From datasets to resultssets in Stata

Roger Newson Kings College London, London, UK roger.newson@kcl.ac.uk http://www.kcl-phs.org.uk/rogernewson/ Distributed at the 10th UK Stata Users Group Meeting on 29 June 2004

What are resultssets?

A resultsset is a Stata dataset created as output by a Stata command. It may be simply listed to the Stata log and/or output to a disk le and/or written to the memory, overwriting any pre-existing dataset. If you are a SAS user converting to Stata, then note that Stata datasets do the job of SAS data sets, and Stata resultssets do the job of SAS output data sets.

1.1

Why resultssets?

In general, a Stata dataset, or a dataset in any other format, should contain one observation per thing, and data on attributes of things. For instance, in the medical sector, a dataset might have one observation per patient, and data on the patients baseline characteristics. Alternatively, a dataset might have one observation per visit to a health centre, and data on the state of the patient who made the visit. Usually, the variables in a dataset are either primary key variables, which identify the things corresponding to the observations, or non-key variables, which identify interesting attributes of those things. For instance, if there is one observation per patient, then one of the variables is usually a patient ID number, which identies the observation uniquely. Or, if there is one observation per visit, and a patient may only make one visit per day, then the primary key variables are usually patient ID and visit date. Statisticians (and other data analysts) are typically provided with (or collect for themselves) a dataset with one observation per experimental or observational unit, where a unit may be a patient, a patientday, a country-year, a car model, or some other thing. However, they are typically paid to produce plots for presentation, or tables for publication. To do this, they really need datasets with one observation per plotted data point, or per Y -axis label, or per X -axis label, or per table row. Axis labels and table rows do not often correspond to the original units in the original dataset. For instance, in the auto data, shipped with ocial Stata, there is one observation per car model, and the primary key is the single variable make. Figure 1 gives condence intervals from a regression model measuring dierences in fuel eciency (in miles per gallon) in cars from inside and outside the US, with various 1978 repair records, compared with a reference car, made by a US company and with a repair record of 3. It was created using the eclplot package, which creates condence interval plots, and requires, as input, a dataset with one observation per condence interval to be plotted and data on estimates and condence limits. Table 1 gives the same data (plus P -values) as a table. It was created using the listtex A package, which creates tables which can be cut and pasted into a TEX, L TEX, HTML or word processor document, and requires, as input, a dataset with one observation per table row. Both eclplot and listtex are downloadable from SSC, but neither can use directly as input the original auto data, with one observation per car model. Instead, resultsset-generating and resultsset-processing packages are used, taking as input the original auto data, and creating, as output, datasets that can be input to eclplot and listtex. This survey explains how to do this, using example do-les distributed with this document on the conference website at http://www.stata.com/support/meeting/10uk/ . These do-les will work under Stata 8 if the user has installed the required unocial Stata packages, listed at the top of each do-le. These packages are downloadable from the SSC archive at http://ideas.repec.org/s/boc/bocode.html using the Stata ssc command. The full list of required packages comprises the resultsset-generating packages descsave, parmest, xcollapse and xcontract, together with the resultsset-processing packages eclplot, listtex, dsconcat, sencode, sdecode, factext, factmerg and ingap, and the estimation package somersd.

From datasets to resultssets in Stata

Figure 1: Dierences in mileage in the auto data (compared to US car with rep78==3).

Car type: Domestic (n=48) Foreign (n=21) Car group Repair Record 1978: 1 (n=2) 2 (n=8) 3 (n=30) 4 (n=18) 5 (n=11)

5.00

0.00 5.00 10.00 Difference in mileage (mpg)

15.00

Table 1: Dierences in mileage in the auto data (compared with US car with rep78==3). Car group Car type: Domestic (n=48) Foreign (n=21) Repair Record 1978: 1 (n=2) 2 (n=8) 3 (n=30) 4 (n=18) 5 (n=11) Dierence (mpg) 0.00 3.56 1.92 0.05 0.00 0.81 5.38 (95% (ref.) (0.32, (2.78, (2.98, (ref.) (2.06, (0.58, CI) P

6.80) 6.62) 3.07) 3.68) 11.33)

