Stata tip 165: Marginal titles for graph bar,graph dot,graph box,and beyond

1 What to do if an xtitle() is wanted but not allowed?

With some graph subcommands, you can find that you cannot call up an xtitle() option. Why does this happen, and how can you solve the problem?

The graph command in official Stata can be divided into two groups of subcommands. First come all subcommands that could be issued as graph twoway or twoway calls. (Two of those subcommands could be issued even more directly, namely, as commands in their own right starting with scatter or line.) Second come the others, including graph bar and graph hbar, graph dot, graph box and graph hbox, graph pie, and graph matrix.

A double rule applies to most of the subcommands in this second group. graph pie and graph matrix follow their own rules. Otherwise, the graph axis showing the outcomes or responses of interest is always regarded as the y axis, even if it is horizontal. Moreover, the x axis is not defined. The other axis is regarded as categorical.

This rule flouts long-standing mathematical conventions: a vertical axis is often denoted by y, and a horizontal axis is often denoted by x. But the point is to make it easier to flip in Stata between graph bar and graph hbar or between graph box and graph hbox without changing a bundle of options. You might try one subcommand and then wonder whether the paired subcommand would work better. A common problem with graph bar or graph box is that there is not enough space to show categorical axis labels intelligibly or without some awkwardness. Excessive abbreviation, minute font sizes, or text at an ungainly angle are all sometimes used, but such devices are often avoidable. Problems are eased if you try graph hbar rather than graph bar or graph hbox rather than graph box. Experimentation is easy with this double rule: ytitle(), ylabel(), and various other options still apply to the outcome or response axis even if it is horizontal.

graph dot is a little different but still essentially similar. There is no graph hdot subcommand, but the vertical option flips axes, while the outcome or response axis is still regarded as the y axis. Note that this vertical option is undocumented, but it is easy to apply once you know about it.

That said, this double rule raises the question: what do you do if you want a title to go where the x-axis title would have been, given that the xtitle() option is no longer defined? What you can do is documented but a little indirectly.

The documentation does give many examples with standard options title() and subtitle(). Either of those could provide a good answer, especially if you exploit the possibility of moving a title or subtitle to a different clock position.

The main point of this tip is different. The standard title options also include t1title(), t2title(), l1title(), l2title(), b1title(), b2title(), r1title(), and r2title(). Option names here follow systematic rules. t, l, b, or r as the first character implies titles on the top, left, bottom, or right of each graph, respectively. Options occur in pairs, with second character 1 or 2 indicating an outer or inner title, respectively. Usually, one of these title options is sufficient to overcome the lack of an xtitle() option, but there might be a need to use two or more.

If you wish to confirm your understanding of the possibilities, add these options to any graph call:

Note that the text title may be omitted from each option call. Abbreviations such as t1() are sufficient.

These options were available in Stata’s graphics before it was massively rewritten for Stata 8, and they are still available.

The remainder of this tip gives some simple concrete examples. Note that these options are also available for graph pie and graph matrix. With those subcommands, they do not often seem needed or helpful, but they are perfectly legal.

2 Examples with graph bar, graph hbar, and graph dot

The auto data are likely to be familiar to you unless you have just started using Stata. They include data on various cars from the late 1970s.

graph bar, by default, shows means of an outcome variable. We spell that out with the syntax (mean). Figure 1 shows a simple bar chart. Here the ytitle() is shown on the vertical axis, but we specify our own variant text. By default, no axis title appears to explain the over() variable, which here is rep78. We fill that gap with a call to b1title(), yielding an extra title at the bottom of the graph.

Figure 1.

Simple bar chart of mean miles per gallon for various categories of repair record. The point of interest is the use of b1title() to add explanatory text in a title at the bottom of the graph.

We could draw a horizontal equivalent of figure 1 using graph hbar. If you are following along in your copy of Stata, you will find that the ytitle() call need not be changed; that title now appears automatically at the bottom of the graph. But you will need to change the b1title() call to say l1title() instead.

Let us explore a different way to show the same results using graph dot . As mentioned in the previous section, graph dot has an (undocumented) option vertical . Figure 2 shows a dot chart (or dot plot) alternative to figure 1. The title options are unchanged. The specific option exclude0 is used to inhibit the inclusion of 0 as a label on the outcome axis.

