INTERACTIVE MODELFIELD NOTEMETASCIENCE

The Sleeping-Side Bench

A live model accompanying “Two-Thirds of Cats Sleep on Their Left Side, Until Somebody Rechecked”

Study: Field note
Review: Peer-edited by the club review board
Article: 13 sectionsabout 18 min
Files: LaTeX source22 kB

Everything in this dispute comes down to one comparison: how many cats fell left, how many fell right, and how surprising that split is against a coin flip. Model 1 below loads the actual published counts from both papers and runs the same chi-square test on them live, so you can watch why one dataset comes out p < 0.001 and the other four don't. Model 2 is ours: flip your own virtual cats from a population you control, and see how often pure chance produces a result that looks, on paper, like a discovery.

χ² = Σ (observed − expected)² / expected expected = n / 2 each side p = erfc(√(χ²/2))

Model 1. Run the real studies

Pick a dataset. The bars and the statistic below them are computed from the published left/right counts, not simulated — this is the actual arithmetic behind Figure 2 in the field note, laid out so you can check it yourself.

Dataset

n: 301left: 192right: 109% left: 63.8%χ²: 22.90p: <0.001

Isparta et al. 2025: a departure from 50:50 this size, at this sample size, would happen by chance alone well under once in a thousand times.

* The Instagram arm reports a mean and standard deviation across accounts (t = 0.836, p = .407) rather than a single pooled count; the bars shown for it use the reported mean percentage applied to n = 60 as an illustrative equivalent count, not the paper's own test statistic.

Model 2. Flip your own cats

This part is entirely the club's own simulation, not a re-analysis of either paper. Set a sample size and a true population-wide chance of landing left, then flip. Each click draws that many independent virtual coin flips and drops one mark on the strip below at the resulting percentage left. Marks are grey when the result isn't significant at the conventional 0.05 threshold, and coloured when it is — purely from the sampling noise you dialled in.

Sample size301True % left in the population50%Overlay a real result

this trial — left: –right: –% left: –χ²: –p: –

Pick a sample size and a true percentage, then flip.

Flips so far: 0
Flagged “significant” (p<.05): 0 (0.0%)
Expected by chance at this threshold: ≈ 5%, if the true % is exactly 50

Set true % left to 50 and sample size to 301, then flip a few dozen times. Almost none of your marks will land anywhere near the amber overlay line at 63.8% — which is the whole reason the original result reads as a real signal rather than noise. Now drop the sample size to 20 or 30 and watch how much wider the noise band gets on its own, with nothing about the underlying population changing at all.

What the two models are and aren't showing

Model 1 is arithmetic on real numbers: it will always agree with the field note, because it's running the same test on the same published counts. Model 2 is a coin, not a cat. It cannot tell you whether cats really lean left, why YouTube and Reddit and Instagram might sample different populations of photogenic cats, or what it means that only a repeated-observation design can measure an individual trait rather than a population snapshot. What it can show, honestly, is the part of the story that's easiest to lose sight of in the argument over whose result is “right”: sample size alone changes how much noise a true 50:50 population throws off, and a result's significance is never separable from how hard it was even possible to find one, one way or the other, at that n.