.032 .42 .98 .57 .076

Resultsset-generating programs

Some programs which generate resultssets are summarized in Table 2. The descsave package is an extended version of describe (see [R] describe), and creates a resultsset with one observation per variable, and data on variable attributes, which are the names, storage types, display formats, value labels and variable labels of the variables, and (optionally) characteristics (see [U] 15.8 Characteristics or [P] char). The parmest package contains two programs, parmest and parmby. parmest creates a resultsset with one observation per parameter of the most recent set of estimation results (see [U] 29 Estimation and post-estimation commands or [P] ereturn), and data on general parameter attributes, which include the names, estimates, standard errors, condence limits and P -values of the parameters. parmby calls a user-specied estimation command once for each by-group (or once only, if the by() option is absent), and creates a resultsset with one observation per parameter (or per parameter per by-group, if the by() option is present), and data on the same parameter attributes as output by parmest. xcollapse and xcontract are extended versions of the ocial Stata programs collapse and contract, respectively (see [R] collapse and [R] contract). SAS programmers may note that descsave does the job of PROC CONTENTS in SAS, xcontract does the job of PROC FREQ, xcollapse does the job of PROC MEANS and PROC SUMMARY, and parmest and parmby do the job of the OUTEST= option of SAS estimation procedures. The other resultsset-generating programs are part of ocial Stata, and are documented in [R] statsby, [R] bootstrap, [R] simulate and

From datasets to resultssets in Stata

[P] post, respectively. statsby calls a Stata command once for each by-group, and creates a resultsset with one observation per by-group and data on general statistical results stored by the called Stata command in its returned results in r() or e() (see [R] Saved results, [P] return and [P] ereturn). The post package is a general low-level utility, allowing users to write their own resultsset-generating programs. Table 2: Some resultsset-generating programs available in Stata. Program Downloadable from SSC: descsave parmest parmby xcollapse xcontract Ocial Stata: collapse contract statsby bootstrap simulate post The resultsset contains: one observation per... and data on... variable estimated parameter parameter per by-group by-group combination of variable values by-group combination of variable values by-group bootstrap sample simulation replication anything variable attributes parameter attributes parameter attributes basic summary statistics frequencies and percentages basic summary statistics frequencies general statistics bootstrap sample statistics simulation results anything

The program in example1.do demonstrates the generation of resultssets by descsave, parmby, xcollapse and xcontract, if the user has installed the required packages (listed at the top of the do-le). It uses as input the auto data, shipped with Stata and accessed using the sysuse command. The user can open it in the Stata do-le editor, click on the icon labelled Do current le, and return to the Stata Command window. Alternatively, the user can type the command do example1 in the Stata Command window. Each resultsset-generating program generates a resultsset which is saved to the memory, overwriting the preexisting data. In each case, the program lists the variables in the resultsset using describe, and lists the observations in the resultsset using list. The more command allows the user to look at the variables or observations before pressing the space bar, which allows the program to continue. Variables in the resultssets are discussed in the online help for each command, and in Newson (2003) in the case of parmby and parmest.

2.1

Common options for resultsset-generating programs

Dierent resultsset-generating programs often have options of the same names with the same functions. Some commonly-used options are listed in Table 3. Their exact syntax is given in the online help for each program, although only parmby has them all. Table 3: Common options for resultsset-generating programs. Option Resultsset-destination options: list() saving() norestore (or replace) fast flist() Resultsset-modifying options: rename() format() idnum() idstr() by() Function List resultsset to log and/or Results window Save resultsset to disk le Overwrite any existing data in memory Fast version of norestore for programmers Global macro accumulating saving() lenames Rename variables in resultsset from default names Give variables in resultsset nondefault formats Value of numeric resultsset ID variable Value of string resultsset ID variable Specify by-groups

The rst ve options specify the destination of the resultsset, and are not mutually exclusive. The list() option species that the resultsset (or a subset of its variables and/or observations) is listed to the Results

