Module 2 · Randomisation & Blinding

How can a coin toss defeat a problem nobody can even see?

Last lesson named the cure for confounding: randomisation. But that should nag at you. Confounding is hidden third factors — things no one has measured, maybe no one has even thought of. So how on earth can flipping a coin neutralise a danger you can't see?

Why chance, of all things?

Put yourself in charge of the trial. You need to split patients into two groups — drug and control — and you want to be fair. So you pick a rule. Watch what every rule quietly does:

• "The sickest patients get the new drug." Now the drug group starts off worse. Any real benefit is hidden; any harm exaggerated.
• "Whoever volunteers gets the drug." But volunteers differ — keener, often healthier, more anxious. The groups aren't comparable.
• "Alternate patients — every other one." Seems fair, until you realise whoever enrols can see the pattern coming and nudge who walks through the door next.

Every rule based on anything — health, choice, order, a clinician's judgement — lets some systematic difference creep in. And a systematic difference between the groups is a confounder waiting to happen.

There is exactly one way to assign treatment that has no agenda, follows no pattern, and favours no trait: pure chance. A coin toss doesn't know who's sick, who's keen, or who's next. That blindness is precisely its power.

The killer feature: the confounders you can't see

Here's the part that feels like magic.

When you assign by chance, the two groups come out alike in every characteristic at once — not just the ones you measured, but the ones you never thought to.

Compare that with the alternative. If you're worried about confounding, you could try to adjust for it: measure age, smoking, weight, and statistically level the groups afterward. But that only works for confounders you (a) thought of and (b) actually measured. The one that wrecks your trial is the one you never knew to look for.

Randomisation doesn't need to know. By splitting purely at random, it balances age and smoking and weight — and the genetic quirk no one has discovered, the dietary habit no one recorded, the unknown unknown. All of them, on average, fall evenly on both sides.

Adjustment can only balance the confounders you can name. Randomisation balances the ones you can't. That is why it sits at the very top of the ladder.

What randomisation does NOT do

One honest caveat, or you'll over-trust it.

Randomisation balances the groups on average — across many trials, or within one large trial. It does not guarantee perfect balance in any single small one.

Toss a coin for 20 patients and you might, by bad luck, land most of the smokers in one group. Randomisation didn't fail — chance simply had room to wobble. Sound familiar? It's our old enemy chance, and the cure is the same: size. The bigger the trial, the more reliably random allocation delivers genuinely balanced groups.

So randomisation defeats confounding — but small trials can still wobble. That's why a good trial publishes a baseline characteristics table: so you can check, with your own eyes, that the groups really did come out alike before treatment began.

The loophole: allocation concealment

A random sequence is only as good as your inability to game it. Picture this: the allocation list is perfectly random — but the doctor enrolling patients can see the next entry before deciding whether to enrol someone.

A doctor who believes in the drug might, without any bad intent, hold off enrolling a frail patient until a control slot comes up — quietly protecting the drug's results. The sequence was random. The allocation was corrupted anyway.

The fix is allocation concealment: hiding the upcoming assignment until the patient is irreversibly enrolled. A central computer or phone line, sealed opaque envelopes — anything ensuring the person deciding "in or out" cannot know which arm comes next.

Randomisation generates a fair sequence. Allocation concealment stops anyone bending it. You need both — and they are not the same thing. (Nor, as you're about to see, are they the same as blinding.)

Each trial below describes how patients were assigned. Tap the verdict — and watch for which safeguard is missing.

The trial isn't over at the starting line

Randomisation and concealment did their job: at the starting line, the two groups are twins. But a race isn't won at the start. Everything that happens after allocation can still pull the groups apart — and none of it is confounding. It's our other enemy: bias.

Think about what changes the moment treatment begins and people know who's getting what:

• Patients who know they're on the exciting new drug may feel better, try harder, report differently — the placebo effect, and more.
• Doctors who know may treat the groups differently — more attention, extra tests, faster referrals for one arm.
• Whoever judges the outcome may score the new-drug patients more generously.

Equal at the start isn't enough. You need the groups treated and measured equally throughout. That is the job of blinding.

Who is blinded, and why each matters

"Blind" simply means kept unaware of who got which treatment. And the useful question is never "was it blinded?" but "who was blinded?" — because each role guards a different leak. (This is one reason to distrust the vague label "double-blind": it gives you a number, not the names.)

