Stata makes it easy to manipulate names, or more specifically, variable names, as in a dataset with three variables for social expenditure called party1 party2 party3. This is common to many empirical preprocessed datasets.

    // example
    mvdecode party*, mv(999)

Furthermore, Stata works like an accountant’s book, so all variables belong to a same data object that never needs to be called beyond loading. This naturally suppresses a lot of possibilities, compensated in part by macros and scalars.

    // example
    loc regressors "age sex"

Macros in particular then branch with loops like the forval and foreach commands to allow more complex data processing. At that level of use, the software is flexible enough for most applied data cleaning.

    // example
    forval i = 1/3 {
      replace socx`i' = socx`i' / 10^6
    }

To access matrix notation, the Stata user needs to move to Mata syntax, while R immediately offers the user to manipulate objects through vectorization. Thinking in these terms is more demanding as there are more possibilities for errors, starting with calls to undeclared objects.

I teach both R and Stata. My experience with social science students is that the additional learning costs of R syntax need to be matched with other benefits to become valuable to them. To me, these benefits lie primordially in the more diverse array of data that R allows to access.

This post will dig a little deeper in this area by presenting limited dependent variable’s type and relevent Stata cracking commands. The methods are:

Truncated Regression

Censored Regression (Tobit model)

Incidental Trucation Regression (sample selection model, Heckman model)

Treatment Effects Model

Instrumental Variable Regression

1, Truncated Regression

Truncated regression deals with truncated data. Then, what is truncated data? It shows like graphs below:

Left truncation(do neglect the numbers in the middle and on the axises):

Right truncation:

Two sides truncation:

If you meet this kind of data distribution, no matter what the actual specific situation you meet, you can use Stata command “truncreg” to solve the problem.

truncreg y x1 x2 x3, ll(#)  (lower limit, Left truncation)

truncreg y x1 x2 x3, ul(#)  (upper limit, Right truncation)

truncreg y x1 x2 x3, ll(#) ul(#)  (lower and upper limits, Two sides truncation)

# is the value of where the truncation happens.

The # should be chosen carefully because it could affect the regression outcomes. The number of obs that out of the # threshold does not affect the regression outcome.

2, Censored Regression (Tobit model)

Censored regression deals with censored data. Then what is censored data? If the data is censored at the upper limit of 5000, then any value that should be larger than 5000 shows to be 5000 in the dataset.

For this kind of data, we use “tobit” command to analysis.

tobit y x1 x2 x3, ll(#)  (lower limit, Left censored)

tobit y x1 x2 x3, ul(#)  (upper limit, Right censored)

tobit y x1 x2 x3, ll(#) ul(#)  (lower and upper limits, Two sides censored)

The # should be chosen carefully because it could affect the regression outcomes. The number of obs that at the # threshold does affect the regression outcome.

3, Incidental Trucation Regression (sample selection model, Heckman model)

Incidental trucation regression, or sample selection, means the data have some kind of bias. Kinds of bias are elaborated in books of Epidemiology. To correct the bias and get the effect of factors on the whole population, we use “heckman” command.

heckman y x1 x2 x3, select(z1 z2)    /*using MLE method, and the dependent variable of selcetion equation is y*/

heckman y x1 x2 x3, select(z1 z2) twostep    /*using twostep method, and the dependent variable of selcetion equation is y*/

heckman y x1 x2 x3, select(w=z1 z2)    /*using MLE method, and the dependent variable of selcetion equation is w. Variable w won’t show up in the main(second step, if twostep) regression equation, even you select it purposely.*/

Things to remember, any un-observed y should be set into missing.

4, Treatment Effects Model

Treatment effects model is derived from heckman model and mainly used to realize the mission of program evaluation.

It has two merits compared with Heckman model mentioned above: 1, In the heckman model, only part of dependent variables can be observed. 2, In the heckman model, intervention variable, which is the potential dependent variable of selection equation, cannot get into the main regression equation.

The command for this model is “treatreg”:

treatreg y x1 x2 x3 w, treat(w=z1 z2) [twostep]   /*w is usually the intervention variable, which defines the treatment group and control group*/

The rho variable in the output is important, the hypothesis test of it in the bottom of the output tells us whether it’s appropriate to choose this model. (It should be non-zero.)

