10 <arch 2008
Eric Rasmusen, erasmuse@indiana.edu
This file is now at http://www.rasmusen.org/a/stata-rasmusen.txt

SOME STATA COMMANDS I USE

This file is for notes on STATA commands that I use. I use Stata 7.0
on my desktop. Stata 9 is on  the  Indiana U. Libra and Steel
computers.

---------------------------------------------------------------------

HELP: FINDIT
 If I type Findit rreg from the command screen, I will get Web files
that are useful, including, e.g. UCLA statistics help files.


---------------------------------------------------------------------

 BASIC STATA
   *jkjlkjlk is a way to do comments

cd d:/mystuff/mydirectory  * Gets to the directory with your data

insheet  using crimedata.csv *Inputs your spreadsheet data


 ds  *Lists all the variable names. use the option ,fullnames to get the
labels too.

summarize *summary statistics for all variables

regress crime  population prison
 *Regresses crime on population and prison

xi:regress crime   prison i.state i.year
*Regresses crime on   prison  and dummies for states and years

generate lnprison = ln(prison)
 *Creates a new variable with a math formula

replace logprison = 1000*lnprison if logprison== 1
 *replace an observation of the variable if the condition is met



---------------------------------------------------------------------
HISTOGRAM DATA FIGURES GRAPHS FORMAT

The "scheme" for graphics called lean2 must be installed from teh web.
Try stata help for lean2 to find out how it's easy). THen use these
commands:

set scheme lean2;

histogram flunks if judge==0,  discrete  freq    ylabel(,nogrid)
xtitle("Flunks by Lawyers")  saving(flunks-l,replace) ;
*This one has no y-axis  grid lines;

histogram flunks if judges>0,  discrete  freq    xtitle("Flunks--
Lawyers") ytitle("First line of the   title" "Continues to a second
line")  saving(flunks-l,replace) ;
*This one has  y-axis     grid lines, and a special y-title
with two lines of text;


graph export flunks-l.ps, replace;

Stata will make the y-axis label vertical writing, rather than
horizontal. I go to Adobe Illustrator to change that.


Use help twoway_options to find out how to use the options.


---------------------------------------------------------------------

USING WEIGHTS IN A REGRESSION

* If I want to weight   observations  1 unless the judge variables
equals 1, and
then use .04;

generate wj = 1;
 replace wj = .04 if judge==1  ;


dprobit  judge utokyo ukyoto     Flunks413 Post93 Post93UT Post93UK
Post93
Fl413    [pweight=wj]



----------------------------------------------------------------------

 MERGING IN A NEW VARIABLE INTO A STATA DATASET

 Mark tells me that this command does it, putting the new variables in
wherever the

joinby prefecture using "file name.dta"

----------------------------------------------------------------------

FORMATTING REGRESSION OUTPUT: OUTREG AND EXCEL

Stata's standard regression output has too many decimal places and
some useless statistics. To convert one or more regressions  to a
table form of the kind customary in economics, use one of two methods:
the Outreg program, or cutting and pasting to Excel.

 The best way to do it is to use outreg, with output in a file like
temp.tex (with long enough lines, few carriage returns). In  the
temp.tex file change all the left parentheses to yyy. Then load it
into Excel simply by opening temp.txt and choosing comma delimiters.
Then copy into WORD or LATEX. THen change all the yyy to left
parentheses.


FIRST, OUTREG:

The add-on program outreg.ado is  available from:

http://econpapers.repec.org/software/bocbocode/s375201.htm

Documentation is at:

http://www.kellogg.northwestern.edu/researchcomputing/docs/outreg.pdf

    I think you can download and install it in one step by issuing the
command in stata

ssc install outreg

 In my PC's Stata 7.0, tho, that doesn't work. Instead, first type
 net from http://fmwww.bc.edu/repec/bocode/o/
 and then type
  net install outreg


Here is how to use the program.
After each regression command such as "regress yvar xvar x2 x3 x4"
type,  for the first regression,

outreg using table1.txt, replace  bdec(2)

and for every succeeding similar regression,

outreg using table1.txt, append  bdec(2)

 What this does is to add the latest regression results as a new
column in  a table in the file table1.txt. The top row will  have the
Y variable (which can differ). Coefficients and t-stats will be to 2
decimal places. The tabs won't work
properly, so the table's spacing won't be right, though.

