The minimum detectable effect (MDE) is the smallest true difference between control and variant that a test has a stated probability, its power, of detecting as significant. Teams pick an MDE before a test to decide how many visitors it needs, or calculate it from available traffic to see what size of change the test could possibly find.
Relative and absolute MDE
MDE can be stated as an absolute change in percentage points or as a relative change. If a page converts at an illustrative 5%, a relative MDE of 10% means detecting a move to 5.5%, which is an absolute change of 0.5 points. Calculators differ in which they ask for, so check before entering numbers.
relative MDE = (variant rate − control rate) / control rate
absolute MDE = variant rate − control rate
control 5.0%, variant 5.5%
relative = 10%, absolute = 0.5 pointsChoosing an MDE
Set the MDE at the smallest effect that would change a business decision or be worth the cost of shipping. Halving the MDE roughly quadruples the sample size needed, so asking to detect very small effects on low-traffic pages leads to impractically long tests.
Work backward too: given the traffic you can send in a few weeks, what MDE can you achieve? If the answer is a very large lift, only bold changes are worth testing on that page.
What MDE does not mean
An MDE is not a prediction of the result. A test sized for a 10% relative lift can still find a smaller real effect, though with lower probability, and it can find nothing if the change has no effect. A non-significant result means the effect, if any, was probably smaller than the MDE, not that it was zero.
Example
Sizing a test from traffic
Hypothetical survey platform Hollowreach can send its trial signup page an illustrative 20,000 visitors over four weeks. Plugging that into a calculator shows the test could reliably detect only a fairly large relative lift. The team drops a minor copy tweak from the queue and instead tests a redesigned form that removes three fields.
Common mistakes
Mixing relative and absolute values
Entering a relative MDE where a calculator expects absolute points produces a wildly wrong sample size. Read the input label.
Picking an MDE to make the test short
Inflating the MDE until the duration looks acceptable just means the test can only detect effects that are unlikely to happen.
Questions
It depends entirely on traffic, baseline rate and what the business considers worthwhile. Pick yours from the decision you need to make, not a generic figure.
