Fixed or random, which factor is which
Today let's work out which parts of a study design you get to call fixed on purpose, and which ones are genuinely random, because that one decision changes which statistical test is the correct one to run.
A pharmaceutical company is testing a new blood pressure drug. It picks four exact doses to test: 0 milligrams, which is the placebo, then 10, 20 and 30 milligrams. The trial runs at twenty clinics, drawn at random from a much larger pool of clinics that could have taken part.
Each clinic enrols two patients at every one of the four doses, 160 patients in all. Every patient's result is the same measurement: the drop in diastolic blood pressure, in mmHg, two weeks after starting treatment.
Here is each patient's drop in blood pressure, grouped by the dose they received, pooled across all twenty clinics.
The boxes climb as the dose goes up. The question this lesson answers is whether that climb is something you can trust, or something that depends on which twenty clinics happened to run the trial.
What makes a factor fixed, and what makes it random
A factor in a study design is fixed when you chose its exact levels on purpose, and the question you care about is only ever about those same levels. The four doses in this trial, 0, 10, 20 and 30 milligrams, are fixed. The company picked these exact amounts deliberately, and the result it wants is about these four doses, not some other set a different company might have tried.
A factor is random when its levels are themselves a sample, drawn from a larger population you want to generalize past. The twenty clinics are random. They were drawn from a much bigger pool of clinics that could have taken part, and the company does not care about these twenty clinics specifically. It cares about how the drug performs across clinics in general.
Here is a simple test that tells the two apart. Imagine running the whole trial again from scratch. Would you pick the same levels on purpose, or would a fresh random draw hand you different ones?
For dose, you would pick the same four amounts again, because those are the exact doses the company wants data on. For clinic, a second trial would draw a different random set of twenty clinics, not necessarily the same twenty. That difference, the same levels on purpose against a fresh random draw of them, is what separates a fixed factor from a random one.
This sounds like a small distinction. It is not: it decides which number belongs in the denominator of an F-test.