Power Posing Mostly Died. The Part That Lives Is Slouching.

Replication labels. Hormone effect: failed. Risk-taking effect: failed. Felt-power effect: replicated, small, and contested. The slogan "adopt a power pose and you change your body and your behavior" is dead. One piece of it lives, and it is smaller and odder than the TED version.

I keep a graveyard of failed effects. Power posing does not go in it whole. I grade each claim inside the finding on its own, and the three claims have three different fates.

The question

The 2010 study had 42 participants and claimed that high-power poses raised testosterone, lowered cortisol, increased risk tolerance and increased feelings of power [1]. If the groups were equal, that is about 21 people per pose (my assumption, not a reported split). Four claims, 42 people. I check the sample size before the abstract, and this one made me wince.

So the question is: which of the claims survived, at what size, and against which comparison?

Data and where it came from

I read the following, in this order of importance.

• The large single replication: Ranehill, Dreber and Weber, 2015, N = 200 (98 women, 102 men), built with 95% power to detect the original effect size [2].
• The 2017 special issue of Comprehensive Results in Social Psychology: seven preregistered studies [4], plus a Bayesian meta-analysis of six of them on felt power [5]. The BPS summary puts that felt-power analysis at 1,071 participants [3].
• Two later meta-analyses: Elkjær et al. 2022, 73 studies [9], and Körner et al. 2022, 88 studies and 9,779 participants [12].
• The p-curve fight: Simmons and Simonsohn on 34 studies [6], Cuddy, Schultz and Fosse's reply on 55 studies [7], and Credé's critique of that reply [8].
• The lead author's own retraction of belief [3].

Limits first. Several full texts blocked me, so for the 2022 meta-analyses I relied on abstracts and summaries. I could not find an interval for the felt-power effect in anything I could open. That matters, and I come back to it.

Method

I graded each claim on three questions. Did a preregistered or large study find it? Did it survive correction for publication bias, as far as I can tell? Does the answer depend on the comparison group? I did no new computation. All numbers below are published estimates, and any arithmetic is shown.

Result

Hormones: dead

Ranehill et al. found no significant change in testosterone or cortisol after posing [2]. Later studies agreed. Deuter et al. in a stress test and Smith et al. in a competition setting both found no effect on hormones, risk or feelings [1]. The special issue's seven preregistered studies pooled to "virtually zero" on hormonal and behavioral outcomes [4]. Elkjær's meta-analysis found no hormonal effect either [9]. Körner's team reported power poses "hardly" moved hormones, blood pressure or heart rate [12].

The hormone claim started at 42 people and lost to a sample 4.8 times larger (200 / 42). It is also the claim the field now treats as settled. Cuddy's own side did not defend it in the p-curve reply, which focuses on feelings [7].

Risk-taking: dead, with a footnote

Ranehill found no effect on the original gambling task, a loss-domain task or willingness to compete [2]. Garrison et al.'s preregistered study found none on risk [1]. Körner's authors say the behavioral effect "almost disappeared" after correcting for publication bias [11].

The footnote: this is a correction by a statistical model, and bias corrections can be wrong. I hold the "dead" label at about 0.9 confidence, not 1.0.

The retraction

In September 2016, Dana Carney, the original lead author, wrote: "I do not believe that 'power pose' effects are real" [3]. I will say this with some warmth: a researcher dropping her own result is rarer than it should be, and it is useful. But her statement came before the later meta-analyses and does not rest on them. It is a judgment about the original study's evidence.

Felt power: alive, small, and tangled

Here the story turns. Ranehill replicated the self-report effect even while the hormones and risk failed [2]. The Bayesian meta-analysis of six preregistered studies found "very strong evidence" for an effect on felt power [5]. Körner et al. reported a self-report effect of about g = 0.35 (a search summary of the abstract, with no interval available to me) [10][12]. Elkjær et al. found a pooled affect-and-behavior effect of g = 0.36 [9]. Körner's authors describe the effect on feelings as real but modest [11].

My working draft said "roughly d 0.2 to 0.3". I could not source that range, and the figures I could find sit a little higher, near 0.35. I am dropping my number and using theirs. I also have no interval for either one, so treat "0.35" as a point estimate with an unknown width. For scale: g = 0.35 means the average posed person scores about a third of a standard deviation above the comparison group. That is a small-to-moderate effect, about what many survey-style self-reports produce when people know what the study wants.

Sensitivity: which assumption moves the result most

The result moves most with the choice of comparison group. This is the point most readers miss.

Assumption 1: what is the control? Elkjær's meta-analysis split the comparison. Contractive (slumped) poses against a neutral control: g = 0.45. Expansive (power) poses against a neutral control: g = 0.06 [9]. Their conclusion is blunt: the effects are driven by the absence of contractive displays, not the presence of expansive ones [9]. So the headline 0.36 is a mix of two very different numbers.

Credé made the same logical point about the Cuddy p-curve reply: a negative effect of a contractive pose is not evidence for a positive effect of an expansive pose [8]. Körner's authors add the problem behind it: most studies had no neutral-position group, so the comparison is usually "power pose versus slump" [11]. That design cannot tell "power helps" from "slumping hurts".

Weighted by this, the claim "power posing makes you feel powerful" sits near zero against neutral (0.06, no interval available to me). The claim "slumping makes you feel worse than standing tall" has the support. That is a different sentence from the TED talk's.

Assumption 2: do participants know the hypothesis? In the Bayesian meta-analysis, evidence fell from very strong to moderate when only participants unfamiliar with the effect were kept [5]. The FORRT review flags demand effects, where people report what they think the study expects [12]. A self-report of "I feel powerful" is exactly the measure a demand effect would inflate. I cannot say how much of g = 0.35 is demand. I can say the direction of the bias is known.

Assumption 3: which p-curve do you trust? Simmons and Simonsohn found a flat p-curve on 34 studies, meaning no evidential value [6]. Cuddy, Schultz and Fosse used 55 studies and found clear evidential value, especially for feelings [7]. These are not two views of one dataset. They differ in study selection, and Credé argues the reply's result rests on the contractive-pose contrasts [8]. This one is open, and I will not pretend a winner. My lean is that p-curve answers "is anything real in here?" and the answer for feelings seems to be yes. It does not tell us the size or the cause.

Assumption 4: publication bias. Körner's authors say bias mattered for the behavioral difference between body displays and nearly erased it [11]. I do not have a bias-adjusted interval for the self-report effect, so I cannot say how much 0.35 shrinks. Given my general view that small-study meta-analyses overstate effects unless corrected, I would expect it to shrink somewhat. That is an expectation, not a finding.

Ranking by effect on the answer: the control group (0.45 versus 0.06) moves it most, then demand, then bias correction. The p-curve dispute moves the verdict on "is anything there" but not the size.

What is left standing

Hormone claim: dead, 42 people against 200 and several preregistered replications. Risk claim: dead, with a bias-model footnote. Felt power: alive at about g = 0.35 in meta-analyses, interval unknown to me. Against a neutral control, the expansive-pose version is near g = 0.06 and the slump version near g = 0.45, so the surviving piece is mostly "slouching makes people report feeling less powerful", and self-report may be partly demand.

I was wrong about one number in my own plan. I expected d of 0.2 to 0.3, and the published pooled figures sit near 0.35. If a preregistered study with a neutral control and a hidden hypothesis found expansive poses beating neutral by an interval clearly above zero, I would move felt power from "contested" to "survives". Until then, the honest verdict is partial survival of a smaller claim, not the claim anyone put on stage.

Readers still cite the TED version. Cite the Ranehill result instead [2].