A common claim in time-perception research is that attention changes where people put the boundary between "short" and "long" without changing how sharply they can tell durations apart. In psychophysics terms, attention shifts the point of subjective equality (the duration judged "long" half the time) but leaves the Weber ratio unchanged. The Weber ratio is the slope-based measure of timing precision. The claim matters for theory. If attention only moves a decision boundary, then cross-domain effects such as large digits making durations seem longer are criterion shifts, not changes to sensory timing. Yet the evidence for "no precision change" rests on null results in small samples.
Those small-sample nulls are weaker than they look. The two parameters are not measured equally well. A psychometric curve's midpoint is pinned down by relatively few trials. Its slope is much noisier per participant, and it is easily distorted by lapses, meaning occasional inattentive responses. A null result on the slope from two dozen participants can therefore reflect measurement weakness rather than an absent effect. This proposal asks whether the "bias but not blur" dissociation survives a design built to detect a precision change if one exists. It also asks how participants and trials should be divided between few-people-many-sessions and many-people-one-session designs to answer that question.
Core Claim
The accuracy/precision dissociation reported for attentionally mediated number–time interactions is currently uninterpretable on the precision side. A design can settle it decisively at fixed total trial cost by combining three elements. First, probe durations are placed adaptively where they carry information about the slope rather than the midpoint. Second, a hierarchical psychometric model estimates lapses separately from slope. Third, the study crosses a deeply sampled cohort with a broadly sampled one. A built-in positive control, a manipulation expected to degrade timing precision, will show whether the design can detect a precision change at all.
The new contribution is not "run more participants." It is a measurement-first test of a specific null claim. The design also separates two ways attention could act on duration judgments: by moving the criterion, as Shukla's attentional account implies, or by trading off sensory precision across moments, as the temporal attention literature shows for visual sensitivity.
The Tension in the Literature
Two lines of work disagree about what attention does, and they use different measures.
On one side are the Shukla–Bapi experiments, which motivate this proposal. Numerical Magnitude Affects Accuracy but Not Precision of Temporal Judgments used a 550 ms standard with 250–850 ms comparisons and the digits 1, 5 and 9 in N = 24 participants. It reports a midpoint of roughly 523 ms for "9", 554 ms for "1" and 582 ms for "5", with a main effect at p < 0.001. For the Weber ratio it reports no difference (p > 0.05, Bayes factor $B_{10}$ = 0.146). Attention mediates the influence of numerical magnitude on temporal processing showed that the dilation appears for positive digits when trials are blocked, vanishes for negative digits, and turns into a positive-versus-negative bias when trials are intermixed. The authors argue from this that attention shifting along a mental number line, rather than shared magnitude hardware, drives the effect. Number-time interaction: Search for a common magnitude system in a cross-modal setting adds that the interference appears only when attention is directed to both dimensions at once. All three are behavioral studies of conventional size. None reports a precision effect size or an equivalence bound, and no reported result found on Lacuna gives a Weber-ratio effect size for these manipulations.
On the other side, Attention flexibly trades off across points in time shows that voluntary temporal attention raises perceptual sensitivity (d′) at the attended moment and lowers it at neighboring moments. Anticipatory and evoked visual cortical dynamics of voluntary temporal attention traces this to MEG signatures in N = 10, including a roughly 157° phase shift of a 2 Hz modulation and a post-target gain increase for the first target only. These results concern orientation discrimination, not duration categorization. If the same trade-off operated on interval timing, attention should change precision. That would contradict the Shukla interpretation unless magnitude-induced attention and cued temporal attention are different processes.
Other work constrains the design. The Effects of Subthreshold Autistic Traits On Time Perception found Weber-ratio group differences in a 400/1600 ms bisection task with N = 63 split into extreme groups. Precision differences are therefore detectable when large, but the study's dichotomization discards the continuous variation a broad sample could model. Expansion and Compression of Time Correlate with Information Processing in an Enumeration Task shows that duration expansion tracks enumeration accuracy only outside the subitizing range. A dual task can therefore change processing efficiency, which makes it a candidate precision manipulation.
The sample-size literature supplies the logic. Scaling laws for language encoding models in fMRI is a study of how brain-prediction accuracy grows with model size and scan hours. It reports that a tenfold increase in per-subject fMRI data improved encoding performance by 122%, from N = 3 deeply scanned subjects. Its population generalizability is untested. A functional connectome phenotyping dataset is a 194-person brain-imaging and behavior dataset. It documents how small cohorts produced inconsistent brain–behavior links. We borrow the deep-versus-wide framing and apply it where it can be tested cheaply and exactly: per-person psychometric parameters.
Research Program
The model is a psychometric function with an explicit lapse term:
$$p(\text{long}\mid d) = \tfrac{\lambda}{2} + (1-\lambda)\,\Phi\!\left(\frac{d-\mu}{\sigma}\right).$$
Here $d$ is the comparison duration, $\mu$ is the point of subjective equality, $\sigma$ is the spread (with Weber ratio $\sigma/\mu$ up to a constant), $\lambda$ is the lapse rate, and $\Phi$ is the standard normal cumulative distribution. The Fisher information about $\mu$ peaks at $d \approx \mu$. Information about $\sigma$ peaks near $d \approx \mu \pm \sigma$ and vanishes at $\mu$, which follows from differentiating $\Phi$ with respect to each parameter. Comparison durations placed uniformly, or bunched near the standard as is common, therefore inform the midpoint well and the slope poorly. A drop in attention that adds lapses flattens the fitted curve and can masquerade as a precision change, or hide one.
