5 Practical Rules of Thumb for Power & Sample Size

5 Practical Rules Thumb Power Sample Size

Image by Editor

Statistical power and sample size calculations can feel overwhelming when you’re planning a study. While specialized software gives you precise numbers, having some practical rules of thumb helps you quickly assess whether your study design is in the right ballpark. Here are five straightforward guidelines that work across most research situations.

1. The “Magic Number” of 30

For most basic analyses, aim for at least 30 observations per group. This threshold isn’t arbitrary. It’s a practical guideline from the Central Limit Theorem. With n=30, the sampling distribution of the mean becomes sufficiently close to normal for most common distributions, making standard statistical tests reliable.

This rule applies whether you’re comparing two groups, running a regression, or conducting other common analyses. With 30 observations in each condition, you’re comparing, your t-tests and confidence intervals will generally perform well, even if your original data show moderate departures from normality (though extreme outliers or severe skewness may still require alternative approaches).

Keep in mind that this is a starting point, not a universal solution. Some situations need more observations, especially when you’re looking for small differences or working with highly variable data.

2. Bigger Effects Need Fewer Subjects

If you’re looking for a large, obvious difference, you need fewer participants than if you’re hunting for a subtle one. It’s easier to spot an elephant in your living room than a mouse.

When planning your study, consider how big a difference you expect to find. Compare a new teaching method that dramatically improves test scores with one that produces modest gains. The dramatic improvement requires a smaller sample size to detect reliably.

This relationship between effect size and required sample size is one of the most important factors in study planning. A common mistake is using sample sizes appropriate for detecting large effects when your research question involves more subtle differences.

3. 80% Power is the Standard

Power of 80% means you have an 80% chance of detecting a real effect if it exists. This has become the conventional minimum target across most fields.

Why 80% and not higher? It’s a balance between being thorough and being practical. You could aim for 95% power, which sounds better, but that might require an unrealistically large sample that’s impossible to collect within your time and budget constraints.

Think of power as your study’s sensitivity. An 80%-powered study will successfully detect a true effect four times out of five. The remaining 20% represents the risk you accept of missing a real effect due to random chance. Most researchers and review boards consider this an acceptable tradeoff.

4. Doubling Your Sample Doesn’t Double Your Power

Power increases with sample size, but not in a straight line. To double your power from 40% to 80%, you might need to quadruple your sample size. To increase it from 80% to 90%, you might need to double your sample again.

This happens because statistical tests depend on the standard error, which decreases with the square root of sample size. This means that to detect an effect half as large, you need roughly four times the sample size. The practical implication? Adding participants to an already well-powered study produces diminishing returns, while underpowered studies need substantial increases to reach acceptable levels.

This non-linear relationship also means that small pilot studies (say with 10 participants per group) often have very low power, sometimes below 30%. Going from 10 to 20 participants per group helps, but you’ll typically need that jump to 30 or more to reach conventional power levels.

5. Plan for 20% Attrition

If you need 100 complete data points, recruit 120 people. Participants drop out, data gets lost, or people don’t follow protocols. Building in a buffer prevents your study from being underpowered when real life happens.

This rule applies regardless of your research setting. In clinical trials, patients miss appointments or discontinue treatment. In online surveys, respondents abandon questionnaires halfway through. In longitudinal studies, participants move away or lose interest over time.

The 20% buffer is conservative for some contexts and might be insufficient for others. Studies with longer durations, more demanding protocols, or hard-to-reach populations often see higher attrition rates. Adjust this percentage based on similar research in your field.

Applying These Rules Together

These five rules work best as a package. Start with the magic number of 30 as your baseline. Adjust upward if you’re looking for small effects or need higher power. Then add your attrition buffer. This quick calculation gives you a reasonable target for most studies.

Here’s an example: you’re comparing two groups and want to detect a moderate effect with 80% power. Start with 30 per group, giving you 60 total. Add 20% for attrition, bringing you to 72 participants. Round up to 75 for a clean number, and you’ve got a solid recruitment target.

These rules provide a starting framework that helps you avoid severely underpowered studies. For final study planning (especially with grants or high-stakes research), use dedicated power analysis software for precise calculations. But for initial planning and quick feasibility checks, these guidelines will steer you in the right direction.

2 Replies to “5 Practical Rules of Thumb for Power & Sample Size”

These are five sensible rules of thumb, all independently true, and yes they can/should be applied together.
There is one important interaction that hasn’t been addressed, it is rule 1 (The “Magic Number” of 3, and rule 2, the expected sample size (Bigger Effects Need Fewer Subjects) rule.
For a two-tailed independent t-test at N=30 per group (distributions are “approximately normal”, at alpha= .05 and considering rule of “80% Power is the Standard”, then it is important to remember (based on power analysis) that the required effect size is d=0.73, almost a large effect size.
This is and unreasonable expectation in many real life situations. In my own field, in default of evidence, we would postulate small-medium in most expectations.

