The Placebo Problem

In a controlled trial, one arm receives the thing being tested and another receives an inert version matched as closely as possible on everything else. The point of the second arm is not that it does nothing. It is that it does everything except the one thing under test, so the difference between the arms isolates that thing.

Three meanings of one word

Confusion arises because the same word is used for the inert item, for the improvement seen in the arm that receives it, and for a supposed mechanism by which expectation produces change. Only the first is uncontroversial. The improvement in a control arm is a sum of several things, and expectation is only one of them.

The others are ordinary and often larger. Many conditions fluctuate and people enter studies when they are worst, so the next measurement tends to be better regardless. Measurement itself changes what is reported. Participants in trials receive attention, monitoring and structure. Reporting to a researcher differs from reporting to nobody.

Why the distinction is not pedantic

Studies designed to separate these components, by including an arm that receives no treatment at all alongside the inert-treatment arm, generally find that a substantial part of what is loosely called the placebo effect is accounted for by fluctuation and measurement rather than by expectation. That result does not eliminate expectation effects, which are real and measurable in some domains, but it does cut the popular version of the claim down considerably.

The consequence for reading trials

A trial estimates a difference. Reporting the size of the improvement in the treatment arm alone, which is common in secondary coverage, describes something the design was never built to estimate. When a summary quotes only how much the treated group improved, the number is not the trial's result, whatever else it may be.

Permalink: /notes/the-placebo-problem