Part 1: Reviewed as the clinical trial that it is, the power posing paper should never have been published.
Has too much already been written about Amy Cuddy’s power pose paper? The conversation should not be stopped until its focus shifts and we change our ways of talking about psychological science.
The dominant narrative is now that a junior scientist published an influential paper on power posing and was subject to harassment and shaming by critics, pointing to the need for greater civility in scientific discourse.
Attention has shifted away from the scientific quality of the paper and the dubious products the paper has been used to promote and on the behavior of its critics.
Amy Cuddy and powerful allies are given forums to attack and vilify critics, accusing them of damaging the environment in which science is done and discouraging prospective early career investigators from entering the field.
Meanwhile, Amy Cuddy commands large speaking fees and has a top-selling book claiming the original paper provides strong science for simple behavioral manipulations altering mind-body relations and producing socially significant behavior.
This misrepresentation of psychological science does potential harm to consumers and the reputation of psychology among lay persons.
This blog post is intended to restart the conversation with a reconsideration of the original paper as a clinical and health psychology randomized trial (RCT) and, on that basis, identifying the kinds of inferences that are warranted from it.
In the first of a two post series, I argue that:
The original power pose article in Psychological Science should never been published.
-Basically, we have a therapeutic analog intervention delivered in 2 1-minute manipulations by unblinded experimenters who had flexibility in what they did, what they communicated to participants, and which data they chose to analyze and how.
-It’s unrealistic to expect that 2 1-minute behavioral manipulations would have robust and reliable effects on salivary cortisol or testosterone 17 minutes later.
-It’s absurd to assume that the hormones mediated changes in behavior in this context.
-If Amy Cuddy retreats to the idea that she is simply manipulating “felt power,” we are solidly in the realm of trivial nonspecific and placebo effects.
The original power posing paper
Carney DR, Cuddy AJ, Yap AJ. Power posing brief nonverbal displays affect neuroendocrine levels and risk tolerance. Psychological Science. 2010 Oct 1;21(10):1363-8.
The Psychological Science article can be construed as a brief mind-body intervention consisting of 2 1-minute behavioral manipulations. Central to the attention that the paper attracted is that argument that this manipulation affected psychological state and social performance via the effects of the manipulation on the neuroendocrine system.
The original study is in effect, a disguised randomized clinical trial (RCT) of a biobehavioral intervention. Once this is recognized, a host of standards can come into play for reporting this study and interpreting the results.
All major journals and publishers including Association for Psychological Science have adopted the Consolidated Standards of Reporting Trials (CONSORT). Any submission of a manuscript reporting a clinical trial is required to be accompanied by a checklist indicating that the article reports that particular details of how the trial was conducted. Item 1 on the checklist specifies that both the title and abstract indicate the study was a randomized trial. This is important and intended to aid readers in evaluating the study, but also for the study to be picked up in systematic searches for reviews that depend on screening of titles and abstracts.
I can find no evidence that Psychological Science adheres to CONSORT. For instance, my colleagues and I provided a detailed critique of a widely promoted study of loving-kindness meditation that was published in Psychological Science the same year as Cuddy’s power pose study. We noted that it was actually a poorly reported null trial with switched outcomes. With that recognition, we went on to identify serious conceptual, methodological and statistical problems. After overcoming considerable resistance, we were able to publish a muted version of our critique. Apparently reviewers of the original paper had failed to evaluate it in terms of it being an RCT.
The submission of the completed CONSORT checklist has become routine in most journals considering manuscripts for studies of clinical and health psychology interventions. Yet, additional CONSORT requirements that developed later about what should be included in abstracts are largely being ignored.
It would be unfair to single out Psychological Science and the Cuddy article for noncompliance to CONSORT for abstracts. However, the checklist can be a useful frame of reference for noting just how woefully inadequate the abstract was as a report of a scientific study.
CONSORT for abstracts
Hopewell S, Clarke M, Moher D, Wager E, Middleton P, Altman DG, Schulz KF, CONSORT Group. CONSORT for reporting randomized controlled trials in journal and conference abstracts: explanation and elaboration. PLOS Medicine. 2008 Jan 22;5(1):e20.
