Replication, Explained: Why Experiments Repeat Themselves

Replication is running the same factor combination more than once — a full, independent repeat of the run, not a second reading of the same part — so the experiment carries its own measurement of natural process noise. Without it, an experimenter can see that two conditions differ but has no honest way to say whether the difference is signal or scatter.

How it works

Every process varies run to run, even with all settings held identical. Replicates capture that variation directly: the spread among repeats of the same condition is pure error, the yardstick against which factor effects are judged. An effect that towers over the pure error is real; one that sits inside it is noise wearing a costume. Replication also sharpens the estimates themselves, since averaging repeated runs shrinks the uncertainty of every effect in the experiment.

The distinction that trips teams is replication versus repeated measurement. Measuring one molded part three times tells you about the gauge, not the process. A true replicate starts from scratch — new setup, new material, new run — in randomized order alongside everything else. How many replicates are enough depends on the smallest effect worth detecting: subtler signals demand more repeats, and a sample-size calculation before the experiment beats regret after it.

A worked example

A machining cell tests two factors against surface roughness in a four-run factorial. Run once, the experiment shows a difference between conditions — but is a difference of that size unusual, or just Tuesday? The team replicates the design: eight runs, each combination twice, fully randomized. The pairs of repeats disagree with each other by a certain spread, and that spread becomes the benchmark. One factor’s effect stands far outside it and is declared real; the other’s sits comfortably within it and is retired. The same data pattern, judged against measured noise instead of intuition, produces a defensible decision.

Practical pointers

• Budget replicates from the start — an unreplicated experiment often has to be rerun anyway
• Randomize replicates along with everything else; back-to-back repeats understate the true noise
• Never present repeated measurements of one unit as replicates — they measure the gauge, not the process
• If full replication is unaffordable, replicate the center point several times for an economical noise estimate

Replication is how an experiment learns the difference between a discovery and a Tuesday.

Judging effects against noise is the core skill of experimental analysis, and it is developed in our Black Belt program ($499, roughly 60 hours of material, with a 150-question proctored exam and one free retake included). Green Belt ($299, 35 hours) lays the statistical groundwork — standard deviation, sampling, hypothesis tests — that makes the logic of replication second nature.