Is "Relative Pressure" Choking a Measurement Artifact? Per-Person Deadline Curves for Expert Motor Failure

A recent result says that elite athletes keep their edge under a fixed deadline but fall apart when the deadline is scaled to their own fast baseline. That would mean expertise breaks at a relative pressure threshold, a limit set by each person's operating speed rather than by any absolute clock. The finding is surprising and influential, but the way it is produced invites a simpler explanation. Each deadline is computed from a noisy baseline measurement. People who looked fast at baseline were partly lucky, so a deadline built from their score is harder for them than intended, and they appear to "choke." On top of that, baseline speed, expertise group, and the relative deadline are nearly the same variable, so a regression model cannot tell their effects apart. The expert group is also a set of survivors, people who stayed in the sport for a decade, so any group difference mixes training with who was kept.

We propose to replace the single baseline-scaled deadline with a dense ladder of fixed deadlines given to every participant. From that ladder we fit each person's full curve of error rate against time allowed and read their pressure threshold off it as a parameter, not as a difference score. If expert failure under relative pressure survives this measurement, the relative-threshold claim becomes much stronger. If it disappears, a widely repeated story about expertise is largely regression to the mean. Either outcome changes how pressure studies should be designed.

What the file supports, and what it does not

The research file covers the pressure literature in detail and gives the textbook definition of the variance inflation factor (VIF). VIF measures how much a regression coefficient's variance grows because that predictor is predictable from the others: $VIF_i = 1/(1-R_i^2)$, where $R_i^2$ is the share of predictor $i$'s variance explained by the remaining predictors. The file is cut off partway through the VIF section. It contains no material on the limits of single-case studies or on retrospective selection bias, and it has no section on closest existing work. Those two directions therefore enter this proposal as methodological constraints drawn from the design of the cited pressure studies, not as literatures I can cite. I found no paper in the file that audits relative-pressure designs for regression-to-the-mean or collinearity artifacts. That is the gap claimed here, but I cannot rule out that such work exists outside the file.

The target finding comes from The Role of Time Constraints in Athletes' Egocentric Mental Rotation Performance. The study tested 24 elite divers and 23 non-athletes on a test where participants judge rotated images of a body. Divers were faster at baseline (896 ms vs. 1062 ms). Under absolute time pressure they kept lower error rates at the hardest rotation angles, rotating at 176°/s versus 163°/s. Under deadlines compressed relative to each group's own baseline, their error rates rose sharply. The file gives no numeric error rate for the relative condition. This paper motivates the proposal and is also the design we scrutinize.

Other pressure studies in the file supply outcome measures and show the same structural risks. Monkeys exhibit a paradoxical decrease in performance in high-stakes scenarios found choking in three rhesus monkeys only when rewards were both large and rare. Monkey N's success rate dropped 25.2%, and reaches undershot the target with prolonged corrective homing. The authors note that pre-movement and holding phases varied idiosyncratically across animals. This is the single-case problem in miniature: three animals with partly individual signatures. Limiting motor skill knowledge via incidental training protects against choking under pressure found an inverted-U failure at $20 stakes for explicitly instructed sequence learners and no failure for incidental learners. Accuracy rose from $5 to $10 and collapsed at $20. Stakes there are absolute, so the paradigm is a natural place to ask whether a per-person threshold exists. Back to feedback: aberrant sensorimotor control in music performance under pressure showed that pressure selectively harms pianists' timing when auditory feedback is delayed by 80 ms, while leaving isolated perception and tapping intact. Its mechanism, a rise in reliance on feedback, predicts a graded curve against pressure rather than a cliff at one deadline. Linear Predictive Coding for Acute Stress Prediction from Computer Mouse Movements supplies a continuous physiological-style outcome: the damping frequency of arm movement, a stiffness marker that rises under stress. That marker is reliable only on long movements.

Core Claim

The claim is that a relative pressure threshold should be estimated as a per-person parameter of a deadline–error curve fitted to fixed deadlines. Inferring it from performance at a single deadline scaled to a noisy baseline is the wrong approach. We model each person $i$'s error probability at deadline $d$ as

$$P(\text{error} \mid d) = \lambda_i + (1-\lambda_i)\,\frac{1}{1+\exp\big((d-\theta_i)/s_i\big)},$$

where $\lambda_i$ is the person's floor error rate with unlimited time, $\theta_i$ is the deadline at which errors climb halfway toward chance (their threshold), and $s_i$ is how abruptly they fail. "Expertise fails under relative pressure" then becomes a precise hypothesis. Experts should have a steeper collapse (smaller $s_i$), or a threshold $\theta_i$ that sits closer to their own baseline time $b_i$, than novices do. A difference in $\theta_i$ alone would not count, because it only says experts are faster.

The artifact we expect in the standard design follows from measurement error. The relative deadline is $d_i = c\,\hat b_i$, where $c$ is a fixed compression fraction and $\hat b_i = b_i + e_i$ is the measured baseline, equal to the true baseline $b_i$ plus noise $e_i$. With baseline reliability $\rho$ (the share of measured baseline variance that is true-score variance), the expected true baseline of someone measured at $\hat b_i$ is pulled toward the group mean by a factor of $\rho$. Participants who looked fastest therefore receive deadlines tighter than intended relative to their true speed. That shortfall is a group-level artifact whenever groups are selected on speed or differ in reliability.