Journal and conference abstracts should contain sufﬁcient information about the trial to serve as an accurate record of its conduct and ﬁndings, providing optimal information about the trial within the space constraints of the abstract format. A properly constructed and well-written abstract should also help individuals to assess quickly the validity and applicability of the ﬁndings and, in the case of abstracts of journal articles, aid the retrieval of reports from electronic databases.
Even if CONSORT for abstracts did not exist, we could argue that readers, starting with the editor and reviewers were faced with an abstract with extraordinary claims that required better substantiation. They were disarmed by a lack of basic details from evaluating these claims.
In effect, the abstract reduces the study to an experimercial for products about to be marketed in corporate talks and workshops, but let’s persist in evaluating it as an abstract as a scientific study.
Humans and other animals express power through open, expansive postures, and they express powerlessness through closed, contractive postures. But can these postures actually cause power? The results of this study confirmed our prediction that posing in high-power nonverbal displays (as opposed to low-power nonverbal displays) would cause neuroendocrine and behavioral changes for both male and female participants: High-power posers experienced elevations in testosterone, decreases in cortisol, and increased feelings of power and tolerance for risk; low-power posers exhibited the opposite pattern. In short, posing in displays of power caused advantaged and adaptive psychological, physiological, and behavioral changes, and these findings suggest that embodiment extends beyond mere thinking and feeling, to physiology and subsequent behavioral choices. That a person can, by assuming two simple 1-min poses, embody power and instantly become more powerful has real-world, actionable implications.
I don’t believe I have ever encountered in an abstract the extravagant claims with which this abstract concludes. But readers are not provided any basis for evaluating the claim until the Methods section. Undoubtedly, many holding opinions about the paper did not read that far.
Forty-two participants (26 females and 16 males) were randomly assigned to the high-power-pose or low-power-pose condition.
Testosterone levels were in the normal range at both Time 1 (M = 60.30 pg/ml, SD = 49.58) and Time 2 (M = 57.40 pg/ml, SD = 43.25). As would be suggested by appropriately taken and assayed samples (Schultheiss & Stanton, 2009), men were higher than women on testosterone at both Time 1, F(1, 41) = 17.40, p < .001, r = .55, and Time 2, F(1, 41) = 22.55, p < .001, r = .60. To control for sex differences in testosterone, we used participant’s sex as a covariate in all analyses. All hormone analyses examined changes in hormones observed at Time 2, controlling for Time 1. Analyses with cortisol controlled for testosterone, and vice versa.2
Too small a study to provide an effect size
Hold on! First. Only 42 participants (26 females and 16 males) would readily be recognized as insufficient for an RCT, particularly in an area of research without past RCTs.
After decades of witnessing the accumulation of strong effect sizes from underpowered studies, many of us have reacted by requiring 35 participants per group as the minimum acceptable level for a generalizable effect size. Actually, that could be an overly liberal criterion. Why?
Many RCTs are underpowered, yet a lack of enforcement of preregistration allows positive results by redefining the primary outcomes after results are known. A psychotherapy trial with 30 or less patients in the smallest cell has less than a 50% probability of detecting a moderate sized significant effect, even if it is present (Coyne,Thombs, & Hagedoorn, 2010). Yet an examination of the studies mustered for treatments being evidence supported by APA Division 12 ( http://www.div12.org/empirically-supported-treatments/ ) indicates that many studies were too underpowered to be reliably counted as evidence of efficacy, but were included without comment about this problem. Taking an overview, it is striking the extent to which the literature continues depend on small, methodologically flawed RCTs conducted by investigators with strong allegiances to one of the treatments being evaluated. Yet, which treatment is preferred by investigators is a better predictor of the outcome of the trial than the specific treatment being evaluated (Luborsky et al., 2006).
