As you can see, all we need to do is list the variables to be described, using the i. prefix for factor variables. The by() option specifies a classification variable, with the suboption test to request a test of differences across that variable, based on regression or Pearson’s statistic as indicated in the notes. That’s quite a bit of work with little effort on our part.
We see that the sample has a few more urban than rural women, and that urban women are younger, more educated, and more likely to use contraception (particularly efficient methods) than rural women. Moreover, all three differences are highly significant.
The sample statististics showing the urban/rural split can be omitted using the nosample option. You can also select which statistics to calculate and where to place them using the sample option, type help dtable##sample for details.
There is a continuous option to specify the statistics and/or tests to use for one or more continuous variables. For example if you wanted to use the median and interquartile range as descriptive statistics and the Kruskal-Wallis rank test for education you could use the option continuous(educ, stat(median iqr) test(kwallis)) . Omitting the variable name would apply these choices to all continuous variables. To see a list of all the statistics and tests available for continuous variables type help dtable##cstats and help dtable##ctests .
There is an equivalent factor option to specify the statistics and tests to be used for factor variables. For example you can use Fisher’s exact test, or a test based on ordinal association, such as Kendall’s tau or Goodman and Kruskal’s gamma. Type help dtable##fstats and help dtable##ftests for a full list of statistics and tests available for factor variables.
The dtable command has a large number of options, including several that control table styles. The command creates its own collection called DTable , which allows further customization using collect commands. To learn more type help dtable .
3.2.3 An Alternative Table 1
The code below shows an alternative “table 1” that can be obtained with the table command in both Stata 17 and 18. It shows sample sizes, mean and standard deviations on separate lines for continuous variables, and just percents for factor variables, but no significance tests.
. gen N = 1 . table (var) (area) , /// > stat(count N) /// sample > stat(mean age educ) stat(sd age educ) /// continuous > stat(fvpercent cuse) /// factor > nformat(%5.2f mean sd) nformat(%5.1f fvpercent) /// > sformat((%s) sd) sformat(%s%% fvpercent) style(table-1) ───────────────────┬─────────────────────────── │ Type of area │ Urban Rural Total ───────────────────┼─────────────────────────── N │ 1,683 1,432 3,115 │ Age in years │ 27.43 28.51 27.93 │ (9.41) (10.08) (9.74) │ Education in years │ 5.87 3.25 4.66 │ (3.79) (2.71) (3.58) │ Contraceptive use │ Not using │ 45.8% 68.1% 57.1% Inefficient │ 7.7% 7.0% 7.4% Efficient │ 46.5% 25.0% 35.6% ───────────────────┴───────────────────────────
We first create a new variable called N to obtain sample sizes. We specify the table rows using var , which refers to the variables in the statistics option, and the columns using area . We then request the count for the sample size, the mean and sd for our continuous variables, and the fvpercent for our factor variable.
To control the number of decimals printed we use our old friend nformat , specifying 2 decimals for the mean and standard deviation, but just one for percents. To enclose the standard deviations in parentheses and append a % sign to the percents we use sformat . (If you are puzzled by the %s%% format, note that %s is the placeholder for the string and that to append a % symbol we need to escape it using %% .)
Finally we use the built-in style table-1 , which provides a more compact layout for factor variables and a few other tweaks. Try running the table without the style to see what it does.
3.3 Tables of Estimates
We now turn our attention to tables presenting the results of one or more estimation commands. We will use as an example simple linear regression with the regress command, but the same ideas apply to other models. We could collect the results ourselves using collect as a prefix of the regress command, or even the command option of table , but the
etable command makes things easier.
3.3.1 A Single Regression
If you type etable after a regress command you get a table showing coefficients with standard errors in parentheses, and the number of observations at the bottom. Let us add just a couple of options.
. sysuse auto, clear (1978 automobile data) . quietly regress mpg i.foreign . etable, showstars showstarsnote ───────────────────────────────-- mpg ───────────────────────────────-- Car origin Foreign 4.946 ** (1.362) Intercept 19.827 ** (0.743) Number of observations 74 ───────────────────────────────-- ** pSo foreign cars travel almost 5 more miles per gallon than domestic cars. The option showstars shows the usual significance stars, and showstarsnote adds an explanatory note. The stars may be customized using the stars() option, type help table##starspec to see how.
