Stata tip 159: Absent friends: How to plot what is not present

1 Introduction: Present, nonmissing, missing, and absent

The simple word present may be applied to whatever data are in any Stata dataset we are using. The data present can have missing values; otherwise, they have nonmissing values. Stata has precise rules about what is regarded as missing. There are special codes . and . a to . z for various flavors of missing numeric values. The empty string “ ” is regarded as missing for string values. Otherwise, values in a dataset are nonmissing and take forms such as 42 for a numeric value or “Texas” for a string value.

We sometimes need an extra term for a further category: whatever values might have been included in the data but, in practice, are not present in our data right now. The word absent seems good for this purpose.

For example, the following tabulation shows that in auto. dta, there are no observations with foreign value 1 and rep78 value 1 or 2. Such observations appear to be entirely possible in principle, but they are absent from the dataset in practice.

Once again, note the difference between missing and absent. The option missing insists that tabulate include a display of the frequencies of missing values if there are any. There are five observations with missing values on repair record. Suppose further contrary to all the evidence, that categories 0 and 6 were possible for extremely poo1 and extremely good repair records, respectively. Because such categories are not present in the data, they must be regarded as absent. They are not shown in the table, which is natural enough, but even if value labels were defined for repair record 0 or 6, that alone would not make tabulate show zero frequencies.

So far, this is mostly about terminology. Some users might report that values of 1 and 2 on repair record for foreign cars are missing from the data, but I recommend the term absent as more appropriate.

In this tip, I consider handling of absent values with a focus on graphics, showing how to plot absent values. We will not go beyond what is possible with official commands and functions.

2 Plotting absent values

Sometimes, people want to plot absent values in Stata; that is, they want to plot the values but show that the values are absent. You could be producing a series of graphs that you want to have a consistent format, both for instances in which sometimes absent values really are present and for instances in which sometimes absent values really are absent.

Absence is simple on one level. It means zero for counts or frequencies, for proportions or fractions, for percentages, and (if relevant) for probability densities.

Sometimes, the question is trivial, even to the extent of never being asked. If you have ever started an axis at zero to underline that a variable is positive or extended any axis to show possible values, you have solved the problem. This is generally possible in any application of graph twoway. In auto. dta, miles per gallon (mpg) ranges from 12 to 41. Other cars beyond that dataset might well show performance outside that range. whether better than the maximum observed of 41 or worse than the minimum of 12. So it is easy to imagine wanting to stretch the axis to accommodate other values.

If needed, you can extend any axis in a twoway graph. You can do that explicitly by using, say, yscale() or xscale() as an option. Or you can do that implicitly by using, say, ylabel() or xlabel() as an option. See the help on axis scale options or axis label options if you want more details. On, say, a histogram, absence is shown by bars that do not exist or, if you prefer, bars that do exist but are of zero height and are thus hard to see.

3 Plotting absent categories: The problem

Let’s now get to the nub of a problem that may bite. With auto. dta, as before, we might ask for a bar chart showing repair record frequencies for foreign cars. See figure 1.

Figure 1.

Bar chart showing frequencies of repair record grades for foreign cars. Note that grades 1 and 2 are not shown.

Stata examines the data specified and shows frequencies for the categories of repair record present for foreign cars. As the table in section 1 already showed, these are just grades 3, 4, and 5. Grades 1 and 2 do not appear in the graph. The idea of changing the $iL$ axis (in this case, given the use of graph bar) will not get you anywhere. The L axis does not even exist for this command, and changing to graph hbar or graph dot will not help either. There are options to tune the labels of what is regarded as a categorical axis, but they will not help either.

4 Plotting absent categories: Use twoway instead.

Remarks in section 2 already pointed to a solution. You can switch to an appropriate twoway graph command and then exploit the scope to change the axis scale and axis labels. For example, the graph produced by this command is similar to figure 1 but also shows axis labels for repair records 1 and 2. See figure 2.

Figure 2.

Histogram showing frequencies of repair record grades for foreign cars. Note that grades 1 and 2 are shown by request but without bars (or equivalently with bars of height zero). Foreign cars with such grades are absent from the dataset, but we can still request that the corresponding axis labels be shown.