From datasets to resultssets in Stata

window and/or the Stata log, in a style that can be customized by the user using suboptions. saving() species that a resultsset is saved to a disk le. norestore, used in example1.do, species that the resultsset is saved to memory, overwriting any existing dataset. fast is a high-speed version of norestore, mostly for advanced programmers, which takes no action to restore the existing data if the user presses Break. flist() is used with saving(), and species the name of a global macro, to which the lename of the resultsset le will be appended, in order to accumulate a list of resultsset lenames. Resultssets saved to disk can be used as the using dataset by append, merge, joinby and cross, in the same way as other disk datasets. (See [U] 25 Commands for combining data for more on these commands.) A particularly useful package is dsconcat, downloadable from SSC. This takes as input a list of lenames of disk datasets, and concatenates the observations in these datasets, or a subset of these observations and/or of the variables, into the memory, overwriting any existing data. The input datasets are often resultssets, especially parmby resultssets. The list of input les may be stored in a global macro specied by the flist() option, and an example of this practice appears in the online help for parmest. dsconcat is useful if the user estimates parameters for more than one regression model, and wishes to plot or tabulate all these parameters, or an interesting subset of them, in a single plot or table. This ability is an advantage of the resultsset method over alternative and complementary methods of creating publication-ready tables of regression results, using ocial Statas estimates table (see [R] estimates), or using John Gallups outreg or Tony Bradys reformat, both downloadable from SSC. On the other hand, the complementary methods are more instant solutions, if the user only wants to tabulate results from one regression model at a time, does not require plots, and is publishing in a journal which prefers the specic format of estimates table, outreg or reformat. The last ve options modify the variables in the resultsset. The rename() and format() options specify nondefault names and formats, respectively, for variables in the resultsset. The idnum() and idstr() options specify values for a numeric and string resultsset ID variable, respectively, both of which have the same value for all observations in the resultsset. If we concatenate multiple input resultssets, then it is useful to know which input resultsset each observation came from, and the idnum() and idstr() options make this easy. The by() option species a list of by-variables, specifying by-groups. The resultsset will then be concatenated at birth, and contain a resultssubset of observations for each by-group.

2.2

Resultssets are often stored in temples

Resultssets on disk are very often stored in temporary les, produced using either the tempfile command or the tempfile extended macro function. In the Stata world, temporary les are usually viewed as a technical specialist subject, mostly of interest to advanced programmers, mentioned briey in [U] 21.7 Temporary objects, and documented more fully in [P] macro. This is in contrast to the SAS world, where users typically learn about temporary SAS data sets before learning about permanent SAS data sets. Stata has the advantage over SAS that it can do much more data handling in memory. However, this advantage is limited by the fact that it can only do this with one dataset at a time. It is therefore a good idea for Stata users to know at least enough about macros to be able to handle tempfiles. The program in example2.do demonstrates the use of tempfile in combination with dsconcat, descsave and xcollapse. After loading the auto data, we use the tempfile command to tell Stata that we intend to create temporary les with macro names tf0 to tf9, and create the rst one as a descsave resultsset, with one observation for each of 9 quantitative variables, which we might want to compare between US and non-US cars. This resultsset is sorted by variable name using the gsort() option, and stored on disk using the saving() option. We then create the other 9 temporary les as xcollapse resultssets, each with 1 observation per car type (US and non-US) and data on the median value of one of the 9 quantitative variables. Each xcollapse resultsset has a string ID variable named idstr, identifying the quantitative variable to which the medians in that resultsset belong. We then concatenate the resultssets into the memory using dsconcat, and describe the variables in the new concatenated dataset in memory, and list the observations. Finally, we rename the string ID variable idstr to name, sort the dataset by it, merge in the descsave resultsset, and create a new dataset in memory, with one observation per quantitative variable per car type, sorted rst by car type and second by the order of the quantitative variable in the list. This dataset is listed, so that we can compare the medians of the 9 quantitative variables between US and non-US cars. This nal listing should look like this: . by foreign: list order name varlab median, noobs ------------------------------------------------------------------------------------------> foreign = Domestic

From datasets to resultssets in Stata +--------------------------------------------------------+ | order name varlab median | |--------------------------------------------------------| | 1 price Price 4782.5 | | 2 mpg Mileage (mpg) 19 | | 3 headroom Headroom (in.) 3.5 | | 4 trunk Trunk space (cu. ft.) 16 | | 5 weight Weight (lbs.) 3360 | |--------------------------------------------------------| | 6 length Length (in.) 200 | | 7 turn Turn Circle (ft.) 42 | | 8 displacement Displacement (cu. in.) 231 | | 9 gear_ratio Gear Ratio 2.75 | +--------------------------------------------------------+

------------------------------------------------------------------------------------------> foreign = Foreign +--------------------------------------------------------+ | order name varlab median | |--------------------------------------------------------| | 1 price Price 5759 | | 2 mpg Mileage (mpg) 24.5 | | 3 headroom Headroom (in.) 2.5 | | 4 trunk Trunk space (cu. ft.) 11 | | 5 weight Weight (lbs.) 2180 | |--------------------------------------------------------| | 6 length Length (in.) 170 | | 7 turn Turn Circle (ft.) 36 | | 8 displacement Displacement (cu. in.) 101 | | 9 gear_ratio Gear Ratio 3.61 | +--------------------------------------------------------+

. more This time, the 9 quantitative variables are identied, not only by their names, but also by their more informative labels (with units), merged in from the descsave resultsset. Examples of processing multiple parmby resultssets can be found in Newson (2003), and in the online help for parmest.

