Sample Size vs Power in A/B Testing

Learn the difference between sample size and statistical power in A/B testing, how they interact, and why increasing traffic alone does not guarantee reliable results.

Article visual illustration
Twitter LinkedIn

Executive Summary

Sample size is how much data you collect.
Statistical power is the probability that your experiment will detect a real effect if it exists.
Increasing sample size increases power — but only relative to effect size, variance, and significance level.
Many experiments fail because teams confuse having “a lot of data” with having enough power to answer their question.


Why This Confusion Exists

In experimentation discussions, “sample size” and “power” are often used interchangeably — but they are not the same thing.

Teams frequently say:

Yet the experiment still produces no result.

The root cause is almost always the same: power was never evaluated.
Sample size was tracked, but detectability was never assessed.


Core Definitions

Sample Size

Sample size is the number of observations collected per variant in an experiment.

Examples include:

Sample size is a quantity of data, not a measure of experimental sensitivity.


Statistical Power

Statistical power is the probability that a statistical test will correctly reject the null hypothesis when a true effect exists.

Formally: [ \text{Power} = P(\text{Detect effect} \mid \text{Effect is real}) ]

Commonly targeted power levels:

Power answers a forward-looking question:

If there is a real effect of the size I care about, how likely am I to detect it?


The Relationship Between Sample Size and Power

Sample size is one input into power — not the output.

Power depends jointly on:

Increasing sample size increases power, but:

More data does not automatically mean a more informative experiment.


Intuition: A Signal vs Noise View

An experiment attempts to detect a signal buried in noise.

Increasing sample size reduces noise by averaging it out.
Statistical power measures whether the remaining signal is strong enough to be reliably detected.

If the signal is too weak relative to the noise, no reasonable amount of listening will help.


Mental Model

Imagine two overlapping distributions representing the outcomes of your control and variant.

When sample size is small:

As sample size increases:

Power describes how separated these distributions must be before the experiment can reliably tell them apart.


Formal Statistical Framing

In classical hypothesis testing, detectability is governed by the inequality:

Effect size ≥ (z_crit + z_{1−β}) × SE

where ,

Effect Size : The true difference between variants that you care about detecting.

Examples:

Power is always defined relative to a specific effect size.

z₍crit₎ : Significance Threshold

The critical value corresponding to the chosen significance level (α).

This term controls the false-positive rate.

z₍1−β₎ : Power Requirement.

The z-score corresponding to the desired power level.

Higher power requires a larger safety margin for detection.

SE — Standard Error :

The standard error measures how much random variation exists in the estimated metric.

It depends on: - Metric variance - Sample size - Experimental design

Larger sample sizes reduce standard error, which increases power. Intuitively , Think of it this way, the more you see something of, the more sure you are of it. That’s why Larger sample size reduces standard error.

Why Sample Size Alone Is Not Enough?

Two experiments can have the same sample size but very different power.

Power will be lower when: - The metric is highly variable - The baseline rate is extreme - The effect of interest is small - The design introduces dependence or clustering

This is why statements like “we had 50,000 users” are meaningless without context.

Worked Example (With Calculation)

Scenario

You run a conversion rate experiment on a landing page.

Inputs


Step 1: Standard Error

For a binary metric, the standard error is:

$ SE = \sqrt{p(1 - p)\left(\frac{1}{n_1} + \frac{1}{n_2}\right)} $

Using equal sample sizes:

$ SE \approx \sqrt{0.1 \times 0.9 \times \frac{2}{10000}} \approx 0.0042 $


Step 2: Detection Threshold

The combined detection threshold is:

$ z_{crit} + z_{1-\beta} \approx 1.96 + 0.84 = 2.80 $

The minimum detectable effect is:

$ \text{Detectable effect} \approx 2.80 \times 0.0042 \approx 0.0118 $


Interpretation

The experiment can reliably detect effects of about 1.2 percentage points or larger.

Smaller true improvements are likely to go undetected, even if they exist.

Common Misconceptions

How to Use This in Practice?

Before launching any experiment, answer one question:

“What is the smallest effect that would justify acting on the result?”

Then: 1. Choose confidence level and power 2. Estimate variance realistically 2. Compute required sample size or detectable effect

If the numbers do not work, do not run the experiment.

How This Relates to MDE?

Minimum Detectable Effect (MDE) is the bridge between sample size and power.

MDE answers: “Given my sample size, variance, and desired power, what effects can I actually detect?”

To understand this constraint in detail, see
👉 What Is Minimum Detectable Effect (MDE)?

Frequently Asked Questions

Can I increase power without increasing sample size?

Only by:

Is 80% power always sufficient?

No. It is a convention, not a law. Higher-stakes decisions often require higher power.

Why do many real-world experiments have low power?

Because effect sizes are overestimated, variance is underestimated, and feasibility constraints are ignored during planning.