Three people can be kept in the dark, each for a reason:

• The patient — blinding them controls the placebo effect and stops their knowledge colouring how they feel, behave and report. Matters most for anything self-reported.
• The treating clinician — blinding them stops the groups being managed differently: no extra co-interventions, attention or care tilting one arm. (This is performance bias.)
• The outcome assessor — whoever measures or judges the result. Blinding them stops the new-drug group being scored more favourably. (This is detection bias.)

A trial can blind some of these and not others. Naming who was blinded tells you exactly which leak was plugged — and which was left open.

Why placebos exist

A question that puzzles newcomers: why give the control group a fake pill at all? Why not just give them nothing?

Because you cannot blind a patient who knows they got nothing. The moment the control group can tell they're untreated, the patient — and often the clinician — is unblinded, and every bias from the last screen pours back in.

A placebo, an identical-looking dummy with no active ingredient, exists for exactly one reason: to make the two groups indistinguishable, so no one can tell who is who.

The placebo isn't there to deceive for its own sake — it's the device that makes blinding possible. The same logic gives us sham procedures in surgical trials: a mock operation, so the patient can't tell whether they got the real one.

When blinding is hard — and when it matters less

Sometimes you simply can't blind. You can't give someone a fake hip replacement, or hide whether they're doing an exercise programme. Does that sink the trial? Not necessarily — and the reason is elegant.

It depends on the outcome. Blinding matters most for subjective outcomes — pain, fatigue, "feeling better," quality of life — anything filtered through human judgement, where knowing the arm can tilt the answer. It matters far less for hard, objective outcomes like death: a patient's knowledge can't argue with a death certificate.

So an unblinded trial measuring survival can still be trustworthy, while an unblinded trial measuring self-reported pain should make you nervous. And even when you can't blind the patient, you can usually still blind the outcome assessor — have someone who doesn't know the arm read the scans or score the results. Blinded assessment can rescue a trial that couldn't blind anyone else.

The question isn't just "was it blinded?" — it's "was it blinded where it mattered, given what was being measured?"

Each trial below left someone unblinded, and it cost them. Tap who most needed to be in the dark.

The logic of the RCT, in one line

Step back and see the whole machine. Two tools, two enemies, two moments in time:

• Randomisation acts at the starting line. It makes the groups identical at baseline — defeating confounding, including the confounders no one could even name.
• Blinding acts throughout the race. It keeps the groups treated and measured identically after allocation — defeating bias and the placebo effect.

Put them together and you get the single most powerful sentence in clinical research:

If the groups were identical at the start, and treated and measured identically throughout, then any difference at the end must be caused by the treatment itself.

That sentence is the entire reason the randomised controlled trial sits at the top of the evidence ladder. Everything else supports it. These two tools are what make it true.

Why this matters for HTA

When you appraise a trial in a submission, you now have a checklist that cuts straight to its credibility:

• Was it actually randomised — by a real random method, not judgement or alternation?
• Was allocation concealed, so no one could bend the sequence?
• Who was blinded — patient, clinician, assessor? And given the outcome, was that enough?

A trial that ticks all of these is hard to fool. One that doesn't isn't worthless — but every missing safeguard is a door through which bias or confounding could have walked. And a manufacturer will rarely volunteer which doors were left open.

Watch especially for unblinded trials reporting subjective outcomes — the exact combination where bias does its quietest, most flattering work.

Randomisation & blinding, in one breath

• Randomisation assigns by chance because chance has no agenda — and it uniquely balances confounders both known and unknown.
• It works on average; small trials can still wobble (that's chance), so bigger is better — and baseline tables let you check.
• Allocation concealment stops anyone gaming the sequence — distinct from randomisation, and from blinding.
• Blinding keeps the groups treated and measured equally throughout; ask who was blinded — patient, clinician, assessor — not just "was it."
• Blinding matters most for subjective outcomes; placebos and sham procedures exist to make it possible.

Randomisation makes the groups equal at the start. Blinding keeps them equal to the end. Whatever difference is left must be the treatment.

But notice exactly what this buys you: confidence the result is true for the patients in the trial. Whether it's also true for your patients — older, sicker, different from a trial's tidy sample — is a separate question entirely. That gap between "true here" and "true there" is the last piece of M2: internal versus external validity.