The first stage is a simulation study. It will be anchored to the published Shukla 2021a design and to a range of plausible effect sizes, including no Weber-ratio effect. It asks three questions. How often does the standard per-subject logistic fit followed by an ANOVA and Bayes factor yield a $B_{10}$ near 0.146 when a true precision effect of a given size exists? What is the smallest precision effect detectable with 90% power under each allocation of participants and trials? How much does lapse-aware hierarchical modeling with adaptive, slope-targeted probe placement shrink that threshold at equal total trials?
We will preregister a smallest effect of interest for the Weber ratio, a 10% relative change chosen as the magnitude at which a shared-sensory-hardware account would be distinguishable. Inference will use two one-sided equivalence tests alongside the Bayesian posterior.
The second stage is the human experiment. It uses the Shukla 2021a stimuli (standard 550 ms, comparisons 250–850 ms, digits 1/5/9) plus the polarity conditions from 2021b (−1/−9), and adds two manipulations. The scientific manipulation is a Denison-style precue indicating which of two successive intervals will be judged. Under the temporal trade-off account, a valid cue should sharpen precision for the cued interval and blur it for the uncued one. Under the pure-criterion account, it should move only the midpoint. The positive control is a concurrent effortful enumeration load (6–14 dots, following Wutz et al.), which is expected to reduce timing precision. Failure to detect a precision cost there would mean the design is too weak, not that attention spares precision.
Two arms receive the same total trial budget. The deep arm has about 12 participants, each completing many lab sessions. The wide arm has about 300 participants, each completing one online session with the same adaptive procedure. Both arms are analyzed with one hierarchical model.
Baselines
The "do nothing" reference is the analysis pipeline of Numerical Magnitude Affects Accuracy but Not Precision of Temporal Judgments applied at its own scale: N = 24, a fixed set of comparison durations, separate logistic fits for each participant, and a repeated-measures test with a Bayes factor on the Weber ratio. Its reported numbers are the anchor: a midpoint of about 523 ms for "9" versus 554 ms for "1", and Weber-ratio $B_{10}$ = 0.146. We run it on a resampled N = 24 subset of the wide arm, and in simulation, so that it sees identical stimuli. Beating it means reaching a decisive verdict on the Weber-ratio question, either equivalence within the preregistered bound or a credible difference, where this pipeline is inconclusive at the same trial count.
The strongest competitor is a large-N replication with the standard analysis: the same fixed probes and per-subject logistic fits, but the full 300-person wide arm. This is what a skeptical reviewer would fund instead, and it can work, because it gains power by brute force. It receives the same participants and total trials. The proposal beats it if the adaptive hierarchical design reaches the same precision-effect detection threshold with substantially fewer total trials. A pre-stated target is at most half. The proposal also beats it if, at equal trials, it recovers simulated precision effects that the standard pipeline biases toward zero because lapses inflate the fitted spread.
Two ablations isolate the key ingredients. The first keeps the hierarchical model but drops adaptive placement, using uniform probes. The second keeps adaptive placement but fits the model without a lapse term. If either ablation matches the full method, that ingredient is unnecessary, and we will say so.
Why not just…?
Why not just run the original study with 300 people? This is a real threat, and it is the strongest baseline above. Its weakness is not power alone. Fixed probes clustered near the standard carry little information about the slope, and lapses that differ between conditions bias the fitted Weber ratio regardless of N. The simulation stage measures this bias directly. If simulation shows the bias is negligible and the large-N replication is as efficient, the contribution reduces to the replication plus equivalence testing, and we will report that outcome.
Why not just reanalyze the existing data with a Bayesian hierarchical model? We have no access to the original trial-level data. A reanalysis also cannot fix poor probe placement, because the information about the slope was never collected. Re-modeling recovers lapses but not missing slope information.
Why not just use duration reproduction, where variability is measured directly? Reproduction variability mixes perceptual timing noise with motor noise. Work such as The temporal transition zone shows that reproduction also carries pulls toward a set-point that categorization tasks do not. Switching tasks would change the question the Shukla claim answers, which concerns categorization precision.
Why not just image the brain, as in the MEG work? Neural studies at N = 10 inherit the same problem of a noisy per-person parameter, with even fewer participants. Establishing whether the behavioral precision effect exists comes first. It also tells future imaging studies how to allocate subjects and sessions.
Decisive Checks
The design works if the enumeration load produces a credible Weber-ratio increase in both arms. Without that, no precision null can be interpreted.
The scientific question is settled in favor of the criterion account if the digit and polarity manipulations reproduce the midpoint shifts reported by Shukla and Bapi while their Weber-ratio effects fall inside the preregistered equivalence bound. It is settled against that account if a credible precision effect appears. For the temporal precue, a precision gain for cued intervals together with a loss for uncued ones would extend the trade-off of Attention flexibly trades off across points in time to duration categorization. That would show that cued temporal attention and number-line attention act through different routes.
Generalizability is tested by asking whether the deep arm's population-mean estimate falls inside the wide arm's credible interval. A second check is whether between-person variation in the effects, estimated only in the wide arm, is large enough that a 12-person cohort would routinely misestimate the mean. This is the behavioral analogue of the untested cross-subject transfer in deep-phenotyping work.
References
- Numerical Magnitude Affects Accuracy but Not Precision of Temporal Judgments
- Attention mediates the influence of numerical magnitude on temporal processing
- Number-time interaction: Search for a common magnitude system in a cross-modal setting
- Attention flexibly trades off across points in time
- Anticipatory and evoked visual cortical dynamics of voluntary temporal attention
- The Effects of Subthreshold Autistic Traits On Time Perception: The Regulation of Interpersonal Information Association
- Expansion and Compression of Time Correlate with Information Processing in an Enumeration Task
- Scaling laws for language encoding models in fMRI
- A functional connectome phenotyping dataset including cognitive state and personality measures
- The temporal transition zone: A gradual approach to a subjective set‐point within the three‐second time window