String-numeric conversion in resultssets

A lot of the work in resultsset processing is converting string variables to numeric variables and/or numeric variables to string variables. To be plotted, a variable must usually be numeric, even if it corresponds to string axis labels, such as those on the vertical axis of Figure 1. On the other hand, if we want to perform string substitutions on a numeric variable and/or add prexes or suxes, then the variable must be converted to string. For instance, we might want to tabulate condence limits with parentheses and commas (as in Table 1), or add stars to P -values, and neither of these is provided by any existing Stata display format. In practice, it is common to present the same results both as plots (for presentation in meetings) and as tables (for publication in journals). It is therefore important to be able to convert variables between string and numeric with a minimum of programming by the user. Ocial Stata provides the utilities encode and decode for conversion between string variables and numeric variables with value labels, and the utilities destring and tostring for conversion between string variables and numeric variables without value labels. These utilities were discussed by Cox (2002). They all have limitations, especially encode, which encodes string variables to numeric in an alphanumeric order. For this reason, I have developed the programs sencode and sdecode for basic conversions between single numeric variables and single string variables, and the programs factext and factmerg for string-factor conversions, which often involve multiple string variables and/or multiple numeric variables converted in parallel. These programs are all downloadable from SSC.

From datasets to resultssets in Stata

3.1

Basic string-numeric conversion using sencode and sdecode

The programs sencode and sdecode are super versions of encode and decode, respectively, and convert string variables to labelled numeric variables and labelled numeric variables to string variables, respectively. With both programs, the output variable may either be generated as a new variable or replace the input variable, taking over its name, variable label, position in the dataset and characteristics (see [U] 15.8 Characteristics or [P] char). sencode takes, as input, a string variable, and generates, as output, a numeric variable, whose value label in each observation is equal to the value of the input string variable in the same observation. It has a gsort() option, enabling the user to dene numeric codes in ascending or descending order of a user-specied list of variables, as recognised by the gsort command (see [R] gsort). sencode groups the observations into ordered groups according to the gsort() option, and then generates an encoded integer value for each combination of gsort() group and input string value, ordered rst by gsort() group and second by alphanumeric order of input string values within each gsort() group. Usually, each gsort() group contains only one input string value, so the encoding is in ascending order of gsort() group. If the gsort() option is not specied, then it is set in default to be equivalent to the Stata expression n, specifying a gsort() group for each observation, ordered by the position of that observation in the dataset. There is also a manyto1 option, allowing the relation between output numeric values and input string values to be many-to-one. Therefore, it is possible for multiple observations to have the same output numeric value, and for multiple output numeric values to have the same input string value. In the simplest case, if gsort() is not specied, an observation has an output numeric value depending on its position in the dataset (if manyto1 is specied), or on the position in the dataset of the rst appearance of its input string value (if manyto1 is not specied). sdecode takes, as input, a numeric variable, and generates, as output, a string variable. In default, the value of the output string variable in each observation is equal to the value label of the input numeric value in that observation (if such a value label exists), or to the formatted string value of the input numeric value (if the input numeric value is nonmissing and has no label), or to an empty string (if the input numeric value is missing). sdecode therefore converts numbers to strings in a similar way to tabulate. The program in example3.do uses the same quantitative variables as the program in example2.do. However, instead of comparing medians between US and non-US cars, we now compare the quantitative variables as diagnostic predictors of the condition of non-US origin. We do this using the somersd package, introduced by Newson (2000) and discussed by Newson (2002), which computes condence intervals for the Somers D parameter of each quantitative predictor with respect to non-US origin. For each quantitative diagnostic predictor, Somers D is related to the area under the specicity-sensitivity or receiver-operating characteristic (ROC) curve (see [R] roc) by the equation D = 2A 1, where D is Somers D and A is the ROC area. Somers D is therefore a performance indicator equivalent to the ROC area, but has the advantage that it is negative, positive or zero when the median dierence between non-US and US cars is negative, positive or zero, respectively. We start by loading the auto data. We then use parmby to call somersd and create in the memory a resultsset with one observation per Somers D parameter and data on parameter attributes such as condence intervals and P -values (with sensible display formats). We describe the variables in the resultsset, and list all observations, selecting the most interesting variables, which include label, a string variable containing, for each parameter, the variable label of the corresponding predictor variable. We then use sencode to encode this string variable to a numeric variable predictor, with variable labels, which are listed. The numeric codes are ordered in ascending order of the Somers D estimates themselves by specifying the option gsort(estimate parmseq), which orders codes rst by Somers D estimate, and breaks ties by the order of the predictor in the list. Next, we use eclplot to plot the condence intervals for the Somers D parameters, and save the plot in an Encapsulated PostScript le to form Figure 2. Negative predictors of non-US origin appear above the possible non-predictor price, which appears above positive predictors of non-US origin. Note that eclplot has a reverse default vertical axis scale if the user species a horizontal condence interval plot. We then produce a table. First, we use sdecode to convert the numeric condence limits to string variables of the same names, use replace to add parentheses and commas to these string variables, and list the reformatted condence intervals. Next, we use listtex to type the dataset to the Results window in an alien-looking style as follows: . listtex predictor estimate min95 max95, headline("Predictor&Somers D&(95%&CI)") type Predictor&Somers D&(95%&CI) Displacement (cu. in.)&-0.84&(-0.96,&-0.72) Turn Circle (ft.) &-0.79&(-0.94,&-0.64) Weight (lbs.)&-0.75&(-0.91,&-0.59) Length (in.)&-0.72&(-0.89,&-0.56)

