Sample Size & Power Calculator

Statistical power is the probability that your study will detect a true effect if it exists. It's defined as 1 − β, where β is the Type II error rate (the chance of missing a real effect). For example, 80% power means that if there really is an effect of the size you're looking for, you have an 80% chance of finding it statistically significant and a 20% chance of missing it. Higher power is better, but requires larger sample sizes. Think of power as the 'sensitivity' of your study—high power means you're less likely to miss real findings.

A power of 0.80 (80%) is a widely accepted convention in statistics and research methods, balancing the desire to detect true effects with practical constraints on sample size and resources. It means you accept a 20% risk of Type II error (missing a true effect), which is considered reasonable in many fields. Some studies aim for 90% power (especially when missing an effect would be costly), but 80% is the standard for most homework, thesis proposals, and preliminary research. The 80% threshold comes from statistical tradition rather than a strict mathematical rule—it's a practical compromise.

α (alpha) and power control different types of errors. α is the significance level, typically 0.05, and represents the maximum probability of a Type I error (false positive—rejecting the null hypothesis when it's true). Power (1 − β) represents the probability of correctly rejecting the null hypothesis when the alternative is true—it controls Type II error (false negative). α is what you set before the study; power is what you achieve based on your sample size, effect size, and α. Think of α as your 'false alarm rate' and power as your 'detection rate.' Both are important, but they protect against different mistakes.

Effect size quantifies how big the difference or relationship is that you're trying to detect. For means, Cohen's d (standardized difference) is common: d = 0.2 (small), 0.5 (medium), 0.8 (large). For proportions, use the difference p₁ − p₂. For correlations, use the expected r value. To choose an effect size: (1) Look at similar prior studies and see what effects they found. (2) Use domain knowledge—experts often know what's a 'meaningful' difference. (3) Define the smallest effect you care about detecting (minimum meaningful difference). (4) Run pilot data for rough estimates. Be conservative: if in doubt, assume a smaller effect to avoid underpowering your study. Never pick a huge effect just to force a small sample size—it'll backfire if the true effect is smaller.

Increasing sample size (n) increases power, holding effect size and α constant. More data gives you a better chance of detecting a true effect. However, the relationship isn't linear—there are diminishing returns. For example, going from n = 20 to 40 per group might boost power from 50% to 80%, but going from 100 to 200 might only raise it from 95% to 99%. Eventually, adding more participants gives tiny power gains. The calculator helps you find the 'sweet spot' where you achieve adequate power (e.g., 80%) without over-recruiting. Very large samples can detect tiny, meaningless effects, so balance statistical power with practical significance.

One-tailed tests look for an effect in only one direction (e.g., 'treatment is better than control'), while two-tailed tests look for effects in either direction (e.g., 'treatment is different from control'). One-tailed tests require smaller sample sizes for the same power because they concentrate the α in one tail of the distribution. However, one-tailed tests are only appropriate when you have strong prior justification for directionality and are willing to ignore effects in the opposite direction. Most homework, research, and standard practice use two-tailed by default. Use one-tailed only if your instructor or field explicitly expects it and you can justify it.

No. This calculator is designed for education, homework, thesis planning, and preliminary exploration—not as the sole basis for high-stakes clinical trials, regulatory submissions, or business-critical experiments. Real-world studies often involve complexities (interim analyses, multiple endpoints, non-normal distributions, stratification, etc.) that require expert statistical guidance and specialized software (e.g., nQuery, PASS, G*Power). Use this calculator to learn power concepts, check homework answers, and explore trade-offs, but consult professional statisticians for real clinical trials, medical device studies, pharmaceutical research, or large-scale business experiments. Transparency: this is a learning tool, not a replacement for rigorous study design.

If the true variability (standard deviation σ) is larger than you assumed in your power calculation, your actual power will be lower than planned—you're more likely to miss the effect (Type II error). For example, if you planned for σ = 10 but the true σ = 15, your calculated n might give you 80% power in theory but only 60% in reality. To protect against this: (1) Use conservative (larger) estimates of σ based on pilot data or prior research. (2) Consider inflating your sample size by 10–20% as a buffer. (3) Report your assumptions transparently: 'Assuming σ = 10, we require n = 64. If σ is larger, power will be lower.' This acknowledges uncertainty and prepares readers for the possibility of null results.

