Minimum Detectable Effect (MDE) Explained for A/B Testing (2026)

The minimum detectable effect is the input that decides, before a single user sees your test, whether the experiment is even capable of answering your question. Teams obsess over p-values and confidence intervals after the fact, but the MDE - set at design time - quietly determines how much traffic you burn and how long you wait. Set it carelessly and no amount of clever analysis saves you. This is what it means and how to choose it well.

The minimum detectable effect (MDE) is the smallest true effect your test is powered to detect reliably. It is a design decision you make up front, not a number the test reports back.

What “detectable” actually means

Say you set an MDE of a 10 percent relative lift at 80 percent power. You are declaring: “size this test so that if the real improvement is 10 percent or larger, I have an 80 percent chance of catching it as statistically significant.” Effects bigger than your MDE are easy to catch. Effects right at the MDE you catch about 80 percent of the time. Effects smaller than the MDE will often slip through as “not significant” - not because they are not real, but because you did not build the test to see them.

That last point is the one teams miss. A non-significant result on an underpowered test does not prove the variants are the same. It means you chose not to spend the traffic required to detect an effect that small. The MDE is where you made that choice.

Why the MDE is the master lever

The MDE is not just one input among four - it dominates the sample size calculation because of how the math is shaped. Required sample size scales with one over the MDE squared.

MDE (relative lift, 5 percent baseline)	Sample per variation	Relative traffic
20 percent	~8,150	1x
10 percent	~31,200	~4x
5 percent	~124,000	~16x

Halving your MDE roughly quadruples the traffic and the runtime. Going from a 20 percent MDE to a 5 percent one is not a modest tightening - it is a 16-fold increase in the data you need. This single relationship explains why ambitious teams end up with tests that never conclude, and why low-traffic sites should chase big, bold changes rather than subtle tweaks.

How to choose the MDE honestly

The right way to set an MDE runs in one direction only - from business value to traffic, never the reverse.

1. Start from what is worth shipping. Ask: what is the smallest lift that would justify building, testing and maintaining this change? If a 3 percent lift on checkout is real money and a 1 percent lift is not worth the code, your MDE is around 3 percent. That is a product judgment, not a statistical one.
2. Compute the sample size that MDE requires using your real baseline rate, 5 percent significance and 80 percent power.
3. Check it against your traffic. Divide the required sample by your weekly visitors to that page. If the answer is more weeks than you can wait, you have a choice to make - not a calculation to fudge.

The cardinal sin is reverse-engineering the MDE from the traffic you have. Teams do this constantly: they see they can only run for two weeks, back out the MDE that fits, and pretend that was the plan. If that MDE is a 40 percent lift, you have quietly agreed to only ever detect enormous effects and to miss every realistic one - usually without saying so out loud.

Common MDE mistakes

• Setting it too small out of greed. “Even a 1 percent lift matters” sounds rigorous and commits you to a test needing hundreds of thousands of users per arm. Be honest about the smallest lift you would act on.
• Confusing relative and absolute. A 10 percent relative lift on a 5 percent baseline is 0.5 absolute points. A 10 percent absolute jump would be 5 to 15 percent - a completely different, far easier target. Our confidence intervals guide covers this distinction.
• Ignoring how power interacts with peeking. An underpowered test plus peeking is the worst combination - low ability to find real effects and inflated false positives. See the statistical significance guide.
• Treating the MDE as a prediction. The MDE is not your guess at the effect. It is the detection floor you are willing to pay for. The real effect can be bigger, smaller, or zero.

How variance reduction changes the math

You are not entirely stuck with the traffic your MDE demands. Variance reduction techniques lower the noise in your metric, which lets you detect the same MDE on less traffic - or a smaller MDE on the same traffic.

CUPED uses each user’s pre-experiment behavior to strip out predictable variation, and it is available in Statsig and GrowthBook, both of which pair it with sequential testing. GrowthBook’s warehouse-native engine also runs sample ratio mismatch checks so you know your traffic actually split the way your MDE calculation assumed. Optimizely built its stats engine to stay valid under continuous monitoring, which protects the power you paid for. For choosing among them, see the best experimentation platforms roundup and Statsig vs GrowthBook.

The bottom line

The MDE is the smallest lift your test can reliably catch, and it is a business decision disguised as a statistical one. Set it from the smallest effect worth shipping, compute the sample size it demands, and check that against your real traffic before you start. Remember that halving the MDE quadruples the cost, that relative and absolute lifts are not the same, and that a non-significant result on an underpowered test proves nothing. Choose the MDE with your eyes open and the rest of the experiment - sample size, runtime, and the honesty of the result - falls into place.