The reason for showing axis labels at 1 and 2 is that we want to underline that such values occur in the rest of the data. But as far as Stata is concerned, we could ask for any labels we want. To continue the little fantasy introduced earlier, we could also ask for labels at 0 and 6 if those values were possible in principle but absent in practice

Here is another technique that is sometimes useful. We construct a frequency variable ourselves and then call up twoway bar.

bysort rep78: egen frequency = total(foreign == 1)

twoway bar frequency rep78, barw(0.6) xla 1/5

> title(Repair record for foreign cars)

The results are not shown here, because they are similar to the previous figure Incidentally, because foreign is a (0, 1) indicator variable, in this particular situation,

bysort rep78: egen frequency = total(foreign)

would have worked just as well. The expression foreign ==1 is true, evaluating to l precisely when foreign is 1, and false, evaluating to 0 precisely when foreign is 0. It is crucial to the small technique just used that occurrences of 0  s do not affect the total add 0 to your total so far, and that total is unchanged. So such zeros are not counted by virtue of making no difference to the total. If you are curious to read more about indicator variables, sometimes called dummy variables and sometimes given yet other names, look at Cox and Schechter (2019) and their references.

5 Plotting absent categories with two or more variables: easy

Suppose that we want a bar chart with joint frequencies of two or more variables. Stata makes plotting of absent categories much easier in that case. Commands like

. graph bar (count), over(rep78) over(foreign)

> title(Repair record for domestic and foreign cars)

. graph hbar (count), over(foreign) over(rep78)

> title(Repair record for domestic and foreign cars) saving(ABS4, replace) file ABS4.gph saved

leave holes without (visible) bars precisely where you might want them. So that result is equivalent to the result of tabulate seen in section 1. Stata’s logic here may be already familiar to you.

Figure 3 shows the second of those graphs. Showing the horizontal variant rather than the more common vertical variant is intended as a small reminder that it often does better at displaying categorical labels readably.

Figure 3.

Bar chart showing frequencies of repair record for foreign and domestic cars. By default, absent categories are just holes without visible bars.

Other way round, if you do not want that behavior, then you need to override the default with an extra nofill option. That suppresses any bars with zero height that show absent cross-combinations.

6 Plotting absent categories with two or more variables: challenging

If this seems too easy now, let’s revisit (for the final time) the fantasy about further absent categories 0 and 6 for repair record. Suppose we want to extend figure 3 by also showing those fantastical categories.

It is not elegant, but it may be practical just to create a small dataset for the purpose. Below is the code followed by commentary

. contract foreign rep78, zero

. expand 4 in L

(3 observations created)

. replace _freq = 0 in -4/L

(4 real changes made)

. replace foreign = 0 in -4/-3

(2 real changes made)

. replace foreign = 1 in -2/L

(0 real changes made)

. replace rep78 = 0 in -4/-3

(2 real changes made)

. replace rep78 = 6 in -2/L

(2 real changes made)

The idea is that contract reduces the dataset to the minimum we need as a starting point. Then we add four extra observations: two for repair record 0 and foreign 0 or 1 and two for repair record 6 and foreign 0 or 1. For all of those extra observations, the frequency is zero. Then we can fire up a general command, as shown below. Figure 4 is the result.

Figure 4.

Bar chart showing frequencies of repair record for foreign and domestic cars. The fantastical categories of repair records 0 and 6 are added to show technique as further holes without visible bars.

Fastidious readers may recoil at such a programming style. It is very specific to the example data and the particular problem. Still, getting the job done quickly and easily occasionally overrides any preference of style. As Good (1965, 56) remarked, “real life is both complicated and short, and we make no mockery of honest adhockery”.

Once we have a quick and dirty solution, however, a second look may suggest a more general approach. The example problem is concocted. If you have a problem with sometimes absent categories, it may well be more complicated and capricious. Occasional or rare categories may appear and disappear from dataset to dataset.

The need is to fill in absent categories with observations in a transient dataset with zero frequencies. But why not flip the problem around? Set up a template dataset with all possible categories and zero frequencies. Then combine that with real data, using a rule that positive frequencies override zero frequencies. Here is how that might work.

First,set up a template

. clear

. set obs 14

Number of observations (_N) was 0, now 14.

. generate foreign = mod(_n, 2)

. egen rep78 = seq(), from(0) to(6) block(2)

. generate _freq = 0

. save template

file template.dta saved

In your real problem, the template might well be more elaborate. Often, categories are best presented with value labels.

Now go back to the original dataset. Get the frequencies we need and combine them with the template.

The resulting graph is not shown here. It is identical to figure 4.

The code just given is still a sketch. The approach could be developed or varied in several ways. You would typically want to save the graph for later use. Users familia1 with frames might prefer to use frames. If such a need is a common problem for you, setting up a do-file might be the next step.