Earlier my colleagues and I had argued for the non-accumulative nature of evidence from small RCTs:
Kraemer, Gardner, Brooks, and Yesavage (1998) propose excluding small, underpowered studies from meta-analyses. The risk of including studies with inadequate sample size is not limited to clinical and pragmatic decisions being made on the basis of trials that cannot demonstrate effectiveness when it is indeed present. Rather, Kraemer et al. demonstrate that inclusion of small, underpowered trials in meta-analyses produces gross overestimates of effect size due to substantial, but unquantifiable confirmatory publication bias from non-representative small trials. Without being able to estimate the size or extent of such biases, it is impossible to control for them. Other authorities voice support for including small trials, but generally limit their argument to trials that are otherwise methodologically adequate (Sackett & Cook, 1993; Schulz & Grimes, 2005). Small trials are particularly susceptible to common methodological problems…such as lack of baseline equivalence of groups; undue influence of outliers on results; selective attrition and lack of intent-to-treat analyses; investigators being unblinded to patient allotment; and not having a pre-determined stopping point so investigators are able to stop a trial when a significant effect is present.
In the power posing paper, there was the control for sex in all analyses because a peek at the data revealed baseline sex differences in testosterone dwarfing any other differences. What do we make of investigators conducting a study depending on testosterone mediating a behavioral manipulation who did not anticipate large baseline sex differences in testosterone?
In a Pubpeer comment leading up to this post , I noted:
We are then told “men were higher than women on testosterone at both Time 1, F(1, 41) = 17.40, p < .001, r = .55, and Time 2, F(1, 41) = 22.55, p < .001, r = .60. To control for sex differences in testosterone, we used participant’s sex as a covariate in all analyses. All hormone analyses examined changes in hormones observed at Time 2, controlling for Time 1. Analyses with cortisol controlled for testosterone, and vice versa.”
The findings alluded to in the abstract should be recognizable as weird and uninterpretable. Most basically, how could the 16 males be distributed across the two groups so that the authors could confidently say that differences held for both males and females? Especially when all analyses control for sex? Sex is highly correlated with testosterone and so an analysis that controlled for both the variables, sex and testosterone would probably not generalize to testosterone without such controls.
We are never given the basic statistics in the paper to independently assess what the authors are doing, not the correlation between cortisol and testosterone, only differences in time 2 cortisol controlling for time 1 cortisol, time 1 testosterone and gender. These multivariate statistics are not very generalizable in a sample with 42 participants distributed across 2 groups. Certainly not for the 26 females and 16 males taken separately.
The behavioral manipulation
The original paper reports:
Participants’ bodies were posed by an experimenter into high-power or low-power poses. Each participant held two poses for 1 min each. Participants’ risk taking was measured with a gambling task; feelings of power were measured with self-reports. Saliva samples, which were used to test cortisol and testosterone levels, were taken before and approximately 17 min after the power-pose manipulation.
And then elaborates:
To configure the test participants into the poses, the experimenter placed an electrocardiography lead on the back of each participant’s calf and underbelly of the left arm and explained, “To test accuracy of physiological responses as a function of sensor placement relative to your heart, you are being put into a certain physical position.” The experimenter then manually configured participants’ bodies by lightly touching their arms and legs. As needed, the experimenter provided verbal instructions (e.g., “Keep your feet above heart level by putting them on the desk in front of you”). After manually configuring participants’ bodies into the two poses, the experimenter left the room. Participants were videotaped; all participants correctly made and held either two high-power or two low-power poses for 1 min each. While making and holding the poses, participants completed a filler task that consisted of viewing and forming impressions of nine faces.
The behavioral task and subjective self-report assessment
Measure of risk taking and powerful feelings. After they finished posing, participants were presented with the gambling task. They were endowed with $2 and told they could keep the money—the safe bet—or roll a die and risk losing the $2 for a payoff of $4 (a risky but rational bet; odds of winning were 50/50). Participants indicated how “powerful” and “in charge” they felt on a scale from 1 (not at all) to 4 (a lot).