3.3.2 Comparing Two Regressions
To compare two or more regressions all we have to do is save the results of each one using estimates store (before they are overwriten by the next regression) and then pass the list of stored estimates to etable .
. gen gphm = 100/mpg . quietly regress gphm i.foreign . estimates store unadjusted . quietly regress gphm i.foreign weight . estimates store adjusted . etable, estimates(unadjusted adjusted) column(estimates) /// > cstat(_r_b) cstat(_r_z, sformat((%s))) /// > note(test statistic in parentheses) showstars showstarsnote ───────────────────────────────--─────────-- unadjusted adjusted ───────────────────────────────--─────────-- Car origin Foreign -1.005 ** 0.622 ** (-3.29) (3.11) Weight (lbs.) 0.002 ** (13.74) Intercept 5.318 ** -0.073 (31.92) (-0.18) Number of observations 74 74 ───────────────────────────────--─────────-- ** pHere we compare the efficiency of foreign and domestic cars before and after adjusting for weight. Our measure of efficiency is gallons per 100 miles or gphm rather than the usual mpg , because it has a more linear relationship with weight. To get the defaults try etable estimates(unadjusted adjusted) . Here we added a couple of options.
The option column(estimates) specifies that we want the columns to be labeled with the name of the estimates rather than the name of the dependent variable, which is the default.
The cstat option (short for coefficient statistics), lets you select which statistics to display. Type help etable##cstat to see a complete list. Here we selected the coefficient ( _r_b ) and the test statistic ( _r_z ). To make sure the test statistic is in parentheses we use the sformat option of cstat to specify (%s) , where %s is a placeholder for the string, just as we did earlier in Section 3.2.1. We also use the note option of etable to indicate exactly what’s shown.
There is also a mstat option (short for model statistics) that lets you select model statistics to display, such as the number of cases, R-squared, Akaike’s information criterion, and others. Type help etable##mstat to see a list. Try adding R-squared to the previous table.
3.3.3 Regressions with Different Outcomes
Our last example compared regressions with the same outcome and different predictors. It is also possible to compare regressions with different outcomes and the same predictors (or at least some overlap). The table below compares regressions of weight and length using four and three predictors, respectively, with foreign cars as the reference cell for car origin:
. quietly regress weight ib1.foreign price rep78 headroom . estimates store weight . quietly regress length ib1.foreign price rep78 . estimates store length . etable, estimates(weight length) eqrecode(weight=both length=both) /// > mstat(N) mstat(r2) showstars showstarsnote ─────────────────────────────────--──────────-- weight length ─────────────────────────────────--──────────-- Car origin Domestic 893.057 ** 29.353 ** (137.788) (5.013) Price 0.140 ** 0.003 ** (0.017) (0.001) Repair record 1978 -47.367 -0.211 (61.474) (2.347) Headroom (in.) 222.060 ** (61.361) Intercept 1048.304 ** 147.845 ** (320.826) (11.229) Number of observations 69 69 R-squared 0.76 0.56 ─────────────────────────────────--──────────-- ** pThe essential new option here is eqrecode() which ensures that coefficients for the same predictor with different outcomes appear in the same row. Try running the command without this option to see the default. This option is also essential if you run a multivariate regression. At the bottom of the table we listed R-squared for each regression, but you already knew how to do that, right? Did you notice that to keep the number of observations you have to add mstat(N) ?
The etable command creates a collection called ETable which becomes the current collection and can then be modified and/or exported. Type help etable to learn more.
3.4 Collection Tables
Let us move now to an example where we will collect the results of standard Stata commands ourselves. We want to calculate Tukey’s five number summary, namely the minimum, first quartile, median, third quartile and maximum. These statistics are all computed by summarize with the detail option. We would like to do this for several variables.
The collect command can be used as a prefix to gather the results stored by a general command in r() or by an estimation command in e() . You can find out exactly what a command has stored by typing return list after a general command such as summarize , or typing ereturn list after an estimation command. But don’t worry, collect will gather everything. So here is our table:
. sysuse auto, clear (1978 automobile data) . collect clear . quietly collect, tags(cmdset[mpg]): summarize mpg, detail . quietly collect, tags(cmdset[length]): summarize length, detail . quietly collect, tags(cmdset[weight]): summarize weight, detail . collect style autolevels result min p25 p50 p75 max . collect label levels result /// > min "Min" p25 "Q1" p50 "Md" p75 "Q3" max "Max", modify . collect layout (cmdset) (result) Collection: default Rows: cmdset Columns: result Table 1: 3 x 5 ───────┬────────────────────────── │ Min Q1 Md Q3 Max ───────┼────────────────────────── mpg │ 12 18 20 25 41 length │ 142 170 192.5 204 233 weight │ 1760 2240 3190 3600 4840 ───────┴──────────────────────────This will require a bit of explanation. We start by clearing the collection system with collect clear .
