Andy Edmonds · Leif Azzopardi
March 2026
Does mousedown→mouseup duration — the time a finger holds a button during a click — carry a cognitive signal?
Andy collected data on politically-charged factual questions via CrowdFlower (n=291, 2015). Leif collected data on perceptual speed tests via Prolific using the treconomics platform (n=227, 2022). Both datasets measured click hold duration alongside task variables.
There is signal. It's small in normal interactions — a 7ms task effect between baseline clicking and word search, a 4ms expertise × question-type interaction — but it's real and it replicates across independent samples.
It may be too small to matter in practice, given the confounds that dwarf it:
- Input device: trackpad vs. mouse produces a 16ms gap — biomechanically grounded, larger than most cognitive effects.
- Age: a clear gradient from 87ms (25–34) to 137ms (65+), reflecting motor slowing.
- Individual motor habits: within-person SD of 30–50ms swamps between-condition differences.
Click duration is correlated with the harder-to-measure pre-click behavior — approach dynamics, dwell time, cursor trajectory — that does carry strong signal. Pre-click dwell discriminates correct from incorrect selections by 91ms (p < 10⁻²⁵). Click duration alone doesn't. But the two are measuring adjacent phases of the same decision pipeline: evaluation (pre-click) → commitment (during click). Maybe the combination gets us somewhere that neither achieves alone.
The most interesting finding is what click duration is independent from:
- NASA-TLX: Self-reported cognitive load doesn't correlate with click duration. Whatever the click is measuring, people can't introspect on it.
- Correctness: Click duration doesn't predict whether a lexical decision was right or wrong — but pre-click dwell does. These are different signals at different stages.
This independence is actually encouraging. It means click duration isn't just a noisy proxy for something we already measure well. It's accessing a different layer — pre-reflective motor hesitation at the moment of commitment, not deliberation beforehand or evaluation afterward.
The strongest effects come from identity-relevant content, not task difficulty. Non-voters hold 19ms longer on "I will vote in the next election" than on "I voted in the last election." Self-reported political experts differentiate facts from opinions in their click latency; non-experts don't. These are small-N observations (n=19–42 per cell), but the direction is consistent and the theoretical framing — commitment uncertainty, not cognitive load — makes sense of when the signal appears and when it doesn't.
Pre-click dwell time is where the strong cognitive signal lives (91ms correct/incorrect discrimination). But dwell time is hard to instrument reliably: you need to define "arrived at target" from continuous cursor movement, handle hover-without-intent, distinguish reading from aiming, and deal with variable target sizes. It's a post-hoc analysis metric, not something you can cheaply capture in production.
Click hold duration — mousedown to mouseup — is the opposite. Trivial to instrument (two event listeners), sub-millisecond precision via performance.now(), works on every element, no heuristics needed. The signal is small but real, and critically, it's stable enough to serve as a within-subject baseline.
The approach dynamics extension is the bridge. By capturing cursor velocity in a ring buffer during normal mousemove events, we get a practical approximation of the pre-click approach phase: deceleration profile, course corrections, pause-before-commit. This isn't dwell time exactly, but it's instrumenting the same motor phase that dwell time measures — and it's just as easy to deploy as click duration itself. The hope is that click hold + approach velocity together recover enough of the pre-click decision signal to be useful, without the instrumentation burden of true dwell time measurement.
The core library ships at 2KB and is deployed in production across 18 blog posts on scrutinizer.app. Data flows to PostHog alongside ReadingDepth (paragraph-level absorption tracking) and standard analytics.
| Signal | Library | What it captures |
|---|---|---|
| Click hold duration | ClickSense | mousedown→mouseup latency (ms), target element, drag filtering |
| Approach dynamics | ClickSense | Pre-click cursor velocity, deceleration, corrections, pause duration |
| Paragraph absorption | ReadingDepth | Per-paragraph dwell time vs expected read time (238 WPM baseline) |
From 117 events with approach data across 4 hold duration buckets:
- Deceleration is monotonic across hold buckets: -0.0011 (quick) → -0.0052 (deliberative). Slower clicks approach more carefully.
- Deliberative clicks (120-160ms) show a 316ms pre-click pause — 2.4x longer than other buckets. This is decision cost, not motor cost.
- Course corrections peak in the "normal" bucket (12.0), not deliberative. Normal-speed clicks course-correct more, suggesting active aiming. Deliberative clicks arrive slowly and pause — they don't aim harder, they wait.
- PostHog dashboard: Hold Duration × Approach Dynamics
- HogQL queries:
queries/analysis.sql— 8 queries (duration distributions, per-user baselines, target breakdown, session drift, approach dynamics) - Library source: github.com/andyed/clicksense
The question remains whether click duration + approach dynamics, combined, produce a practical signal that neither achieves alone — and whether within-subject baselines and identity-relevant contexts strengthen it enough for a contribution.
Recent work shows that gaze and mouse cursor aren't redundant — the relationship between them carries cognitive signal that neither captures alone:
-
Stone & Chapman (2023) — "Unconscious Frustration: Dynamically Assessing User Experience using Eye and Mouse Tracking." Proc. ACM Human-Computer Interaction (ETRA). Eye-mouse coordination breakdown reveals UX friction points that neither signal detects independently. Webcam eye tracking sufficient. "Quantified coordination of unconscious behaviors" successfully identified friction with "minimal cost." doi:10.1145/3591137
-
Zhu, Shi, Song & Cai (2023) — "Integrating Gaze and Mouse Via Joint Cross-Attention Fusion Net for Students' Activity Recognition in E-learning." Proc. ACM IMWUT, 7(3). Cross-attention fusion of gaze + mouse achieves 94.87% F1 on 8-class activity recognition, 7.44% improvement over either signal alone. The attention mechanism learns when to trust which modality. doi:10.1145/3610876
-
Huang, White & Buscher (2012) — "User See, User Point: Gaze and Cursor Alignment in Web Search." CHI '12. Cursor and gaze diverge during reading and scanning — the cursor trails, leads, or parks while the eyes continue. doi:10.1145/2207676.2208591
-
Chen, Anderson & Sohn (2001) — "What Can a Mouse Cursor Tell Us More? Correlation of Eye/Mouse Movements on Web Browsing." CHI '01 Extended Abstracts. Mouse position correlates with gaze during browsing (CMU). doi:10.1145/634067.634234
Connection to ClickSense: These papers characterize the gaze-cursor delta during the evaluation phase (scanning, reading, foraging). ClickSense captures the commitment phase (click dynamics, approach deceleration). Together they cover the full decision arc: evaluate (gaze leads cursor) → decide (gaze-cursor converge) → commit (click hold + approach dynamics). The combination — pre-click gaze-cursor coordination + click-moment motor behavior — may recover more of the decision signal than either approach alone.
Data: Edmonds (CrowdFlower, 2015), Azzopardi & Edmonds (Prolific/treconomics, 2022). Library: github.com/andyed/clicksense. Production deployment: scrutinizer.app/blog