An imagined bewildered review from someone accustomed to evaluating clinical trials
Although the authors don’t seem to know what they’re doing, we have an underpowered therapy analogue study with extraordinary claims. It’s unconvincing that the 2 1-minute behavioral manipulations would change subsequent psychological states and behavior with any extralaboratory implications.
The manipulation poses a puzzle to research participants, challenging them to figure out what is being asked of them. The $2 gambling task presumably is meant to simulate effects on real-world behavior. But the low stakes could mean that participants believed the task evaluated whether they “got” the purpose of the intervention and behaved accordingly. Within that perspective, the unvalidated subjective self-report rating scale would serve as a clue to the intentions of the experimenter and an opportunity to show the participants were smart. The manipulation of putting participants into a low power pose is even more unconvincing as a contrasting active intervention or a control condition. Claims that this manipulation did anything but communicate experimenter expectancies are even less credible.
This is a very weak form of evidence: A therapy analogue study with such a brief, low intensity behavioral manipulation followed by assessments of outcomes that might just inform participants of what they needed to do to look smart (i.e., demand characteristics). Add in that the experimenters were unblinded and undoubted had flexibility in how they delivered the intervention and what they said to participants. As a grossly underpowered trial, the study cannot make a contribution to the literature and certainly not an effect size.
Furthermore, if the authors had even a basic understanding of gender differences in social status or sex differences in testosterone, they would have stratified the study with respect to participate gender, not attempted to obtain control by post hoc statistical manipulation.
I could comment on signs of p-hacking and widespread signs of inappropriate naming, use, and interpretation of statistics, but why bother? There are no vital signs of a publishable paper here.
Is power posing salvaged by fashionable hormonal measures?
Perhaps the skepticism of the editor and reviewers was overcome by the introduction of mind-body explanations of what some salivary measures supposedly showed. Otherwise, we would be left with a single subjective self-report measure and a behavioral task susceptible to demand characteristics and nonspecific effects.
We recognize that the free availability of powerful statistical packages risks people using them without any idea of the appropriateness of their use or interpretation. The same observation should be made of the ready availability of means of collecting spit samples from research participants to be sent off to outside laboratories for biochemical analysis.
The clinical health psychology literature is increasingly filled with studies incorporating easily collected saliva samples intended to establish that psychological interventions influence mind-body relations. These have become particularly applied in attempts to demonstrate that mindfulness meditation and even tai chi can have beneficial effects on physical health and even cancer outcomes.
Often inaccurately described as as “biomarkers,” rather than merely as biological measurements, there is seldom little learned by inclusion of such measures that is generalizable within participants or across studies.
Let’s start with salivary-based cortisol measures.
A comprehensive review suggests that:
- A single measurement on a participant or a pre-post pair of assessments would not be informative.
- Single measurements are unreliable and large intra-and inter-individual differences not attributable to intervention can be in play.
- Minor variations in experimental procedures can have large, unwanted effects.
- The current standard is cortisol awakening response in the diurnal slope over more than one day, which would not make sense for the effects of 2 1-minute behavioral manipulations.
- Even with sophisticated measurement strategies there is low agreement across and even within studies and low agreement with behavioral and self-report data.
- The idea of collecting saliva samples would serve the function the investigators intended is an unscientific, but attractive illusion.
Another relevant comprehensive theoretical review and synthesis of cortisol reactivity was available at the time the power pose study was planned. The article identifies no basis for anticipating that experimenters putting participants into a 1-minute expansive poses would lower cortisol. And certainly no basis for assuming that putting participants into a 1-minute slumped position would raise cortisol. Or what such findings could possibly mean.
But we are clutching at straws. The authors’ interpretations of their hormonal data depend on bizarre post hoc decisions about how to analyze their data in a small sample in which participant sex is treated in incomprehensible fashion. The process of trying to explain spurious results risks giving the results a credibility that authors have not earned for them. And don’t even try to claim we are getting signals of hormonal mediation from this study.
Another system failure: The incumbent advantage given to a paper that should not have been published.
