Which visual channels can carry a continuous numeric column, and which can carry only a column of distinct categories?

## Structure has to match on both sides A **binding** is the statement that a **visual channel** — a property of a drawn mark such as position along an axis, length, area, colour or marker shape — varies with a given column. Whether that statement is well formed depends on a single question: does the channel have the same structure as the column? Columns come in two relevant shapes here: - **Continuous**, where values are ordered and there is always something between two of them — a price, a duration, a temperature. - **A set of distinct categories**, where values differ but nothing lies between them — a region, a payment method, a device kind. Channels have exactly the same two shapes. Position, length and area are ordered and infinitely divisible: halfway along an axis is a real value. Marker shape is neither ordered nor divisible: there is no shape between a triangle and a square, and no reason a square should be "more" than a triangle. ## Channel by channel | channel | continuous numeric column | column of distinct categories | |---|---|---| | Position along an axis | Yes — the default carrier | Yes — one slot per value | | Length from a common baseline | Yes | Only as the quantity measured per category | | Area or mark size | Yes, though read imprecisely | Poor — invents a magnitude the data lacks | | Graded colour | Yes — a reader can interpolate | Poor — nothing to interpolate between | | A set of distinct colours | No — a reader cannot interpolate | Yes, up to a handful | | Marker shape | No | Yes, up to a handful | Two things in that table are worth saying out loud: - **Colour is two channels sharing one name.** A graded colour scale carries a continuous column; a set of distinct colours carries a category column. Which of the two a binding produces depends on the column, and confusing them is one of the commonest binding errors. - **The unordered channels have a capacity limit.** A reader can hold a handful of shapes or distinct colours in mind. Past that they stop reading the channel and start ignoring it, so the binding is technically valid and practically dead. ## What actually happens when the structures do not match This is where designs in this family genuinely differ, and asserting one behaviour as the rule is a mistake: 1. **Some designs refuse the binding**, because the channel's scale type is checked against the column's type and a continuous column has no discrete scale to be fitted to. 2. **Some enumerate the distinct values instead**, treating your thousands of prices as thousands of categories — which either exhausts the available shapes or produces a legend as long as the dataset. 3. **Some accept it and quietly do something reasonable**, such as cutting the column into a few bands, which is a statistic nobody asked for and which is invisible on the page. The practical upshot is that you cannot tell from the finished picture which happened. If a shape binding rendered at all, check how many distinct shapes came out and whether that number bears any relation to the number of distinct values in the column. ## The reverse mismatch Binding a category column to a channel that is ordered and continuous is the mirror-image error and it is quieter, because it always renders. A set of regions carried by mark area gives every region a size, and a reader reads size as magnitude — so the chart asserts that one region is more than another when the column says nothing of the kind. Nothing errors; the picture simply makes a claim the data does not support. ## A short checklist before you bind - Is the column ordered, and are there values between its values? That decides which half of the table you are in. - Does the channel you picked have both properties too? If it has fewer, the binding loses information; if it has more, the binding invents some. - How many distinct values will reach an unordered channel? If it is more than a handful, the binding will render and not be read. - If the picture rendered but you are unsure which behaviour you got, count the distinct visual values against the distinct data values. ## What an interviewer is listening for That you explain the rule by structure rather than by memory — *shape has no order and nothing in between, so it cannot express more* — and that you handle the failure case honestly. A candidate who says "it throws an error" has learned one tool's behaviour; a candidate who says "that depends, and here is how I would find out which behaviour I have" has learned the concept.