Does Arousal Speed the Internal Clock During a Looming Threat? A Trial-by-Trial Test

When something rushes toward us, we tend to think it will hit sooner than it actually will. Psychologists usually explain this with an "internal clock": a pacemaker that emits ticks and a counter that adds them up. Arousal is assumed to make the pacemaker tick faster, so a fixed stretch of time collects more ticks and feels longer, and a person bracing for impact reacts early. The evidence for this story is indirect. Arousal has been inferred from ratings people gave to static pictures, never measured while the object was approaching. Responses have been button presses, which arousal could speed up simply by making people jumpier. And the effects come from small, staged experiments whose effect sizes are likely inflated.

We propose one study that closes all three holes. We record pupil size, skin conductance, and heart rate on every trial of a looming task. We produce arousal without changing the image, using predictable versus unpredictable threat cues. We use a response format that does not depend on motor speed. Then we fit competing clock models to the data. The design has two stages: a small pilot fixes the analysis, and a preregistered confirmatory sample tests it, sized with a deliberately shrunken pilot estimate. The result is a direct test of whether the pacemaker's tick rate follows moment-to-moment physiological arousal. If it does, it also gives a fitted rule relating arousal to time distortion.

Core Claim

Our central prediction is that measured arousal on a single trial, rather than the stimulus category, determines how early a person thinks a looming object will arrive. The sign of the effect should reverse with predictability. When a threat is signalled in advance, arousal builds before the object appears, the pacemaker speeds up, and estimated time-to-collision (TTC, the time left before an approaching object would reach the observer) shrinks in proportion to pre-onset arousal. When the threat arrives without warning, attention is captured and ticks are lost, so estimates lengthen even though arousal is high.

This crossover extends a result from blank-interval judgments, where Opposing Subjective Temporal Experiences in Response to Unpredictable and Predictable Fear-Relevant Stimuli found that predictable threat stretches and unpredictable threat compresses perceived duration, to the looming case, which that paper did not test. If the crossover holds when motor speed is controlled, it is hard to explain with an attention-gate-only account, with a pure response bias, or with visual differences between pictures.

Relation to Prior Work

Two behavioural studies motivate the proposal. 对威胁刺激的碰撞时间估计 reports that people underestimate TTC for threatening approaching stimuli. The role of arousal in the estimation of time‐to‐collision of threatening stimuli matched arousal across threatening animals and appetitive images such as food and banknotes. The threat effect disappeared under this matching, which points to arousal rather than negative valence as the cause. That study also found no effect for angry faces.

Both studies measured arousal through ratings of static images. Neither recorded physiology during the approach, and neither fully equated low-level image properties across categories. We keep their paradigm and stimulus sets but change how arousal is measured and how it is manipulated.

The speed of our mental soundtracks supports the idea that internal tempo tracks everyday alertness, measured with wearables. It motivates treating arousal as a continuous, trial-varying quantity rather than a two-level factor. It does not connect tempo to TTC.

Two methodological papers supply methods. The staged-evidence view (pilot for sensitivity, confirmation for specificity) shapes the two-stage design. Correcting Human Labels for Rater Effects in AI Evaluation contributes item-response-theory corrections, which model each rater's personal leniency or severity. We apply these to the participant-level arousal ratings, so that a person's habit of rating everything as "very arousing" is not mistaken for a faster clock.

Robotic work treats TTC as an objective optical quantity. Biologically Inspired Collision Avoidance Without Distance Information does so explicitly. Reactive Collision Avoidance for Safe Agile Navigation adds a fixed sigmoid mapping from TTC to a safety weight. These papers are downstream consumers, not competitors. A fitted arousal-to-distortion law is the kind of empirical input such heuristics currently lack, but building a controller is out of scope here.

Research Program

On each trial, a participant sees an approaching object for about one second before it disappears. The participant then judges when it would have arrived. The object is shown against isoluminant backgrounds, and phase-scrambled versions are matched in spatial frequency and contrast. The images come from the Animal Pictures Dataset (threatening versus non-threatening animals), the arousal-matched food and banknote set, and NimStim faces, all as used in the arousal study above, plus IAPS images for replication against the predictability study.

Arousal is manipulated independently of the image. A cue signals that a loud aversive noise may follow on this trial (predictable threat), or the noise can occur without a cue (unpredictable), or there is no threat context at all. Crossing this manipulation with image category separates "the picture is a snake" from "the observer is aroused."

