There was some discussion in the comments thread here about axis labels and zero. Sometimes zero has no particular meaning (for example when graphing degrees Fahrenheit), but usually it has a clear interpretation, in which case it can be pleasant to include it on the graph. On the other hand, if you’re plotting something that varies from, say 184 to 197, it would typically be a bad idea to extend the axis all the way to zero, as this would destroy your visual resolution.
The advice we usually give is: If zero is in the neighborhood, invite it in. This came up on the blog once, but it was buried in a longer post, so I thought I’d give it a stand-alone here.
Good heuristic! Empirical data suggests that too narrow of a range for the y-axis causes the same problems as too wide a range (albeit in the opposite direction).
In Witt 2019, across bar graphs, line charts, and graphs with error bars, subjects estimates of effect size were impacted. Too much extra space, and subjects reported smaller effect sizes. To little extra space, and subjects reported larger effect sizes. The paper proposed a y-axis range 1.5 SDs of the data points.
https://open.lnu.se/index.php/metapsychology/article/view/895
This is a cute name, but there’s a problem with the premise. Bar charts of course should always have the y based at zero, because doing otherwise would distort the ratios. The same argument can extend to line graphs, whenever they omit the areas under their curves, which means anytime you are tempted to distort in this way.
There are exceptions to practice: we see medical oscilloscopes, stock charts, and certain other common graphs without a zero baseline. This might seem to argue that any line graph is okay without a zero. Yet such graphs are designed to be read by seasoned experts who understand and assume the nonzero base and who have little vertical space or need for zero.
In any case in which the state of the art of a given graph is not implicit to the average reader of the graph, zero should always appear, or else the base value should be clearly indicated by greater emphasis, such as at the lower left somehow. I believe the best way to solve your problem would be to always include a zero but to add a “break” indication along the line — as used in engineering drawing, including graphing — two horizontal or diagonal lines, or a little lightning zigzag, on the y axis line just above the zero mark. Cheap but honest.
(I invented the sparkline; I have drawn a fair number of graphs in the last 50 years.)
“if you’re plotting something that varies from, say 184 to 197”
This seems like the textbook example of why to always have the y-axis start at 0. If it’s, say, the price of oil, and you plot it with a a y-axis extending from 184 to 197, then a price increase from 185 to 195 looks like a 10x spike in oil price, when in reality it’s a ~5% increase in oil price.
Do you find log scales acceptable? In that case, one can use a log scale for the y axis, showing the range [184 197], and the objection goes away.
Adede:
In the oil price example, it’s important to have a long enough time scale for there to be some perspective. In the case of oil prices, if you plot over a few years, the dynamic range of the data is enough that zero is, indeed, in the neighborhood.
When plotting zip codes near Columbia, we get 10027 and 07020, so that zero sits on the borderline of the neighborhood—a hard decision shall be make.
Yuling:
I would plot not using numerical zip code but instead using some relevant zip code level predictor such as estimated average income in the zip code. If the average incomes have a high enough dynamic range, I’d have the graph go all the way down to zero.
Interesting! I always included zero if it was anywhere near any plot elements.
I never thought about why.
Upon reflection, I think the reason for me is cognitive. It’s a critical anchor. And, if would be a distraction NOT to have it.
I made some additional comments here: https://junkcharts.typepad.com/junk_charts/2022/01/start-at-zero-or-start-at-wherever.html
In some cases (where zero isn’t a meaningful value), I’d suggest benching the axis. The axis isn’t compulsory!