Dual axes can fabricate correlation that does not exist

What it is

The previous leaf's four free parameters, in careless or calculating hands, become a forgery machine: want "the indicator improving together after the policy"? Tune both axes' endpoints until the lines converge after that date. Want "the rival's growth dragging us down"? Tune the other way and get scissors. Visual synchrony or divergence has no binding whatsoever to statistical correlation — two uncorrelated random walks can be tuned into "perfect coupling" in minutes of axis fiddling. Readers are exquisitely sensitive to lines co-varying (the Gestalt common fate), so the product of axis parameters gets read as a data relationship — the sharpest form of dual-axis misleading.

Why it happens

Co-variation perception is automatic: two lines rising and falling together, holding spacing, rarely crossing, and the impression of "relatedness" arises without reasoning — normally a reliable relational cue under a single axis (where co-variation under one scale is real); dual axes cut the link between cue and conclusion: perception fires as usual, though the precondition (a shared scale) is gone. Forgery is therefore cheap and hard to disprove after the fact: a reader checking the data finds a zero correlation while the chart still "looks" correlated, and the author can plead "the axis parameters were conventional." The only innocent use keeps the two lines visibly apart so that no co-variation perception arises — which has already given up the dual-axis "shared frame" selling point.

Where it stops holding

This leaf owns the mechanics of fabrication and its detection; recognizing deliberate deployment of the technique and its consequences belong to the ethics family. Detection routine: on seeing a dual-axis chart, check the two axes' units first (same units need no dual axis), then ask "does this chart survive different axis parameters?", and finally verify true correlation with a scatterplot — the scatterplot is the referee for correlation questions; a dual-axis chart never is. A reverse reminder: a dual-axis chart that "happens" to show a real correlation still proves nothing, because readers cannot distinguish coincidence from tuning.

Applying it

• Reading discipline: any relational conclusion from a dual-axis chart is downgraded to "to be verified," citable only after a scatterplot or correlation check.
• Making discipline: when the message is a correlation between two series, dual axes are banned; draw the scatterplot directly (encode time as colour gradient or trajectory if needed).
• Verification: take two series known to be uncorrelated (random numbers do) and let the team try to "manufacture" correlation by tuning axis parameters; manufacturable within five minutes completes the visceral training for this leaf.

Related

• Same group: U3.04.1 On dual axes, relative placement is an arbitrary choice · U3.04.3 Prefer facets over dual axes
• Nearby: U2.05.1 Scatterplots suit the relation between two continuous variables · U2.05.3 Correlation is not causation
• Search terms: spurious correlation · dual axis · common fate · fabricated trend