Even when publication is based on inadequate editorial oversight and review, any likelihood or correction is diminished by published results having been blessed as “peer reviewed” and accorded an incumbent advantage over whatever follows.
A succession of editors have protected the power pose paper from post-publication peer review. Postpublication review has been relegated to other journals and social media, including PubPeer and blogs.
Soon after publication of the power pose paper, a critique was submitted to Psychological Science, but it was desk rejected. The editor informally communicated to the author that the critique read like a review and teh original article had already been peer reviewed.
The critique by Steven J. Stanton nonetheless eventually appeared in Frontiers in Behavioral Neuroscience and is worth a read.
Stanton took seriously the science being invoked in the claims of the power pose paper.
Carney et al. (2010) collapsed over gender in all testosterone analyses. Testosterone conforms to a bimodal distribution when including both genders (see Figure 13; Sapienza et al., 2009). Raw testosterone cannot be considered a normally distributed dependent or independent variable when including both genders. Thus, Carney et al. (2010) violated a basic assumption of the statistical analyses that they reported, because they used raw testosterone from pre- and post-power posing as independent and dependent variables, respectively, with all subjects (male and female) included.
^Mean cortisol levels for all participants were reported as 0.16 ng/mL pre-posing and 0.12 ng/mL post-posing, thus showing that for all participants there was an average decrease of 0.04 ng/mL from pre- to post-posing, regardless of condition. Yet, Figure 4 of Carney et al. (2010) shows that low-power posers had mean cortisol increases of roughly 0.025 ng/mL and high-power posers had mean cortisol decreases of roughly 0.03 ng/mL. It is unclear given the data in Figure 4 how the overall cortisol change for all participants could have been a decrease of 0.04 ng/mL.
Another editor of Psychological Science received a critical comment from Marcus Crede and Leigh A. Phillips. After the first round of reviews, the Crede and Philips removed references to changes in the published power pose paper from earlier drafts that they had received from the first author, Dana Carney. However, Crede and Phillips withdrew their critique when asked to respond to a review by Amy Cuddy in a second resubmission.
The critique is now forthcoming in Social Psychological and Personality Science
Revisiting the Power Pose Effect: How Robust Are the Results Reported by Carney, Cuddy and Yap (2010) to Data Analytic Decisions
The article investigates effects of choices made in p-hacking in the original paper. An excerpt from the abstract
In this paper we use multiverse analysis to examine whether the findings reported in the original paper by Carney, Cuddy, and Yap (2010) are robust to plausible alternative data analytic specifications: outlier identification strategy; the specification of the dependent variable; and the use of control variables. Our findings indicate that the inferences regarding the presence and size of an effect on testosterone and cortisol are highly sensitive to data analytic specifications. We encourage researchers to routinely explore the influence of data analytic choices on statistical inferences and also encourage editors and reviewers to require explicit examinations of the influence of alternative data analytic specifications on the inferences that are drawn from data.
Dana Carney, the first author of the has now posted an explanation why she no longer believes the originally reported findings are genuine and why “the evidence against the existence of power poses is undeniable.” She discloses a number of important confounds and important “researcher degrees of freedom in the analyses reported in the published paper.
Coming Up Next
A different view of the Amy Cuddy’s Ted talk in terms of its selling of pseudoscience to consumers and its acknowledgment of a strong debt to Cuddy’s adviser Susan Fiske.
A disclosure of some of the financial interests that distort discussion of the scientific flaws of the power pose.
How the reflexive response of the replicationados inadvertently reinforced the illusion that the original pose study provided meaningful effect sizes.
How Amy Cuddy and her allies marshalled the resources of the Association for Psychological Science to vilify and intimidate critics of bad science and of the exploitation of consumers by psychological pseudoscience.
How journalists played into this vilification.
What needs to be done to avoid a future fiasco for psychology like the power pose phenomenon and protect reformers of the dissemination of science.
Note: Time to reiterate that all opinions expressed here are solely those of Coyne of the Realm and not necessarily of PLOS blogs, PLOS One or his other affiliations.