Physiology is recorded throughout. Pupil diameter is sampled at high rate and corrected for luminance-driven change using the scrambled-image trials. Skin conductance and heart-rate change serve as slower, luminance-independent checks. From these we compute a pre-onset arousal index and an early-approach arousal index for each trial.

Responses take two forms, in separate blocks. The first is the standard open-loop button press. The second is a two-alternative judgment: participants say whether a probe tone came before or after the moment the object would have arrived. From these choices we estimate a point of subjective arrival, which does not depend on how fast the person can move.

The analysis compares hierarchical generative models fitted to both response types. In the pacemaker model, tick rate is $\lambda_t = \lambda_0 \exp(\beta a_t)$, where $\lambda_0$ is the participant's baseline rate, $a_t$ is the trial's arousal index, and $\beta$ is the arousal gain. Predicted subjective arrival scales roughly as $1/\lambda_t$, because a fixed reference count is reached sooner when ticks come faster. The competing models are a constant-rate model with an attention gate that loses a fraction of ticks, a combined model, and a motor-bias model in which arousal shifts only when the button is pressed.

The pilot, with roughly two dozen participants, is used only to settle preprocessing, arousal-index definitions, and model parameterization. It also yields an effect-size estimate that we shrink toward zero before running the power analysis for the preregistered confirmatory sample. This is our explicit guard against the winner's curse typical of preliminary findings.

Baselines

The naive reference point is the no-distortion model. It predicts responses from objective TTC with participant-level scaling and noise, and it includes no arousal or category terms. Any model worth keeping must beat it on held-out participants.

The strongest existing account is the category-level model of The role of arousal in the estimation of time‐to‐collision of threatening stimuli, which predicts TTC bias from static arousal ratings of each picture category. We fit it to exactly the same confirmatory trials, with the same hierarchical structure and the same number of free parameters per participant as our pacemaker model.

The file does not give a reported number for that model's effect size or fit. We found no reported result on Lacuna for the TTC-underestimation magnitude in either threat-looming study. Our comparison therefore relies on held-out log-likelihood and on effect direction, not on matching a published value.

"Beating" this baseline means two things. First, the trial-level physiological model must improve held-out log-likelihood over the category model, both within image category and within threat-cue condition. Second, the arousal gain $\beta$ must have the same sign in the button-press and two-alternative response formats.

The attention-gate model and the motor-bias model receive identical priors and fitting budgets, and each must lose on its own diagnostic pattern, described in the next section.

The key ablation shuffles each participant's arousal indices across trials within a condition. This keeps each condition's arousal distribution but breaks the link between a given trial's physiology and that trial's response. If the shuffled model performs as well as the unshuffled one, the trial-level claim is wrong.

Why not just…?

Why not just use the static arousal ratings, as prior work did? Category ratings cannot tell whether the clock follows arousal or merely follows the category. The null result for faces in the prior study already shows that ratings and time distortion come apart. The shuffle ablation and the within-category analysis settle this directly: if the ratings were enough, trial-level physiology would add nothing.

Why not just an attentional gate with a constant-speed clock? This is a serious alternative. The predictability study argues against a gate-only account because predictable and unpredictable threat push perceived time in opposite directions (Opposing Subjective Temporal Experiences), but that evidence comes from blank intervals. A gate model can produce later responses under unpredictable threat. It cannot produce earlier responses that grow with pre-onset arousal under predictable threat when attention is held constant. If the fitted combined model assigns $\beta \approx 0$, we will report that the gate account wins.

Why not just say arousal makes people press the button sooner? This is the most likely confound in all prior TTC work, and the file contains no evidence against it. For that reason the two-alternative tone judgment is central, not optional. Under a motor-bias account, arousal effects appear in button presses and vanish in the tone judgments. Under a pacemaker account, they appear in both.

Decisive Checks

The main test is preregistered and has three parts. First, in the confirmatory sample, the pacemaker model's held-out log-likelihood must exceed that of the category, gate, and motor models. Second, the arousal gain $\beta$ must be positive in both response formats. Third, the predictability crossover must appear in the point of subjective arrival from the tone judgments, not only in response times.

A secondary check tests whether the effect survives low-level image control. If the arousal effect persists for phase-scrambled images under threat cues, it cannot be a spatial-frequency artifact.

A final check asks whether $\beta$ estimated from the pilot shrinks in the confirmatory sample. We report the size of that shrinkage as a measure of how much preliminary estimates in this literature should be discounted. Any of three outcomes would falsify the pacemaker account in its standard form: $\beta$ disappears in the two-alternative block, the shuffled model ties the unshuffled one, or no crossover appears.