From datasets to resultssets in Stata Trunk space (cu. ft.)&-0.47&(-0.69,&-0.24) Headroom (in.)&-0.37&(-0.61,&-0.14) Price&0.15&(-0.14,&0.45) Mileage (mpg)&0.46&(0.19,&0.72) Gear Ratio&0.87&(0.76,&0.98) . more

The typed output (which appears bright yellow in the Results window) can then be copied from the Results window, pasted into a Microsoft Word document, selected inside Microsoft Word, and converted to a Microsoft Word table using the menu sequence Table->Convert->Text to Table. Alternatively, it can be pasted into a Microsoft Excel document and converted to Microsoft Excel columns using the menu sequence Data->Text to Columns. Note that listtex can also produce tables that can be cut and pasted into TEX, A L TEX or HTML documents, usually with even less work, because Microsoft Word requires users to do their own table justication, autoformatting and italicizing. Figure 2: Somers D for quantitative diagnostic predictors of non-US origin in the auto data.

Displacement (cu. in.) Turn Circle (ft.) Predictor of nonUS origin Weight (lbs.) Length (in.) Trunk space (cu. ft.) Headroom (in.) Price Mileage (mpg) Gear Ratio

1.00 0.75 0.50 0.25 0.00 0.25 0.50 Somers D (95% CI)

0.75

1.00

3.2

String-factor conversion using factext and factmerg

Often, the predictor variables in an estimation are dummy variables, indicating membership of a group dened by a value of a categorical variable. Such categorical variables are known as factors in the classical Rothamsted terminology, as used by the statistical packages GLIM and Genstat. The dummy variables are usually produced by xi (see [R] xi), but may also be produced by tabulate (see [R] tabulate), or by the desmat command of John Hendrickx (introduced in Hendrickx (1999), and updated in Hendrickx (2000, 2001a and 2001b)). Their variable labels are usually of the form "factor name==value" (1)

(where factor name is a valid Stata variable name and value is a value of the variable), and will be saved in parmest and parmby resultssets in the variable label, if the label option is specied. To produce plots like Figure 1, it would be useful to regenerate these categorical variables in the resultssets. The factext program allows us to do this. It takes, as input, a list of string variables, which defaults to a single variable named label, as found in a parmby output. It generates, as output, a set of factors, with names extracted from the left-hand side of input string values of the form (1), and values extracted from the right-hand side of the same input string values. These output factors are usually numeric.

