There is no single minimum sample size that guarantees statistical significance. The number you need depends on how large the effect you’re trying to detect is, how confident you want to be in your results, and what type of analysis you’re running. That said, concrete numbers do exist for common scenarios, and a power analysis can tell you exactly what you need before you collect a single data point.
Why There’s No Universal Minimum
Statistical significance means your result is unlikely to have occurred by chance alone. The conventional threshold is a p-value below 0.05, meaning there’s less than a 5% probability the result is a fluke. But reaching that threshold isn’t just about how many people or observations you include. It depends on four interconnected factors: your significance level (that 0.05 threshold), your statistical power, the size of the effect you’re looking for, and your sample size. Change any one of these and the others shift with it.
Power is the probability that your study will detect a real effect when one exists. The accepted standard is 80% power, meaning you have an 80% chance of catching a true effect and a 20% chance of missing it. Some researchers aim for 90% power in high-stakes studies, which requires even more participants. The effect size is simply how big the difference or relationship is that you’re trying to measure. A new drug that cuts recovery time in half is a large effect. A supplement that improves test scores by 2% is a small one. Small effects need far more data to distinguish from random noise.
Sample Sizes for Common Study Types
The most useful way to answer “how many do I need?” is to look at real numbers for standard statistical tests at 80% power and a 0.05 significance level. These figures come from calculations based on Cohen’s effect size conventions, which classify effects as small, medium, or large.
For a study comparing two groups (like a treatment group vs. a control group using an unpaired t-test with equal group sizes):
- Large effect (d = 0.8): 28 per group
- Medium effect (d = 0.5): 66 per group
- Small effect (d = 0.2): 402 per group
For a study measuring the correlation between two variables (total sample size):
- Large correlation (r = 0.5): 32 total
- Medium correlation (r = 0.3): 88 total
- Small correlation (r = 0.1): 800 total
For regression analysis, where you’re predicting an outcome from multiple variables, the required sample size also depends on how many predictors you include. Detecting a medium effect requires roughly 52 participants plus the number of predictor variables. A small effect pushes that to 392 plus the number of predictors.
The pattern is clear: as the effect shrinks, the required sample size explodes. Detecting a small effect can require 10 to 25 times more participants than detecting a large one.
Where the “30” Rule Comes From
You may have heard that 30 is the minimum sample size for meaningful statistics. This comes from the Central Limit Theorem, a foundational principle stating that when you have at least 30 observations, the distribution of sample averages approximates a normal (bell-shaped) curve regardless of how the underlying population is distributed. At 30 observations, the t-distribution (used in many common statistical tests) becomes essentially identical to the normal distribution.
But 30 is not a magic number for significance. It’s the point where certain mathematical assumptions start holding up, not a guarantee that your study has enough power to detect anything meaningful. A sample of 30 is perfectly adequate if you’re looking for a large effect, but woefully insufficient for a small one. For pilot studies testing the reliability of a questionnaire or measurement tool, 30 participants is generally considered sufficient. For hypothesis testing in a full-scale study, you almost always need a proper power calculation.
What Happens When Your Sample Is Too Small
An underpowered study doesn’t just fail to reach statistical significance. It actively distorts the research landscape. When your sample is too small, you increase your risk of a Type II error: concluding that no effect exists when one actually does. Studies with low power find fewer true effects, which means promising treatments, interventions, or relationships can be dismissed as ineffective based on inadequate evidence. This can stall entire lines of research, with real effects remaining undiscovered simply because early studies weren’t large enough to detect them.
The flip side is equally problematic. When underpowered studies do happen to reach significance, the effects they report tend to be inflated. This is because only unusually large (and unrepresentative) results clear the significance bar when the sample is small. Other researchers then fail to replicate those inflated findings, contributing to what’s known as the replication crisis.
How to Calculate Your Required Sample Size
The standard approach is called a power analysis, and you run it before collecting data. You need to specify three things: your desired power level (typically 80%), your significance threshold (typically 0.05), and the smallest effect size you want to be able to detect. The calculation then tells you how many participants you need.
The most widely used tool for this is G*Power, a free software program that handles sample size calculations for t-tests, chi-square tests, regression, ANOVA, and other common methods through a simple graphical interface. You select your test type, plug in your parameters, and it returns the required sample size. Most university statistics departments recommend it, and it’s straightforward enough that you don’t need programming experience.
If you don’t have a specific effect size in mind, you can use Cohen’s conventions (small, medium, large) as starting points, or look at prior studies on your topic to estimate a realistic effect size. Using the “medium” category is a common default when no prior data exists, but this is a rough guide. If the true effect turns out to be smaller than you assumed, your study will be underpowered.
Sample Sizes in Clinical Drug Trials
Clinical trials follow a structured progression where sample sizes grow with each phase. Phase 1 trials, which primarily test safety and dosing, typically enroll 20 to 80 participants. Phase 2 trials expand to a few hundred patients and begin evaluating whether the treatment works, though these studies still aren’t large enough to confirm effectiveness on their own. Phase 3 trials, the pivotal studies that determine whether a drug gets approved, involve 300 to 3,000 participants.
These ranges reflect the logic of power analysis at scale. Early phases look for large, obvious signals (severe side effects, clear dose responses) that don’t require huge samples. Later phases need to detect more subtle differences in effectiveness between the new treatment and existing options, which demands much larger groups.
Continuous vs. Categorical Data
The type of data you’re collecting also affects how many observations you need. When you’re measuring something continuous, like blood pressure or test scores, the required sample size depends on the variability in your measurements and how precise you need your estimate to be. More variability in the data means you need more observations to pin down the true average.
When you’re measuring proportions or categories (like the percentage of people who respond to a treatment), the math works differently. If you have no prior estimate of what the proportion might be, the most conservative approach is to assume it’s 50/50, which produces the largest possible required sample size. This guarantees you won’t undercount, though it may lead to collecting more data than strictly necessary. As your prior estimate moves away from 50% in either direction, the required sample size drops.