Research Program

The program has three parts. First, a simulation study generates synthetic participants from the curve above, with known thresholds and realistic baseline reliability. It runs the original relative-deadline analysis and measures how often that analysis reports "expert relative choking" when none exists. In parallel it computes VIFs among expertise group, measured baseline, and relative deadline, to show when the joint regression cannot separate them. The simulation tells us what effect size the artifact alone can produce.

Second, a new behavioral study reruns the mental body rotation paradigm with divers and matched non-athletes. Each participant gets an untimed baseline split into two independent blocks. One block sets any relative deadline. The other serves as an independent baseline for analysis, which breaks the coupling that produces regression to the mean. The deadline ladder spans both groups' operating ranges. Each person's curve is fitted in a hierarchical Bayesian model that partially pools individuals within group. Every participant is thus treated as a replicated single case, and individual-level conclusions are still possible. This addresses the single-case limitation the monkey study illustrates.

Third, the same curve-fitting protocol is ported to two further paradigms from the file. One is the incidental-versus-instructed sequence task, using a ladder of monetary stakes in place of deadlines. The other is a mouse-reaching task with long trajectories, where damping frequency is a second outcome alongside error. The test is whether per-person thresholds behave consistently across a cognitive-motor task, a sequence task, and a biomechanical marker.

Retrospective selection gets a sensitivity analysis rather than a cure. We will recruit, where possible, former competitive divers who left the sport, and test whether their curves sit between current divers and non-athletes. If they do, part of the "expert" advantage reflects who survived rather than what training does.

Baselines

The naive reference point is a simple comparison of expert and novice error rates at each condition, with no baseline adjustment. It shows the raw pattern any explanation must account for.

The strongest existing method is the original relative-time-pressure analysis of The Role of Time Constraints in Athletes' Egocentric Mental Rotation Performance. It compresses deadlines by a fixed fraction of the group's baseline and compares error rates. The file reports its baseline speeds (896 ms vs. 1062 ms), rotation rates under absolute pressure (176 vs. 163 °/s), and decision times (457 vs. 515 ms). There is no reported numeric error rate for the relative condition on Lacuna. We will reimplement the analysis exactly on our new data, using the same participants, trials, and compression fraction, so the only difference is the analysis. Our method beats it if two conditions hold. In simulation, its false-positive rate for expert relative choking must exceed nominal levels while ours stays nominal. On real data, the split-baseline curve estimates must yield materially different, and more test–retest-stable, conclusions about experts.

The third baseline is ANCOVA, meaning analysis of the relative-condition errors with measured baseline entered as a covariate. It is the textbook fix a statistician would reach for.

The key ablation removes the split baseline: we fit the same curves but compute relative deadlines from the same baseline block used in analysis. Any change between the two versions isolates the contribution of breaking the coupling between deadline-setting and evaluation. A second ablation fits curves without hierarchical pooling, to show how much individual estimates depend on pooling.

Why not just…?

Why not just add baseline as a covariate? This is a real threat to the proposal's necessity. When groups differ systematically on a covariate measured with error, ANCOVA under-adjusts and still leaves group differences that are pure artifact. The groups here do differ, by 166 ms at baseline. The simulation compares ANCOVA directly to the split-baseline curve method across reliability levels. If ANCOVA recovers the true thresholds equally well at realistic reliabilities, we will report that the simpler fix suffices.

Why not just run a larger sample? Regression to the mean is a bias, not noise, so more participants make the wrong answer more confident. The simulation will show the artifact's size is unchanged as N grows.

Why not just drop the collinear variable when VIF is high? Dropping baseline or group because their VIF exceeds a conventional cutoff answers a different question. The scientific question is exactly how expertise and baseline speed jointly determine failure. We use VIF diagnostically, to show when the original design cannot identify the effect, not as a pruning rule.

Why not just use many absolute deadlines and compare groups? This alternative is close enough to plausibly work as well that the proposal is built on it. The deadline ladder is that design. What we add is per-person curve fitting, which turns the relative threshold into an estimable parameter ($\theta_i$ relative to $b_i$, and $s_i$). A plain group comparison at each deadline cannot say whether a given expert fails nearer their own limit.

Proposed Experiments and Decisive Checks

The first decisive check is the simulation. If, with no true relative-choking effect and baseline reliability in a plausible range, the original analysis reports expert relative choking at a rate comparable to a real effect, the published finding is not identifiable from that design. The second check is the rerun with divers: does the expert advantage in $\theta_i - b_i$ or in $s_i$ persist under split baselines? Persistence supports a true relative threshold. Disappearance, alongside a large drop between same-block and split-block estimates, supports the artifact account.

The third check is cross-paradigm. Per-person thresholds from the stakes ladder in the sequence task should show the instructed-learner collapse reported in Limiting motor skill knowledge via incidental training protects against choking under pressure as a steeper $s_i$, and incidental learners should show flat curves. In the reaching task, damping frequency from Linear Predictive Coding for Acute Stress Prediction from Computer Mouse Movements should rise in step with the error curve on long trajectories. The fourth check is test–retest reliability of $\theta_i$ and $s_i$ across sessions. If they are not reliable, per-person thresholds are not a usable construct, which would itself be a useful negative result.