From datasets to resultssets in Stata

String values of the form (1) do not contain enough information to reconstruct a numeric factor complete with its storage type, display format, value labels, variable labels and characteristics (see [U] 15.8 characteristics or [P] char). However, factext can be helped by descsave, which is not only a resultsset generator, but also a program generator. descsave has a dofile() option, which allows it to generate a dole, which, if run, can reconstruct the variable attributes of the variables described, assuming that variables of the same name and mode (numeric or string) exist in the current dataset. factext also has a dofile() option, which allows it to run this do-le after extracting the factor values. The factors in the resultsset will then have the same storage types, display formats, value labels and variable labels as the factors of the same names in the input dataset, and also (optionally) the same values of characteristics specied by the user (such as the omit characteristic mentioned in [R] xi). The factext program may cause a resultsset to be wide and sparse by generating a long list of factors, each of which may have missing values in most observations, because most of the parameters do not belong to dummies belonging to that factor. In Table 1 and Figure 1, by contrast, there is a single column of row or axis labels, containing information on both factors (foreign and rep78). The factmerg program converts in the opposite direction to factext, and can be useful in generating such row label variables. It takes, as input, a list of factors (usually numeric), and generates, as output, up to three string variables, containing, in each observation, the name, variable label and string value, respectively, of the rst factor in the list with a nonmissing value. These string variables can be used in further processing to create row label variables for multi-factor plots and tables, such as Table 1 and Figure 1. The program in example4.do demonstrates the use of descsave, factext and factmerg. After loading the auto data, we assign reference levels to the factors foreign and rep78 using the omit characteristic, and use descsave to create a temporary do-le to reconstruct these factors. We then use parmby to replace the dataset in memory with a resultsset, with one observation per parameter of a regression model and data including a string variable label. Next, we use factext to extract the factors foreign and rep78 from this variable, list them, and create a single-factor condence interval plot for the dierences between cars with non-reference and reference repair records, which is saved as Figure 3. After that, we use factmerg to merge the two factors and generate string variables faclab and facval, containing the variable label and value, respectively, of the factor corresponding to each parameter. We can now use generate and sencode to create a labelled numeric row variable cargp, identifying car groups with non-reference car origin type or repair record. Finally, we produce a multi-factor condence interval plot for the dierences in mileage associated with these non-reference car groups as Figure 4. Figure 3: Dierences in miles per gallon by repair record (compared to reference category 3).

1 Repair Record 1978

5.00

0.00

5.00 10.00 Difference (miles per gallon)

15.00

From datasets to resultssets in Stata

Figure 4: Dierences in mileage in the auto data (compared to US car with rep78==3).

Car type: Foreign

Repair Record 1978: 1 Car group

Repair Record 1978: 2

Repair Record 1978: 4

Repair Record 1978: 5

5.00

0.00 5.00 10.00 Difference (miles per gallon)

15.00

Adding reference and gap observations to resultssets

Figure 4 is not Figure 1. The reference categories are absent. The axis labels contain no group numbers. Also, there is little space for group numbers, because the labels are very repetitive, repeating factor labels instead of putting them in gaps on the axis. The rst two problems can be solved by merging xcontract resultssets into the parmby resultsset. The third problem can be solved by adding gap observations to the resultsset, using the ingap package, downloadable from SSC. The ingap package is discussed in Newson (2003). It adds gap observations to the dataset in memory, at a user-specied list of positions within the dataset, or within each by-group if by-groups are specied. Existing variables in the gap observations (apart from by-variables) are set to missing, unless the user species otherwise. Usually, one string variable is specied as a row or axis label variable by the rowlabel() option, and its values in the gap observations are specied by the growlabel() option or by the grexpression() option. The program in example5.do is a more advanced version of example4.do, and uses xcontract and ingap to produce Figure 1 and Table 1. After creating the do-le using descsave, and before creating the parmby resultsset, we create two xcontract resultssets, containing the frequencies of foreign and rep78 in cars with nonmissing repair record (included in the regression). After extracting the factors from the string variable label, we merge in these xcontract resultssets, and then set the condence interval variables estimate, min95 and max95 to zero in observations with reference values for the two factors. After using factmerge to create faclab and facval, we generate the row variable cargp, adding gap observations using ingap in the process. In these gap observations, the value of cargp is set by the grexpression() option of ingap. We then plot the condence intervals against cargp as before, this time creating a better-looking graph (Figure 1). Finally, we use sdecode and replace to convert the condence limits to strings with A parentheses and commas in the right places, and use listtex to list to the Results window a L TEX tabular environment as follows: . listtex cargp estimate min95 max95 p if !missing(factor), rstyle(tabular) type /// > headline("\begin{tabular}{rrrrl}" "\hline" /// > "\textit{Car group}&\textit{Difference (mpg)}&\textit{(95\%}&\textit{CI)}&\textit{P}\\\\" /// > "\hline") /// > footline("\hline" "\end{tabular}") \begin{tabular}{rrrrl} \hline

