Sample Size & Power Calculator

Sample size is the number of people, observations, or data points in your study. If you survey 200 students, your sample size is 200. When your study compares groups, it helps to think per group as well: a treatment group and a control group with 60 people each give a total sample size of 120, but the per-group number (60) is what drives the calculation. This calculator asks for the number per group for independent t-tests and one-way ANOVA, and for the total number for paired t-tests, correlations, and chi-square tests.

Sample size decides what your study can and cannot detect. With too few participants, the study may miss a real effect, its estimates swing around with every data point, and any "significant" result it does find often exaggerates the true effect. Far more participants than needed wastes time and money, and can make trivial, meaningless differences look statistically significant. Choosing the right size in advance (with a power analysis) gives the study a good chance of detecting the effect you care about without spending more than the question requires.

Here, power means statistical power: the probability that your study will detect an effect that genuinely exists. For example, if a new teaching method truly improves test scores, power is the chance your study will find a statistically significant difference. It is calculated as 1 − β, where β (beta) is the probability of a false negative, meaning missing a real effect. So 80% power means a 20% chance of missing. Power is a property of the study design, set by the effect size, the significance level, the sample size, and the variability in the data, and you decide it before collecting anything.

It means that if a real effect genuinely exists, the study has an 80% chance of detecting it and producing a statistically significant result. Imagine running the exact same study 100 times: about 80 of those runs would find the effect, and about 20 would miss it. That 20% miss rate is the trade-off you accept, because chasing 100% power would require an impractically large sample. 80% is the widely accepted minimum in scientific research, clinical trials, and A/B testing, balancing the risk of missing a real effect against the cost of recruiting more participants.

They move together: as sample size increases, statistical power increases. Larger samples make the standard error smaller, which narrows the sampling distributions of your test statistic. With less overlap between the null and alternative distributions, a real effect becomes easier to spot, and the chance of a false negative (Type II error) falls. The gains are not in a straight line, though: power climbs quickly in small samples and levels off once samples are large, so doubling an already large sample buys very little extra power. The effect size matters too: detecting a small, subtle effect needs a much larger sample than detecting an obvious one.

80% is the standard target, and 90% is used for higher-stakes research such as clinical trials, where missing a real effect is more costly. Power is driven by four things, and a power analysis balances them: sample size (larger gives more power), effect size (bigger effects are easier to detect), significance level (usually 0.05), and variability (noisier data lowers power). A study below 50% power is worse than a coin flip at detecting a real effect, and underpowered studies are a common reason research fails to replicate.

A power calculation balances four values:

• Effect size: the size of the difference or relationship you want to detect (Cohen's d for t-tests, r for correlations, w for chi-square, f for ANOVA). Estimate it from prior studies, a pilot test, or the small/medium/large conventions.
• Significance level (α): the false-positive risk, almost always set at 0.05.
• Statistical power (1 − β): the chance of detecting a real effect, usually 0.80 or 0.90.
• Data variability and test: the expected standard deviation (or baseline rates) and the exact test you plan to run, such as a t-test, ANOVA, or chi-square.

Pick the test, set the values, and the calculator returns the smallest sample size that reaches your target power. Then add 10 to 20% for expected dropouts or missing data. This calculator handles t-tests, correlations, chi-square tests, and one-way ANOVA, and can also work backwards from a sample size to the power it achieves.

Four problems, and they compound. Low power: the study has a high chance of missing a real effect (a false negative), so a non-significant result means very little. Instability: with few data points, a couple of outliers can swing the average dramatically, and confidence intervals become so wide you cannot pin down the true value. Inflated results: significant effects that do appear in small studies often exaggerate the true effect size, one reason small studies fail to replicate. Poor generalizability: conclusions from a handful of participants rarely apply to the broader population, and random chance can produce striking "patterns" that do not exist.

Four drawbacks. Cost and time: recruiting, tracking, and cleaning more data takes more budget and effort, and the extra power gained flattens out as samples grow. Trivial significance: very large samples make tests so sensitive that tiny, meaningless differences become "statistically significant", which is why effect sizes matter alongside p-values. Bias is not fixed: a larger sample only reduces random noise; if the sampling method is flawed or unrepresentative, a bigger sample makes the biased answer more precise, not more correct. Ethics: in clinical research, enrolling more participants than necessary exposes extra people to risk or burden without adding proportional value.

It is the old rule of thumb that 30 observations per group is "large enough". It comes from the Central Limit Theorem: with about 30 observations, the distribution of a sample mean is close to a normal bell curve even when the underlying data is skewed, and the t-distribution becomes nearly identical to the normal z-distribution, so the standard tests behave as expected. It is a guideline, not a law: heavily skewed or outlier-ridden data can need far more than 30, and a clean, tightly controlled experiment can justify fewer. For a two-group comparison, roughly 30 per group gives about 80% power against a large effect (Cohen's d of 0.8), but against a medium or small effect it falls well short. The modern approach is to skip the folklore and run a power analysis for your specific test and effect size.

No, and the difference trips up a lot of people. A p-value is calculated from your data after the test: it says how surprising your result would be if there were no effect at all. Power is a property of the study design before you collect any data: the chance the study will detect an effect if one exists. The practical consequence shows up with non-significant results: in an underpowered study, "not significant" tells you almost nothing, because the study probably could not have detected the effect even if it were there. In a well-powered study, a non-significant result is reasonable evidence that any effect is small or absent. High power also does not make a p-value smaller or "more significant": power affects your chances of finding the effect, not how strong the evidence is once you have the data.