Stata is statistical analysis software used commonly in social sciences. It is known for it’s ease of use, robust support for complex survey design, and comprehensive and clear documentation.
Stata (pronounced either of stay-ta or stat-ta, the official FAQ supports both) is primarily interacted with via typed commands written in the Stata syntax. There is a powerful GUI that you can access via midesktop, or you can SSH into one of the university Unix servers and run Stata at the command line (via the stata command).
Once you have Stata up and running, the simplest form of use is as a calculator.
. display 2 + 4The . is often used for documents like this to note a command; whenever you run a command in Stata, the first thing it does is echo back the command you ran prefixed by the ..
The command name is “display” and it outputs the requested expression. This is a trivial example, but display is much more flexible.
Most Stata commands follow the same basic format:
. command <variable(s)>, <options>The command can have sub-commands. The number of variables which needs to be/can be passed obviously varies by command. The order of variables may matter (for example, any regression model treats the first listed variable as the response.)
The options are space-separated words (e.g. robust) or words with options (e.g. level(.90)). Almost all commands support some number of options, most commands do not require any options.
For example, a regression would be as simple as
. regress y x1 x2, noconstantThis is defining a linear regression model predicting y based upon the continuous variables x1 and x2. The noconstant operation forces \(\beta_0 = 0.\)
Stata supports abbreviations for most commands. For example, the regression command could have been typed as:
. reg y x1 x2, noconsI would not recommend using abbreviations in most cases (a lot aren’t as clear as reg), but you may see others use them.
Most commands support an if or in option to operate on a subset of the data. For example,
. regress y x1 x2 if female == 1performs linear regression only on the female subsample. in operates on specific rows, and may be useful for testing slow running code:
. regress y x1 x2 in 1/100The 1/100 syntax is equivalent to writing 1 2 3 4 .... 99 100.
Stata has amazing documentation. Use the help _____ syntax to look up documentation on any command. Help files contain the syntax of the command, a list of options allowed, a description of the command, and a number of examples. Additionally, the very top of each help file contains a link to the appropriate location in the full Stata manual which typically has further examples. Use these help files liberally.
Note that running help on an abbreviation (e.g. help reg) pulls up the help correctly. This is useful when reading others code. Additionally, in the syntax and options listing, you’ll see partial underlining. These indicate the required abbreviations.
One quirk of Stata is that only a single data set can open at a time. If you want to operate on multiple data sets, you either need to be constantly switching between them, or merge them into one file (for help with merging, look into the append and merge commands).
The benefit of having a single data set open is that you never need to refer to it - any command you give must operate on the open data set.
You can access files with:
use to load a local file, or when passed a URL, an online file.sysuse or webuse to access Stata example data sets (often used in Help examples)import excel and import delimited for importing Excel or CSV files.For importation, I’d recommend using the File -> Import menu instead of trying to type the command. The Import dialog box has a live preview as you change settings which can be very useful. After you run the import, the command it generated is echoed and you can save it to run again.
You can save a data set with the save command which works as
. save newfilenamePassing the replace option allows Stata to overwrite an existing file.
If you try and open a new data set when there are un-saved changes in the existing data, Stata will refuse. The clear command will remove the existing data and allow you to load a new data set. You can also pass clear as an option to use/sysuse/webuse/import to save one step.
The preserve and restore commands can be very useful for switching between files or making destructive modifications to an existing data set. preserve takes a snapshot of the data as it currently is, and restore switches back to it. To see an example of it, consider the collapse command. The collapse command generates a summary data set by collapsing the existing data by a given variable and creating summary statistics. For example, if we have some sort of census data,
. collapse (mean) bmi (percent) female, by(state)This would replace the existing data set with a new one with one row per state, containing a variable indicating which state, a variable indicating the average BMI in that state, and a variable indicating the percent of the state which is female.
collapse is a destructive command; after running it the original data (and any unsaved changes) is lost. We can wrap this in preserve and restore to save the descriptive data and recover the original data.
. use data1
. preserve
. collapse (median) x1 x2 (min) x3 (sum) x4, by(group)
. save summarydata
. restoreAfter running the above five lines, “data1” will still be open, but we’ll have a new file “summarydata”.