5, Instrumental Variable Regression

Instrumental variable regression is a method to solve the problem that the independent variable have contemporaneous correlation with error term, when there are some explainary items are not included in the equation. So, this method uses an instrumental variable to solve this problem. This IV is hard to determine because it needs to fit the criteria “the instrument should be correlated with endogenous explanatory variables, but cannot be correlated with the error term”.

Here is the command:

ivregress 2sls/liml/gmm y x1 x2 x3 (x4=iv1 iv2 iv3)  /*x4 is the variable that correlated with the error term*/

ivprobit y x1 x2 x3 (x4=iv1 iv2 iv3), [twostep]   /*using MLE or twostep method*/

Also, it is recognized that because it is hard to find the appropriate instrument, it’s fine to use treatment effects model to insteat IV method.

All the above is the five methods I would like to introduce, later I may revise this post if I find any mistake. If the reader of this article finds any mistake, or has any suggestion, please kindly let me know, thank you!

First, you should have the STATA software and program adds-on. By

help pscore

install (one of the suits, usually the newest one)

Then, after you load the dataset, you can begin to calculate the PSM.

The example given by ”Microeconometrics: Methods and Applications” used global command to set the global variables, if you don’t like it, you can just jump this global part and use real numbers or words names in where there is a $.

global breps 200      /*set 200 to variable breps, and indicate the times bootstrap process will repeat*/

global vars name1 name2 …namen  /*just to save efforts of repeating writing the variable names*/

Next, we come to the core processors of PSM calculating. Pscore and PSM regression.

Pscore:

pscore TREAT $vars, pscore(myscore) comsup blockid(myblock) numblo(5) level(0.005) logit

or

pscore TREAT $vars, pscore(myscore) blockid(myblock) numblo(5) level(0.005) logit

TREAT is the var which indicats where the obs is a treated or a controlled. To perform regression on it, is based on the theory that all the obervations should have the same probability to be treated or controlled. In the other word, it should be a independent variable.

Pscore, blockid, numblo, level, logit(the default is probit model), means to get pscore, block indicator, set number of blocks, and significant level when selecting varialbes, also, by the method of logit model.

The comsup, which is the sole difference between the two command, is used to set whether to use the common supported part, the part that both group, treat and control, share.

PSM regression:

set seed 10101
attnd RE78 TREAT $vars, comsup boot reps($breps) dots logit

set seed 10101
attr RE78 TREAT $vars, comsup boot reps($breps) dots logit radius(0.001)

set seed 10101
atts RE78 TREAT, pscore(myscore) blockid(myblock) comsup boot reps($breps) dots

set seed 10101
attk RE78 TREAT $vars, comsup boot reps($breps) dots logit

The four groups of commands all perform PSM regression, but use different method, specifically, Nearest neighbor matching, Radius matching for Radius=0.001, Stratification Matching, Kernel Matching, respectively.

Boot reps($breps) is showing that we are using bootstrap method and repeat $breps times. Dots is interesting dots in the screen if you like to see. Attention should be taken that different methods using different numbers of observations.

Ok, all above is a simple, short command to perform PSM. Enjoy your research!

I have been deceived by it for such a long time…

Here is the command of applying DID in STATA:

regress EARNS Tdyear2 TREAT dyear2

OLS regresssion, with some dummy variables. Among the command, EARNS stands for the outcome you want to measure, and TREAT and dyear2 stand for the treat and time, respectively. The KEY variable Tdyear2, is the multiplication of TREAT and dyear2, which, therefore distinguish the treat group on the second year with all the other 3 parts of the observations.

So, the paramater of Tdyear2, is the treat effect.

If you use the command of “regress EARNS Tdyear2 TREAT dyear2, robust”, then you can calculate heteroskedastic-robust standard errors.

Things to remember:

DID的基本假设：
1， 共同趋势假设：假设不同组别的时间效应都是相同的。
2， 假定两组的合成部分在变动前后均是稳定的。

From my student files.

    cap pr drop find
    program find, rclass
        qui lookfor `*'
        if "`r(varlist)'" != "" codebook `r(varlist)', c
    end

This returns the output of the codebook command with the compact option instead of the standard lookfor output, which is based on the describe command. The variable labels are less readable, and long variable names still get abbreviated (such a strange idea). Yet it works, and returns the N.

I tried to call the program look, but that is still taken by a deprecated command, lookup, which has been superseded by search. As for the single letter l, it is taken by the list command, so that will not work either. The same holds for q, and ? is sadly not a valid program name.