So, I would suggest the “Magic Number of 30” addresses distributional issues only and is a minimum number; it is trumped by the corollary of rule 2 “smaller effects need bigger numbers” and then the other three rules kick in.

With 30+ years of teaching, supervising, and examining UG, MSc, and doctoral research in psychology, good rules of thumb are important, and rules 4 and 5 are often forgotten and really helpful… but please don’t let a distributional rule (1) overshadow the power-based rule (2).

Edited version:
These are five sensible rules of thumb, all independently helpful, and yes, they can/should be applied together.
There is one important rule interaction that hasn’t been addressed, it is rule 1 (The “Magic Number” of 30) x rule 2, the expected sample size (Bigger Effects Need Fewer Subjects). Let’s look at a few scenarios using GPower
First, for a two-tailed independent t-test at N=30 per group. Distributions are “approximately normal”, at alpha= .05 and considering rule 3 – “80% Power is the Standard”, then it is important to remember (based on power analysis) that the required effect size is d=0.73, almost a large effect size.
Second, for a single group simple bivariate regression/correlation, at N=30 (alpha = .05, one-tailed, power = .80) required effect size is approx.42, closer to a large effect size than medium. With three variables, the combined effect size (deviation from zero f2) would need to be large. For the last step of three in a hierarchical regression (R2 increase), required effect size R2 would also be .28 (large).
These large effect sizes are, in my own experience, an unreasonable expectation in many real life situations. In my own field, in default of evidence to the contrary, we would postulate small-medium effect sizes in most scenarios.
So, I would suggest the “Magic Number of 30” addresses distributional issues only and is a minimum number. It is trumped by the corollary of rule 2 “smaller effects need bigger numbers” and then the other three rules kick in.
With 30+ years of teaching, supervising, and examining UG, MSc, and doctoral research in psychology, good rules of thumb are important.
Rule 3 (80% power) is a reasonable trade-off. Rules 4 (unequal groups inflates N required vs equal groups) and 5 (data loss/attrition) are often forgotten and really helpful… But please don’t let a fundamentally sensible (minimum) distributional rule (1) overshadow the (critical) power-based rule (2).

Professor Mark Freeston says:

November 17, 2025 at 5:52 pm

These are five sensible rules of thumb, all independently true, and yes they can/should be applied together.
There is one important interaction that hasn’t been addressed, it is rule 1 (The “Magic Number” of 3, and rule 2, the expected sample size (Bigger Effects Need Fewer Subjects) rule.
For a two-tailed independent t-test at N=30 per group (distributions are “approximately normal”, at alpha= .05 and considering rule of “80% Power is the Standard”, then it is important to remember (based on power analysis) that the required effect size is d=0.73, almost a large effect size.
This is and unreasonable expectation in many real life situations. In my own field, in default of evidence, we would postulate small-medium in most expectations.

So, I would suggest the “Magic Number of 30” addresses distributional issues only and is a minimum number; it is trumped by the corollary of rule 2 “smaller effects need bigger numbers” and then the other three rules kick in.

With 30+ years of teaching, supervising, and examining UG, MSc, and doctoral research in psychology, good rules of thumb are important, and rules 4 and 5 are often forgotten and really helpful… but please don’t let a distributional rule (1) overshadow the power-based rule (2).

Mark Freeston says:

November 17, 2025 at 6:35 pm

Edited version:
These are five sensible rules of thumb, all independently helpful, and yes, they can/should be applied together.
There is one important rule interaction that hasn’t been addressed, it is rule 1 (The “Magic Number” of 30) x rule 2, the expected sample size (Bigger Effects Need Fewer Subjects). Let’s look at a few scenarios using GPower
First, for a two-tailed independent t-test at N=30 per group. Distributions are “approximately normal”, at alpha= .05 and considering rule 3 – “80% Power is the Standard”, then it is important to remember (based on power analysis) that the required effect size is d=0.73, almost a large effect size.
Second, for a single group simple bivariate regression/correlation, at N=30 (alpha = .05, one-tailed, power = .80) required effect size is approx.42, closer to a large effect size than medium. With three variables, the combined effect size (deviation from zero f2) would need to be large. For the last step of three in a hierarchical regression (R2 increase), required effect size R2 would also be .28 (large).
These large effect sizes are, in my own experience, an unreasonable expectation in many real life situations. In my own field, in default of evidence to the contrary, we would postulate small-medium effect sizes in most scenarios.
So, I would suggest the “Magic Number of 30” addresses distributional issues only and is a minimum number. It is trumped by the corollary of rule 2 “smaller effects need bigger numbers” and then the other three rules kick in.
With 30+ years of teaching, supervising, and examining UG, MSc, and doctoral research in psychology, good rules of thumb are important.
Rule 3 (80% power) is a reasonable trade-off. Rules 4 (unequal groups inflates N required vs equal groups) and 5 (data loss/attrition) are often forgotten and really helpful… But please don’t let a fundamentally sensible (minimum) distributional rule (1) overshadow the (critical) power-based rule (2).