The following commands may be useful for exploring the data:
list prints the data to the output window. list <variable name> prints only the requested variables. Combine with if and in to make the output more reasonable.browse (or edit) to browse (or edit) the data in an Excel-style window. Combine with if and in. (Only works in GUI.)describe displays some information about the variable, including it’s type (string or numeric) and whether it has labels attached to it.summarize, codebook, tab, mean display some descriptive statistics.Stata scripts are called “Do-files”, named for their extensions (.do). In the GUI, you can open a new Do-file for editing by typing doedit. You can run Do-files interactively by highlighting the desired code and hitting the “Execute (Do)” (Windows) or “Do” (Mac) buttons.
Do-files can also be run from start to finish by using the do command. Alternatively, if you are accessing Stata at the command line via SSH, you can simply run stata mydofile.do to launch Stata, run the Do-file, and exit.
Stata uses some programming terminology in unique ways. Specifically:
regress, use, display.log(salary), log() is the function).In addition to the one data principal, Stata also operates on one estimation command at a time. An estimation command is any command which produces some inference, for example mean to obtain the mean and a confidence interval, or regress as we discussed above. save or preserve are not estimation commands. After running an estimation command, Stata provides access to postestimation commands and returned objects.
Estimation commands support commands that follow them and use their results. For example, after a regression command we can obtain the AIC and BIC by running
. estat icNote that we do not refer to a specific model. You can see a list of all postestimation commands supported by running help ____ postestimation. For example, help regress postestimation.
After running an estimation command, the command may create several objects which can be examined. Estimation commands are either of R-class or E-class (the distinction is not very important). What is important is that E-class commands store things in the ereturn while R-class commands store in the return.
For example, after running regress, the object r(r2) contains the \(R^2\) for the model.
. regress y x
. display e(r2)Matrices can also be returned, but cannot be accessed with display, instead we use matrix list:
. matrix list e(V)will return the variance-covariance matrix of the estimators.
You can see a full list of the returned objects via return list or ereturn list. The help file for each command will also describe what the command returns.
We will see how these objects can be manipulated after we discuss macros
Macros in Stata use a simple substitution evaluation system. We define a macro with the local command.
. local myvars x1 x2 x3
. regress y `myvars'The local command stores whatever is past the name of the macro, then when it is referenced (via the backtick and single quote) is substituted into the command and the command executed. For example, when you run those regress line above, Stata will replace “`myvars’” with “x1 x2 x3” and then execute the regression model.
Macros can also store numeric values and can be operated on.
. local x 3
. display `x'
. local y = `x' + 2
. display `y'You’ll notice the first local call contains no equal sign, whereas the second does. The equal forces immediate evaluation. You can see the difference by running:
. display `y'
. display "`y'"
. local y2 `x' + 2
. display `y2'
. display "`y2'"You can store returned objects and operate or display them.
. regress y x
. local r2 e(r2)
. display "The model R^2 is " `r2'Matrices can be stored as well
. regress y x
. matrix v = e(V)
. display "The variance/covariance matrix:"
. matrix list vNote that when referring to matrices, no tick/quote is needed (however, you cannot refer to it as generally as you can a macro).
Loops are another place where macros are used often. The syntax is very similar to other languages.
foreach <macro name> of <list of numbers/words/variables> {
..
}The list can be:
numlist 1/5 or numlist 1 4 29 192a b cvarlist x y zFor example, to regress over a series of outcome variables:
foreach var of varlist y1 y2 y3 {
regress `var' x1 x2
estimates store reg_`var'
}The estimates store command saves the regression results so that you can restore them later using estimates restore to make them the most recent estimation command, or use a command such as outreg to produce an output document.
Alternatively, if the variables names are that clean, you could loop over just the number:
foreach i of numlist 1/3 {
regress y`i' x1 x2
estimates store reg_y`i'
}drop/keep - Removing columns (using drop <varname>) or rows (using drop if <conditional>).generate - Creating new variables using syntax generate <newvarname> = <expression>. The expression can depend on other variables.replace - Modify an existing variable. Commonly used with if.destring/tostring - Convert strings that are really numbers to strings, and vice-versaencode/decode - Convert strings with words to numbers with associated labels, and vice-versa.Mata is a matrix programming language which is part of Stata. It has two primary uses:
While objects from Stata can be passed into Mata and vice-versa, most of the time we can think of Mata as completely independent of Stata.