From datasets to resultssets in Stata \textit{Car group}&\textit{Difference (mpg)}&\textit{(95\%}&\textit{CI)}&\textit{P}\\ \hline Car type:&&&&\\ Domestic (N=48)&0.00&(ref.)&&\\ Foreign (N=21)&3.56&(0.32,&6.80)&.032 \\ Repair Record 1978:&&&&\\ 1 (N=2)&1.92&($-$2.78,&6.62)&.42 \\ 2 (N=8)&0.05&($-$2.98,&3.07)&.98 \\ 3 (N=30)&0.00&(ref.)&&\\ 4 (N=18)&0.81&($-$2.06,&3.68)&.57 \\ 5 (N=11)&5.38&($-$0.58,&11.33)&.076 \\ \hline \end{tabular} . more

10

A This typed output was cut and pasted into Table 1 of the L TEX version of this document. The large program in example5.do may seem to be a very complicated way of creating a small plot and a small table. However, there are often economies of scale if the user needs to create large plots and tables, or a lot of plots and tables, or multiple versions of the same plot or table. An alternative approach to drawing plots with gap labels is to use the egen function axis(), which is part of the egenmore package, written by Nicholas J. Cox and downloadable from SSC (see [R] egen). This approach has the feature that it does not add gap observations to the resultsset. This feature may or may not be an advantage, depending mainly on whether or not the user wishes to use the same resultsset as input both for eclplot and for listtex. It is often a good idea to use the preserve and restore commands (see [P] preserve) before and after a sequence of resultsset-modifying commands such as those in example5.do.

Acknowledgements

I would like to thank Nicholas J. Cox of Durham University, UK, for coining the term resultsset and for many helpful comments on this document; Christopher F. Baum of Boston College, MA, USA for distributing on SSC the resultsset-generating and resultsset-processing programs demonstrated here (and many more); and all the members of the Stata community who have given me their constructive advice on developing these programs.

References

Cox, N. J. 2002. Speaking Stata: On numbers and strings. The Stata Journal 2(3): 314329. Hendrickx, J. 1999. dm73: Using categorical variables in Stata. Stata Technical Bulletin 52: 28. In Stata Technical Bulletin Reprints, vol. 9, 5159. College Station, TX: Stata Press. Hendrickx, J. 2000. dm73.1: Contrasts for categorical variables: update. Stata Technical Bulletin 54: 7. In Stata Technical Bulletin Reprints, vol. 9, 6061. College Station, TX: Stata Press. Hendrickx, J. 2001a. dm73.2: Contrasts for categorical variables: update. Stata Technical Bulletin 59: 25. In Stata Technical Bulletin Reprints, vol. 10, 914. College Station, TX: Stata Press. Hendrickx, J. 2001b. dm73.3: Contrasts for categorical variables: update. Stata Technical Bulletin 61: 5. In Stata Technical Bulletin Reprints, vol. 10, 1415. College Station, TX: Stata Press. Newson, R. 2000. somersd Condence intervals for nonparametric statistics and their dierences. Stata Technical Bulletin 55: 4755. In Stata Technical Bulletin Reprints, vol. 10, 312322. College Station, TX: Stata Press. Post-publication update downloadable from http://www.kcl-phs.org.uk/rogernewson/ as of 20 April 2004. Newson, R. 2002. Parameters behind nonparametric statistics: Kendalls tau, Somers D and median dierences. The Stata Journal 2(1): 4564. Pre-publication draft downloadable from http://www.kcl-phs.org.uk/rogernewson/ as of 20 April 2004. Newson, R. 2003. Condence intervals and p-values for delivery to the end user. The Stata Journal 3(3): 245 269. Pre-publication draft downloadable from http://www.kcl-phs.org.uk/rogernewson/ as of 20 April 2004.