We then collect the results of summarize mpg, detail , which will produce the statistics we need, using quietly to skip displaying them. We also ask the system to tag the results with the name of the variable being summarized, which unfortunately is not stored with the results. Fortunately Stata creates a dimension called cmdset for our commands, which are just numbered 1, 2, and 3. The tags option creates a more informative tag, using the name of the variable.
Next we define a style. As it happens, summarize, detail produces 19 results and we don’t want them all, just the five-number summary. The collect style autolevels result command sets the levels of result to the five statistics we want. (Alternatively, you can specify which results to collect, type help collect get to learn more.)
Stata generates labels for practically all the results stored by its commands, for example the label for p25 is “25th percentile”, and by default uses these on the tables. We would like to use shorter labels, in this case “Q1”, hence the collect label levels result command.
The final step is to specify the layout of the table with collect layout , which says we want the cmdset with the variable names in the rows, and the result with the five-number summaries in the columns. The row and column specifications in collect layout must be enclosed in parentheses.
Rather than repeat essentially the same command three times, varying only the name of the variable, we could have used a loop, a concept discussed later in Section 5.2 of this tutorial. That would make it easy to include many more variables in our table.
It is possible to produce similar results using table , as all five summaries are in the list of statistics available, but the idea here was to collect the results ourselves to give you a sense of the power and flexibility of the collection system.
3.5 Customizing Tables
Consider the two-way table in Section 3.1.2, showing contraceptive use by education. We would like to show just the row percents, as we did, but add a column with the total number of observations in each row. One way to do this is to get both the frequencies and percents, and then decide exactly what we want to show and how. We will also modify the header, and remove a vertical border. Try the following commands (you may want to try the first two without quietly to see what happens at each step):
. use https://grodri.github.io/datasets/drsr03x, clear (DRSR03 extract) . quietly table educg cuse, stat(percent, across(cuse)) stat(frequency) . quietly collect layout (educg) /// > (cuse#result[percent] cuse[.m]#result[frequency]) . collect style header result , level(hide) . collect style cell border_block, border(right, pattern(nil)) . collect preview ───────────────────────────────────────────────────────────────────── Contraceptive use Not using Inefficient Efficient Total ───────────────────────────────────────────────────────────────────── Education level 0-2 69.38 5.37 25.25 100.00 503 3-4 59.65 5.45 34.90 100.00 404 5-7 50.00 8.50 41.50 100.00 306 8-18 31.84 14.43 53.73 100.00 201 Total 57.07 7.36 35.57 100.00 1,414 ─────────────────────────────────────────────────────────────────────After using table to tabulate the data, we use collect layout to specify rows with educg and columns with the percents for cuse (using an interaction between cuse and result[percent] ) and the frequency for the total (interacting cuse[.m] with result[frequency] ).
We have used dimensions informally to refer to the rows and columns of a table, but the concept of dimension here is more general, representing all features used to tag the elements of a collection. Type collect dims to list all dimensions of the current collection. Type collect levelsof dimname to list the levels of a dimension, and collect label list dimname to list the labels of the levels. This is how I learned that cuse[.m] had the totals.
Finally we use a couple of collect style commands that aim for a cleaner look; one to remove the labels of the levels of result from the header, and another to omit the vertical border between the row headers and the body of the table. This, by the way, uses yet another dimension called border_block , used to tag cells in the row and column headers, the top-left corner, and the body of the table with the items. Type collect levelsof border_block to list the level names.
This example has barely touched the surface of table customization. To learn more type help collect .
3.6 Exporting Tables
Tables are displayed on your screen but can also be exported in various formats, including HTML, Word documents, Excel documents, LaTeX, PDF, plain text, Markdown and even Stata’s own SMCL format. Type collect export to learn more.
© 2024 Germán Rodríguez, Princeton University