Figure 2.

The results of figure 1 shown differently using a (vertical) dot chart

The main attraction of a dot chart is allowing focus on a comparison of results with each other. Evidently, a bar chart like figure 1 uses 0 as the baseline: statistical results are all shown by bar heights with 0 as the reference base. (By the way, it is often stated as dogma that bars on a bar chart should start at 0, but bars may defensibly start at any relevant base, such as 1 whenever parity is a reference level, 100 for values at a base date of some indexed time series, freezing temperature 32° Fahrenheit, or the mean or another level of some outcome.)

In the case of figure 1, bars start at 0, which matches a common convention. The chart produced may even be the display that researchers or their readers want or expect. But the design wastes space and distracts attention from what is likely to be the question of most statistical and scientific interest: comparison of each group’s result with that of other groups.

A further merit of dot charts is that they can be used with scales such as logarithmic or reciprocal scales on which 0 is not plottable. There might be a case to use 1 or 100 (say) as the reference level, but otherwise, dot charts may work better with such scales.

Such dot charts are often called Cleveland dot plots, which honors the advocacy of William S. Cleveland (1984, 1993, 1994). As often, even an innovative design rarely lacks precedents. See, for example, Snedecor (1937) and later editions until Snedecor and Cochran (1967). For more discussion and examples of dot charts in a Stata context, see Cox (2008, 2024).

Naming these dot charts for Cleveland also serves a purpose of distinguishing them from dot plots that are pointillist displays of frequency distributions and are, in turn, often called Wilkinson dot plots. Stata’s own dotplot command implements the latter.

3 Examples with graph box and graph hbox

Figure 3 is a box plot for the same variables. A box plot shows median and quartiles of an outcome variable. Beyond the quartiles, many different conventions exist, but graph hbox and graph box follow the most common convention, for which the canonical reference is Tukey (1977). Data points are shown individually if and only if they lie at least 1.5 times the interquartile range (IQR) from the nearer quartile and so are equal to or more than the upper quartile +1.5 IQR or equal to or less than the lower quartile -1.5 IQR. Otherwise so-called whiskers are drawn between the quartiles and the most extreme data points not lying beyond those thresholds.

Figure 3.

Box plots for miles per gallon given repair record. The point of interest is the use of l1title() to add explanatory text in a title at the left of the graph

The detail arising from this tip should now be familiar. A call to l1title() adds a title saying more about the over() variable.

A further (entirely optional) twist is that the y axis has been shifted to the top of the graph. That can sometimes look better for graphs with some table flavor. Just as with tables, it is customary to explain columns in headers (Cox 2012).

An extra comment about the data here: The display for repair record 1 may seem puzzling. Drilling down to the data resolves the puzzle. There were just two data points for that category. The precise rules imply that those two data points are reported as the quartiles; the median is reported as halfway between them. There were no other data points to be reported as beyond the quartiles either as individual data points or as implied by whiskers. The scope for box plots to be puzzling is even greater than this, but that really is another story.

If you wanted to use vertical box plots, the difference is essentially that the categorical variable may be explained in a title at the bottom of the graph. Here is command syntax, which you may wish to apply yourself.

4 Conclusion

This tip is a reminder of four pairs of options for placing marginal titles at the top, left, bottom, and right of any graph. Such options may be used generally, but they offer direct solutions to a specific problem. With commands like graph bar, graph dot, graph box, and their horizontal equivalents, you may want to use an xtitle() option but find that such an option is not supported. Now you know of easy workarounds.

Examples to make that concrete have been used to underline other points.

Dot charts can often work as well or better than bar charts, particularly if we focus on comparing values with each other rather than comparing values with some reference base, usually 0, or if logarithmic or reciprocal scales make bar charts problematic.

Such dot charts may be vertical and may explicitly exclude showing 0 if that lies some way outside the observed range.

Charts that have table flavor may look better with horizontal axis material at the top.

Supplemental Material

sj-txt-1-stj-10.1177_1536867X251398633 - Supplemental material for Stata tip 165: Marginal titles for graph bar, graph dot, graph box, and beyond

Supplemental material, sj-txt-1-stj-10.1177_1536867X251398633 for Stata tip 165: Marginal titles for graph bar, graph dot, graph box, and beyond by Nicholas J. Cox in The Stata Journal