is 0.4. If you use the pilot estimate directly, you will size your full study on noise and end up with a dramatically underpowered design. Instead, take the lower confidence bound of the pilot estimate, or use a published meta-analytic estimate from related work.

Forgetting to convert between one-tailed and two-tailed tests. Some older software and textbooks present power tables for one-tailed tests. A one-tailed test at alpha 0.05 uses the same critical value as a two-tailed test at alpha 0.10. If you pull a sample size from a one-tailed table and run a two-tailed test (which most journals require by default), your study is underpowered by design. This tool uses two-tailed alpha throughout.

Neglecting attrition in any study with follow-up measurements. If your outcome is measured at six months and you expect 20% of participants to drop out, recruiting exactly the minimum sample means you finish with only 63 per group analyzable people per group instead of what you planned for — leaving the study powered for a larger effect than the one you designed to detect. Build the dropout buffer in before you recruit, not after you see the final data.

The required sample size per group for a two-sample t-test is derived from the non-central t-distribution. The working formula is: n per group equals two times the quantity (z-alpha over two plus z-beta) squared, divided by d squared. Here z-alpha is 1.96 for the standard alpha of 0.05, z-beta is 0.8416 for the standard power of 0.80, and d is the effect size entered as 0.5. Plugging in those values gives 63 per group per group and 126 participants total.

The formula uses the two-tailed normal quantile for alpha because the test is two-sided — you are looking for effects in either direction. The one-tailed quantile for beta reflects that the power calculation asks only whether the test will detect a departure from the null, not which direction it goes. Both quantiles come from a standard normal approximation, which holds well for the t-distribution at sample sizes returned by typical power calculations.

After computing the base n, this tool inflates by a factor of one divided by (one minus the dropout fraction) to produce the recruitment target. If you enter a 15% dropout rate, you divide by 0.85, which increases the number to recruit so that you still have the target number of completers for analysis. The ceiling function (rounding up to the next whole person) is applied at every step — you cannot recruit fractional participants.

The formula implemented here assumes equal group sizes and equal variances (the pooled t-test). Both assumptions are optimistic. When group sizes are unequal, the harmonic mean of the group sizes enters the formula, and total n for a given power increases — sometimes substantially if the imbalance is severe. When variances are unequal (Welch's t-test), the effective degrees of freedom shrink, and you need slightly more participants than this formula returns. For a conservative planning estimate with expected imbalance, compute the harmonic mean of your target allocation and substitute it for n-per-group.

The normal approximation to the t-distribution quantiles is exact only as n approaches infinity. For very small true n values — under about fifteen per group — the normal approximation understates the required sample, and you should use a direct non-central t-distribution solver. The practical consequence: if this tool returns a very small sample, treat the output as a lower bound and verify with software that implements the exact non-central t CDF.