A chart draws spend and signups in one frame, each against its own vertical range — what does their apparent agreement establish?
answer
- who chose the two ranges
- a shift and a stretch each
- agreement is manufactured, not measured
- a derived axis asserts nothing new
basics
~20 sNothing about the data. Two vertical ranges chosen independently can make almost any two sequences appear to agree, because each range shifts and stretches its own sequence. The apparent match is a property of the chosen ranges, and the page never shows the choice.
solid answer
~50 sA second axis at its own scale means two sequences drawn in one frame, each read against its own vertical range, so their apparent agreement is a choice of ranges rather than a finding. Each range is a low value and a high value that someone picked, and mapping a sequence onto it shifts the curve and stretches it — two free parameters per sequence, which is enough to pin two features wherever you like. Align the peaks, or make both curves start and end together, and almost any pair will appear to track. What *does* survive is timing: the horizontal dimension is genuinely shared, so "this one turned before that one" is readable. The gaps, the crossings and the apparent co-movement are not. Say what varies by asking who chose each range and what would change if one were widened.
go deeper
Recall that two sequences on two separately chosen vertical ranges can be made to look like they agree. Treat the apparent match as a decision someone made about the ranges, not as something the data showed.
Explain the mechanism: each range contributes a shift and a stretch, two free parameters are enough to align two features, and the page never records which ranges were available. Separate the timing reading, which survives, from the gap and the crossing, which do not.
In review, ask who chose each range and what changes if one is widened, then propose the alternative — a mark per period placed by both measures — and state what that alternative gives up. Note that some designs refuse the arrangement on purpose.
The call worth owning is whether the arrangement is allowed in work that leaves the team at all. Banning it outright is cheap to enforce and costs you the legitimate timing comparison; allowing it with a disclosure rule costs review attention every week.
## What a second axis at its own scale is Two sequences are drawn in one frame, each read against its own vertical range — one curve against the left range, one against the right. The **mark** here is whatever is drawn for each row, a line or a bar, but the interesting object is the pair of ranges, because their apparent agreement is a choice of ranges rather than something measured. ## Why agreement is nearly always achievable 1. An axis range is a **low value and a high value**: two numbers someone picked, whether by hand or by a rule the tool applied. 2. Mapping a sequence onto that range **shifts** it and **stretches** it. Shift and stretch are two free parameters, and each sequence gets its own pair. 3. Two free parameters are enough to pin two features of a curve wherever you like: put both peaks at the same height, or make both curves start and end at the same level. So for almost any pair of sequences, a pair of ranges exists that makes them appear to track. The picture is evidence that such a pair exists. It is not evidence about the data. And because the page shows only the ranges that were used, never the alternatives that were available, a reader has no way to discount for the choice. Three consequences follow, and they are the ones a reviewer should say out loud: - **A crossing point means nothing.** Where the curves cross moves as soon as either range moves, so "signups overtook spend in August" is a statement about the axis settings. - **A widening gap means nothing.** The gap is the distance between two independently chosen mappings, measured in no unit at all. - **Timing does survive.** The horizontal dimension really is shared, so "this one turned two weeks before that one" is a legitimate reading and one of the few honest reasons to overlay at all. ## The honest alternatives | what the reader should see | what to draw | what it costs | |---|---|---| | whether the two move together | one point per period, positioned by both measures | the order through time disappears unless it is encoded somehow | | relative movement in comparable terms | both sequences on one range, each expressed as change from a common starting point | the original units are gone; only relative movement remains | | both levels, honestly | two charts sharing the horizontal dimension, one above the other | more vertical space, and the reader pairs positions by eye | The first is the one people skip and the one that answers the question actually being asked. If the claim is "these two move together", a mark per period placed by both measures states it directly and cannot be flattered by a range. ## When a second axis is defensible The rule is not "never two vertical axes" — it is **never two vertical axes chosen independently**. A second axis derived by a stated transformation of the first asserts nothing new: the same quantity in other units, or a count alongside its share of a fixed total. Both axes then describe the very same marks, there is no second mapping to choose, and no pair of ranges can be selected to flatter the picture. That is a labelling convenience, not a comparison. ## Where the designs disagree This is one of the places the ecosystems take genuinely different positions, and a candidate who knows only one will state a local behaviour as a law. Some interfaces hand you a twinned region that shares the horizontal axis and carries an independent vertical one, so the whole arrangement is a single call and looks like any other option. Others deliberately refuse an independent second scale and will give you only an axis derived by a stated transformation of the first. **That refusal is not a missing feature; it is a design position about what the chart would otherwise be asserting.** Knowing that the two designs disagree is more useful than memorising either one, because it tells you the arrangement is contested rather than standard. ## What an interviewer is listening for The answer that scores is mechanical, not moralising. Say that each range contributes a shift and a stretch, that two parameters per sequence are enough to manufacture agreement, and that the page does not record the choice. Then separate what survives — timing, along the shared horizontal dimension — from what does not: crossings, gaps and the impression of co-movement. Finish with the alternative you would draw and what it costs, and with the one case where a second axis is honest because it is derived rather than chosen.
- The two lines cross in the middle of the year. What does the crossing mean?Nothing on its own. The crossing point is wherever the two independently chosen mappings happen to coincide, and it slides as soon as either range is widened or narrowed. Since the two sequences are usually in different units, there is no sense in which one is "larger" than the other at that point. If the crossing is being read as an event, that is the clearest sign the arrangement has misled someone.
- Is anything on a two-axis chart safe to read?The horizontal dimension, because it is genuinely shared. Turning points, lead and lag, and whether a move in one sequence precedes a move in the other are all legitimate readings, and they are the only respectable reason to overlay. Everything vertical — the gap, the crossing, which curve is higher, how closely they agree — is a consequence of two range choices that the chart does not disclose.
Two rulers laid side by side, one marked in centimetres and one in inches, each stretched until its own marks happen to fill the page. Any two objects can then be made to look the same length. What the picture compares is the rulers, not the objects.
saying these in an interview costs you the question
- Reads the crossing of the two curves as an event
- Says the measures are correlated because the curves track
- Thinks a shared horizontal axis makes the ranges comparable
- Claims the chart is fine because both axes are labelled
- Cannot say what would change if one range were widened