Mata is entered by using the mata command.
. mataWhen in Mata mode, the prompt changes from a . to :. Stata commands will not be accepted inside Mata. To exit Mata, enter the end command.
: endYou can enter mathematical expressions directly into Mata:
: 5-4When inside Mata, you can define and use “variables” in the same sense as R.
: x = 4
: x + 2It also supports functions, e.g.
sqrt(4)log(x)Mata sessions have permanence. If you end a Mata session and then invoke a new session, it retains the same variables.
You can also run a single line of Mata with the mata: preface. After running this command, you will be in Stata, not Mata.
. mata: 2 + 2To define a Mata matrix, we can combine the column-join operator , and the row-join operator \. We can print a matrix by calling it alone.
: M = (1,2\3,4)
: M(Note that I use lowercase letters for scalars and upper case letters for matrices. This is not enforced by Mata and just a preference.)
Matrix operations work as expected. We can:
mata: 4*M, mata: M + 2mata: A + B, mata: A*B (an error if dimensions are not compatible)mata: A'mata: A\B would stack A on top of B (assuming dimension compatibility)mata: A:*Bmata: C = I(5)mata: D = J(4, 2, 0)We can pass matrices between Stata and Mata using the st_matrix() function. This functions works inside of Mata.
Let’s say we run a regression in Stata:
. sysuse auto
. regress mpg headroomWe saw before that e(B) contains the variance/covariance matrix, but let’s obtain the standard errors for the coefficients.
. matrix v = e(V)
. matrix list vTo manipulate these, we can pass them into Mata.
. mata:
: V = st_matrix("v")
: SE = diagonal(sqrt(V))
: st_matrix("se", SE)
: end
. matrix list seThe diagonal function extracts only the diagonal of the matrix. (Confusingly, as opposed to the diag, which only sets off-diagonal elements to 0.)
marginsOne of Stata’s most popular tools is the margins command. It is a very complex command (the help manual for this one command is ~55 pages long), but extremely powerful. We’ll explore two common uses.
First, when running a regression with a categorical variable, one level of the categorical variable is the reference, and all other groups are tested against it. If you wanted to test between variables which weren’t the reference, you need to either change the reference category or develop a contrast. margins does this second step automatically.
. regress y x i.z
. margins z
. margins z, pwcompareFirst, note the i.z variable in the regress command. All Stata models assume by default that variables are continuous. To treat a variable as categorical, preface it with i.: E.g. variable race would be, in the model, i.race.
The first margins command will estimate the marginal mean for each level in z. This is done by taking the original data, assuming in every row of data that z takes on it’s first group, and then using the regression equation to estimate each response and generate the average. Then it assumes each row of data that z takes it’s next value, and so on.
The second margins command adds the pwcompare option which will generate all contrasts of comparison between each level of z. By default it produces a confidence interval for each contrast; you can obtain a p-value by passing pwcompare(pv) instead.
The second use of margins is for the creation of interaction plots. Interactions are entered into regressions via x##z; this includes the main effects of both x and z as well as their interaction. When a variable is involved in an interaction, Stata assumes it is categorical; you can use c. to treat it as continuous.
. regress y c.x##i.z
. margins z, at(x = (1 2 3 4 5))
. marginsplotThis margins call is obtaining the marginal mean of z at each of those 5 values of x. marginsplot is a post-post-estimation command which can be run after margins to produce a plot (another extremely powerful yet complex command, another 35 pages of documentation).
We can test whether the slope is the same is several subgroups with margins as well:
. margins z, dydx(x)
. margins z, dydx(x) pwcompare(pv)The dydx() option estimates the slope on a continuous variable by taking the slope of the regression equation relative to the continuous variable. By passing the categorical variable z we’re asking for the slope in each group, and then testing them against each other next.
Professor Shedden’s Stata Intro
Dr. Henderson’s Stata Notes from a previous semester
Two sets of notes I use for a more beginner-focused, more in depth Stata workshop