No. Power analysis should be done before data collection (a priori), not after (post-hoc). Computing 'observed power' after finding a non-significant result is controversial and generally uninformative—it's directly tied to your p-value and doesn't add new information. Post-hoc power is often used incorrectly to excuse null results ('our power was low, so we can't conclude anything'), which is circular reasoning. Instead, report your results with confidence intervals and effect sizes, which convey both the estimated effect and its precision. If you're concerned about power after the fact, acknowledge it as a limitation and suggest that future studies with larger samples are needed. Always plan power prospectively.

Report power calculations clearly and transparently: (1) State the test type: 'We used two-sample t-test power analysis.' (2) List all parameters: α = 0.05, power = 0.80, effect size d = 0.5 (or specify means, SDs, proportions). (3) Cite justification for effect size: 'Based on Smith et al. (2020), we expect a medium effect.' (4) Report required sample size: 'n = 64 per group (128 total).' (5) Mention assumptions: 'Assumes approximately normal distributions and equal variances.' (6) If relevant, discuss trade-offs: 'If we can only recruit 50 per group, power drops to 70%.' This demonstrates rigorous planning and helps reviewers or instructors assess your study design. Never just say 'we used 64 participants' without explaining why.

There's no universal minimum—it depends on your test, effect size, power, and α. However, as a very rough guideline, most statistical tests become unreliable with fewer than 10–15 observations per group. For power = 0.80, α = 0.05, and a medium effect (d = 0.5 for means), you typically need 60–80 per group. For small effects (d = 0.2), you need 300+ per group. For proportions with small baseline rates or small differences, thousands may be needed. Always use a power calculator rather than guessing. Very small samples (n < 20 per group) rarely have adequate power unless effects are very large. If resources are limited, adjust your expectations: either accept lower power or focus on detecting larger effects (MDE).

Not necessarily. While adequate power (typically 80–90%) is important, extremely high power (e.g., 99%) can be overkill and may detect tiny, practically meaningless effects. With very large samples, even trivial differences become statistically significant (e.g., a 0.1-point difference on a 100-point scale). This can lead to 'statistically significant but practically irrelevant' findings. The goal is to achieve adequate power to detect the smallest effect you care about (minimum meaningful difference), not to maximize power infinitely. Balance power with practical significance, cost, and feasibility. For homework and thesis work, 80% power is typically sufficient. For critical medical trials, 90% might be warranted. But there's no need to aim for 99% unless failure to detect an effect would have severe consequences.

Standard power formulas (like those in this calculator) assume approximately normal distributions for means-based tests. If your data are severely non-normal (e.g., highly skewed, binary outcomes, count data), these formulas may be approximate or less accurate. For proportions, the calculator uses normal approximations that work well with reasonably large samples. For non-parametric tests (e.g., Mann-Whitney U, Wilcoxon), power calculations are more complex and may require simulation or specialized software. As a rough guideline, non-parametric tests often have slightly lower power than parametric tests for the same sample size, so you might increase your sample by 5–15% as a buffer. For homework or preliminary planning, standard power calculations give a reasonable starting point even for mildly non-normal data.

If the required sample size exceeds your budget or resources, you have several options: (1) Accept lower power (e.g., 70% instead of 80%) and acknowledge this limitation. (2) Increase the minimum detectable effect—focus on detecting larger effects that require fewer participants. (3) Use the 'MDE' mode to find the smallest effect you can detect with your available sample, then assess whether that's meaningful in your field. (4) Consider paired or within-subjects designs (if appropriate), which often require smaller samples. (5) Treat your study as a pilot for a larger future study. (6) Collaborate with other researchers to pool data or resources. Always be transparent: report your actual power and explain resource constraints. Don't force a study with 20% power and then overinterpret null results.

Small differences in results across calculators can occur due to: (1) Different approximation methods (some use exact distributions, others use normal approximations). (2) Rounding conventions (some round up sample sizes, others don't). (3) Treatment of continuity corrections for proportions. (4) Different handling of unequal group sizes or allocation ratios. (5) Assumptions about one-tailed vs two-tailed tests. These differences are usually minor (within a few participants) and don't change conclusions. For homework or planning, any reputable power calculator is fine. For critical research, document which calculator and version you used, and report exact assumptions (α, power, effect size, test type). If in doubt, try multiple calculators and see if results converge—they usually do within 5–10%.