Fair comparisons: placebo, blinding and who counts
Randomization made the two groups alike. What can still go wrong?
The dropout lab
Some patients on the drug stop taking it. Do we still count them?
1. Who is counted?
Count everyone is called intention to treat: every patient is counted in the group the coin chose, even if they quit. Drop quitters is called per protocol: only patients who stuck to the plan are counted. Patients who quit get no help from the drug.
2. Settings
Which question? "What does the drug do if taken?" and "What happens when we prescribe it and some patients stop?" are different questions. A trial must say up front which one it answers (ICH E9(R1) calls this the estimand).
One simulated trial, patient by patient
Predict, then check
Blinding: who is kept in the dark?
Real trials
RECOVERY: dexamethasone
Everyone knew who got the drug. Death is less open to opinion than symptoms, but unblinded care can still differ.
Horby et al., N Engl J Med 2021;384:693-704.
PANORAMIC: molnupiravir
26,411 patients were randomized. Hospitalization or death was rare. The trial did not show a reduction in this mostly vaccinated population. Adjusted odds ratio 1.06, 95% credible interval 0.81 to 1.41.
Butler et al., Lancet 2023;401:281-293 A clear "no" is a useful answer.
Regulator's corner
"Preservation of the initial randomisation in analysis is important in preventing bias"
In plain words: analyze patients in the groups the coin chose.
ICH E9, section 5.2.1 (PDF).
Takeaway
- Hide who got what with a look-alike placebo, so hopes cannot color the results.
- Count everyone in the group the coin chose, even people who quit (intention to treat).
- Say up front which question the trial answers: the drug if taken, or prescribing it (the estimand).
- Dropping quitters makes the groups unequal again.