Ego Depletion: Two Multi-Lab Replications Found Nothing, and One Was Led by an Author of the Original

Almost every piece of advice about resisting your phone rests on an assumption most people have never examined: that self-control is a finite quantity, that it drains as you use it, and that this is why you fold at eleven at night having held out all day.

That assumption has a name, a founding experiment, and - unusually for psychology - two enormous preregistered tests designed specifically to check whether it is true.

The founding experiment

Baumeister, Bratslavsky, Muraven and Tice (1998), in the Journal of Personality and Social Psychology 74(5), 1252-1265, ran four experiments. The first is the one everybody remembers.

67 introductory psychology students (31 men, 36 women) sat in a room with freshly baked chocolate biscuits and a bowl of radishes. Some were told to eat the biscuits. Some were told to eat the radishes and leave the biscuits alone. A third group got no food. Everyone then attempted an unsolvable geometric puzzle, and the measure was how long they persisted before giving up.

Condition	Persistence
Radish (resisted the biscuits)	8.35 minutes
Chocolate	18.90 minutes
No food control	20.86 minutes

The people who had resisted temptation gave up in less than half the time. The authors named the phenomenon and defined it precisely:

“A temporary reduction in the self’s capacity or willingness to engage in volitional action (including controlling the environment, controlling the self, making choices, and initiating action) caused by prior exercise of volition.”

They were careful about what they had not shown. Their own limitation, stated in the paper: “we do not have a clear understanding of the nature of this resource”, and they “provided no direct measures of the limited resource and hence no direct evidence that some inner quantity is diminished.”

That caution did not survive contact with the popular version.

The literature grew, and then somebody checked it

By 2010 a meta-analysis of the field pooled 198 published tests and reported an effect of d = 0.62, 95% CI [0.57, 0.67] - large, and apparently secure.

Carter and McCullough (2014), in Frontiers in Psychology 5, article 823, asked whether that pooled figure survived correction for small-study effects. Trim-and-fill brought it down to roughly d = 0.48-0.50. PET regression, which the authors judged the most accurate of the methods they applied, produced an intercept of b₀ = -0.10, not statistically significant.

Their conclusion:

“After controlling for the influence of small-study effects, our results do not support the claim that the depletion effect is meaningfully different from zero.”

They were honest about the method’s limits: they corrected statistically rather than by collecting unpublished studies, because many null findings were never written up and some may have been converted into positive ones through undisclosed analytic choices.

The first multi-lab test

Hagger, Chatzisarantis and 56 co-authors (2016), in Perspectives on Psychological Science 11(4), 546-573, ran the same protocol in 23 laboratories with 2,141 participants.

One detail forecloses the obvious objection. The protocol was not chosen by sceptics. The tasks came from Sripada, Kessler and Jonides (2014), and the paper records that the selection was made “based on the recommendation of Roy Baumeister” - the theory’s originator approved the test before it ran.

The result:

d = 0.04, 95% CI [-0.07, 0.15]

And in the authors’ words: “Results across 23 laboratories revealed small effect sizes for the ego-depletion effect on the primary and secondary dependent variables, with confidence intervals for most laboratories including zero.”

They report their own limitations plainly: task duration may moderate the effect, exclusion rates were high because of accuracy criteria, and effect sizes were moderately heterogeneous across labs despite the standard protocol.

The second one, run by the other side

The response to 2016 was that the letter-crossing task was a weak manipulation. So a second attempt was built to answer that objection.

Vohs, Schmeichel and colleagues (2021), in Psychological Science, ran 36 laboratories with 3,531 participants. Rather than replicate one published study, the depletion tasks were chosen by crowdsourcing among depletion researchers and the participating labs. Two protocols were used: a letter-crossing task with an unsolvable figure-tracing outcome (20 labs, n = 1,852), and a constrained-writing task with a cognitive estimation outcome (16 labs, n = 1,679).

Outcome	Effect
Overall	d = 0.06, 95% CI [-0.02, 0.14]
Figure-tracing duration	d = 0.15 [0.02, 0.29]
Figure-tracing attempts	d = 0.05 [-0.06, 0.17]
Cognitive Estimation Test	d = 0.01 [-0.10, 0.12]

A Bayesian analysis found the data 4.4 times more likely under the null hypothesis than under the depletion hypothesis. The authors’ summary: preregistered analyses “did not find evidence for a depletion effect”, and depletion “is not as reliable or robust as previously assumed.”

The lead author is Kathleen Vohs, a co-author of the 1998 paper that founded the field.

The authors disclose that 30.25% of participants were excluded under the preregistered criteria - a large fraction, and one they flag themselves.

The case for the defence, which is not nothing

This is a live dispute, not a closed one, and the strongest counter-argument deserves stating properly.

Dang (2017), in Psychological Research, meta-analysed 142 effect sizes and found a pooled g = 0.38, 95% CI [0.31, 0.45], falling to g = 0.24 [0.16, 0.32] after trim-and-fill. His central methodological point against Carter and McCullough is that they pooled across depleting tasks that differ enormously, which “conceals a more accurate estimation”, and that PET-PEESE is unreliable given the heterogeneity present (I² = 60.67%).

On the 2016 replication he argues that “cautious attention has to be paid to the effectiveness of the depleting task” used, noting that the letter-crossing task captured only one of the three features of the original manipulation.

His own conclusion is, however, more measured than the field’s summaries of it: “it is not adequate to draw a strong conclusion from the current analysis that the ego depletion effect exists.”

One caution on his subgroup results. Dang reports much stronger corrected estimates for a subset of “reliable” depleting tasks. Selecting the tasks that worked and then estimating the effect within that subset is a procedure that will tend to find an effect whether or not one is there, and we are reporting it as suggestive rather than as a settled figure.

What this changes about resisting your phone

The reason this matters here is that the depletion account is doing quiet work in a lot of digital-wellbeing advice. Save your willpower. Do the hard thing in the morning. You gave in at midnight because you had spent the day’s supply.

If the resource does not behave the way the model says, then that advice is resting on something that two of the largest coordinated experiments in psychology could not detect.

What survives is narrower and, awkwardly, more useful. Self-control failures are real. What is not established is that they are caused by a general reservoir being emptied. And that distinction points somewhere specific: away from rationing an inner quantity and towards changing what is in front of you. This site has documented that the interventions with the better evidence are environmental rather than motivational - reducing the number of available switches rather than steeling yourself against them. That is the basis on which checking behaviour and screen-time reduction were assessed, and it is why the abstinence-versus-restriction evidence in the digital detox literature came out the way it did.

Limitations and disputes

• This is unresolved, not settled. Dang’s 2017 meta-analysis reports a small positive effect after correction. Two multi-lab preregistered studies did not.
• Both replications exclude large fractions of participants - the 2021 study 30.25%, the 2016 study a high proportion on accuracy criteria. Both disclose it; both have been criticised for it.
• Neither replication tested every version of the theory. Depletion has been operationalised dozens of ways, and a null on two protocols is not a null on all of them. This is Dang’s argument and it is a fair one.
• The 1998 authors stated at the time that they had no direct measure of the supposed resource. That has still not changed.
• We have not covered glucose. The claim that depletion is caused by falling blood glucose has its own literature and its own problems, and is not assessed here.
• Nothing here says self-control does not fail. It says the reservoir explanation for why it fails has not held up under coordinated testing.
• This is not advice about managing willpower and does not tell anyone to stop using strategies that work for them.