A chart's vertical limits were set to 0-100, yet 3% of rows exceed 100 — where did those rows go?

## The limits are a property of the picture **The axis range** — the lowest and highest values the axis shows — is a presentation setting. It can be narrower than the data, and when it is, some rows have no mark inside the drawn region. Those rows are still in the table, still in memory, still in whatever is written out afterwards. Nothing about setting a limit is a deletion. What makes this a real production hazard rather than a curiosity is that the resulting picture is indistinguishable from an honest one. The axis stops at 100 because that is the limit; a chart of data that genuinely stops at 100 also has an axis stopping at 100. The page carries no mark, no annotation and no count to tell the two apart. ## Two mechanisms, one picture | | narrowing the window | removing out-of-range rows first | |---|---|---| | where the rows live | still in the data and still considered | still in the data, but not in what the chart saw | | the clipped marks | not drawn inside the region | not drawn, because the rows never reached the mark | | anything the chart derived | computed over every row, including the ones off the picture | computed over the survivors only, so it moves | | what the reader sees | an axis stopping at 100 | an axis stopping at 100 | | what the tool may say | possibly nothing | often a message about removed rows, easy to miss | The second row of consequences is the one that catches people. Suppose a reference line is drawn across the chart for the column's average. Under the first mechanism it sits where the true average is, possibly above the top of the frame and therefore invisible. Under the second it sits somewhere lower, because the largest three percent of rows were not part of the calculation. A colleague who computes the average from the table then reports a number the chart disagrees with, and the disagreement is the diagnostic. ## Why this is harder to catch than a filter you wrote A condition applied to the rows lives in the code that prepares the data, next to everything else that shapes the result, where a reviewer reading the transformation will see it. An axis limit lives in the presentation step, often far from the data work and often set to make an early draft look tidy after one dominating value squashed everything else. It then survives every later change to the data, including the change that made the dominating value normal and something else the new extreme. The limit is also invisible to the checks people usually run. A row count before and after the chart is unchanged, because the chart does not consume rows. Whatever assertions guard the transformation all pass. The only surface on which the problem exists is the picture. ## How to catch it 1. **Compare the highest tick with the column's maximum**, computed from the table. If the tick is lower, rows are outside the picture. Do the same at the bottom. 2. **Read the messages the drawing produced.** Where a design removes rows rather than clipping marks, it usually says so once, in a stream nobody is watching. Where it clips, it may say nothing at all, so silence is not evidence. 3. **Cross-check one number.** Take any summary the chart shows and compute it from the table. Agreement suggests a narrowed window; disagreement points at removal before drawing. 4. **Do not expect clipped marks to pile up at the edge.** Most designs simply do not draw them, though some offer an explicit out-of-range indicator you have to ask for. An empty top edge means nothing either way. ## When clipping is legitimate, and what honesty requires There is a real case for it: a handful of very large values can compress everything else into a band a few pixels tall, and a chart that shows the structure of ninety-seven percent of the rows is more useful than one that shows a flat line and three dots. The limits are then a deliberate editorial choice, and the obligations that come with it are concrete: - say in the caption what the limits are; - say **how many rows** fall outside them and how far they reach, so the reader knows the size of what is missing; - consider drawing the extremes some other way rather than hiding them — an axis on which equal distances mean equal ratios, a non-linear axis where a fixed step across the page is a fixed multiplier rather than a fixed amount, will often fit the whole column in one frame without cutting anything; - re-check the limits whenever the data changes, because a limit chosen for last quarter's distribution is a silent editorial decision about this quarter's. The failure mode to remember is not that clipping is wrong. It is that a clipped chart makes exactly the same claim about completeness as an unclipped one, and only the author knows which it is.