The placebo problem in a trial you run on yourself

The standard objection to worrying about placebo in a self-experiment is that you already know what you took, so there is nobody to deceive. The first half is true and the second does not follow. What contaminates a trial you run on yourself is not deception. It is that you changed something, then measured, and everything else that was going to move the number moved over the same fortnight.

Three things happen at once

Krogsbøll and colleagues pulled out the trials that randomised patients to three arms — no treatment, placebo, and an active intervention — and measured change from baseline in each. Thirty-seven trials, 2,900 patients, eight conditions [1].

The active arm changed by −1.01 (95% CI −1.16 to −0.86), putting spontaneous improvement and placebo at 24% and 20% of the treatment effect — averages the authors qualify by pointing at the width of all three intervals [1]. An uncontrolled before-and-after returns the three added together.

How large the expectation part is

Hróbjartsson and Gøtzsche pooled 202 of 234 randomised trials comparing placebo against a no-treatment control, across 60 clinical conditions. Across 158 trials with continuous outcomes (10,525 patients) the pooled placebo effect was an SMD of −0.23 (95% CI −0.28 to −0.17); across 44 trials with binary outcomes (6,041 patients), a relative risk of 0.93 (95% CI 0.88 to 0.99) [2].

The split that matters when choosing an outcome is who reported it. Patient-reported continuous outcomes gave an SMD of −0.26 (95% CI −0.32 to −0.19); observer-reported outcomes gave −0.13 (95% CI −0.24 to −0.02). Meta-regression found larger placebo effects with patient-involved outcomes, with small trials, and with trials that did not inform patients about the possible placebo intervention [2].

That last association points the other way for self-experimenters. Someone running their own trial is maximally informed, which is the condition under which the review found placebo effects smaller. Smaller is not absent.

The outcome we care most about has no answer

Insomnia is on the review's list of seven conditions studied in three or more trials where no statistically significant placebo effect was found, with confidence intervals the authors call wide [2]. That is not evidence that expectation leaves sleep alone. It is the absence of an estimate for the outcome we spend most of our time measuring.

Knowing it is a placebo may not fix it

The escape route — you cannot fool yourself — is undercut by the open-label placebo literature, where participants are told outright that the pill is inert.

A large estimate on a thin base. Four of the thirteen reviewed studies were rated high risk of bias, mostly for missing outcome data and unblinded outcome assessors; excluding them dropped heterogeneity from I² = 76% to I² = 4%. Every primary outcome pooled was self-reported [3].

What blinding costs

Blinding yourself is not absurd, only narrow. Of 74 randomised n-of-1 trials published between 2011 and 2023, 57 (77.0%) were blinded and 49 (66.2%) placebo controlled, though only 32 (43.2%) used a washout [4]. It works because a capsule has a matched twin. Cold water and an earlier bedtime do not.

Where it can be done, it changes the answer. StatinWISE randomised 200 primary-care patients who had stopped or were considering stopping statins because of muscle symptoms to six double-blind two-month periods of atorvastatin 20 mg or matched placebo [5].

Withdrawals for intolerable muscle symptoms ran to 18 participants (9%) in statin periods and 13 (7%) in placebo periods, and two thirds of those who completed intended to restart statins [5]. Their prior belief was specific, strong, and drawn from their own experience.

What this does not establish

The review rated risk of bias low in only 16 of its 202 trials (8%), and for continuous outcomes reports moderate heterogeneity and an asymmetrical funnel plot.

So −0.23 is not a constant to subtract from your own result, and none of these figures give the placebo effect for your intervention on your outcome. StatinWISE tested one statin at one dose and lost 43% of participants to withdrawal; a group-level null does not establish a null for any individual.

Why this is an n-of-1 problem

A population trial does not remove expectation. It puts the same expectation in both arms and subtracts. You have one arm at a time, so nothing subtracts unless you build the comparison. Randomising the start date and the order of periods is worth doing and does not fix this — it breaks the link between the intervention and time, which protects you from the season, the training block, and the bad week that made you start. That is the −0.24 arm. The −0.44 arm survives it.

So we record whether a trial was blinded, and where it could not be, we report an estimate that still contains expectation rather than netting out a number the literature does not support. An unblinded subjective outcome is an upper bound, and we label it as one.