 To get proper spacing, you   can also outreg into a comma-separated
file,  by adding an option like this:

outreg using table1.txt, replace  bdec(2)       comma


Then cut and paste to MS Word. Select all the text except the notes at
the bottom and click on  Table, Convert Text to Table. It will format
it nicely. Then if you cut and paste to a plain text file, the good
spacing remains.

 If you use dprobit dtobit, etc., you will get only MARGINAL EFFECTS
reported.
That is usually best. Then outreg will automatically report those too.
Otherwise, try:

margin[(u|c|p)] specifies that the marginal effects rather than the
coefficient
    estimates are reported.  It can be used after truncreg,marginal
    from STB 52 or dtobit from STB 56.  One of the parameters u, c, or
p
    is required after dtobit, corresponding to the unconditional,
    conditional, and probability marginal effects, respectively.  It
is not
    necessary to specify margin after dprobit, dlogit2, dprobit2,
    and dmlogit2.



 SECOND, EXCEL
    The other way to  format a regression table is to cut-and-paste
from teh
STATA log on teh screen (not from the *.log file) using EDIT and then
SAVE TABLE
(NOT the regular Save;SAVE TABLE put in tabs between columns.)
  Paste it into a
plain text file, or, perhpas into MS-WORD directly (and convert text
to table
then, I guess). Save the plain text file as temp.txt, then open it in
Excel, and
it will be a nice spreadsheet  with columns and rows that can be cut
and pasted.


Note that in Excel you can also add columns of "(" and ")". Then save
as  a
*.csv file, with comma delimiters. Edit that as plain text, to get the
parentheses in the same columns as their contents,  and upload into
Excel again.


USING BOTH
 Start with temp.txt. Then chagne all the commas to tabs (mabye this
happens
anyway if you don't use the COMMA option in OUTREG). Then make a
spreadsheet
file and pick TEXT as the default for Nubmer Format. Then copy and
past the
temp.txt file to the spreadsheet file. It will save with parentheses
around the
t-stats, instead of making them negative.

BETTER: in the temp.txt file change all the left parentheses to yyy.
Then load
it into Excel simply by opening temp.txt and choosing comma
delimiters. Then
copy into WORD or LATEX. THen change all the yyy to left parentheses.


  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

X-WINDOWS
 If I am trying to use the Indiana U. computer remotely to use Stata
9.1, load
up an x-terminal and then type:


ssh erasmuse@libra.uits.iu.edu

or

 ssh erasmuse@steel.ucs.indiana.edu

and issue the command:

xstata

 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

 /*

  This is a multline comment. No semicolons are
needed.


 */

 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;
  #delimit ;
 * This says that the semicolon denotes the end of a line of command.
All lines
must end with semicolons after this;

log using marginal-effects.log, replace;

set more 1;
 *This should stop the pauses;
 *To keep going regardless of errors, use DO mYFILE, nostop;

   * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

FIXED-EFFECTS REGRESSIONS

 This creates dummies out of verbal or numeric categories.

xi: tobit yvar  xvar1 xvar2 i.categoryvariable, ll(10000) ul(30000);

OR

xi: regress yvar  xvar1 xvar2 i.categoryvariable, robust;



 * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;


*PLOTTING TWO VARIABLES AGAINST EACH OTHER;

graph murder black;

graph murder black, symbol([state]);

graph murder black, symbol([state]) psize(150);

*this makes the symbol size 50 percent larger than the
default

 graph  score cashleft, symbol([rank]) xscale(0, 40000)  yscale(30,
110)

When I tried this in Stata 9.2 in dec. 2007 it didn't work anymore,
though. For a simple scatterplot I  had to use:

twoway scatter exshpt exshare;

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

*DESCRIBING THE DISTRIBUTION OF A SINGLE VARIABLE;

histogram flunks if judge==0,  discrete  saving(flunks-l,replace) ;
*This is for the discrete variable flunks, if the judge variable-0. It
saves the
graph as flunks-l.gph. To convert the graph to TIFF, just right-click
on it in
STATA and save as TIFF. ;

graph winrate, box saving (winrate-box,replace) ;
 * for a box and whiskers plot;

graph winrate, histogram bin(20) saving (winrate-histogram,replace)  ;

kdensity winrate,  saving (winrate-kdensity,replace)  ;





  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

 *MISSING VALUES;

  * FOr missing values in entries, use  a period, .  Other thigns
cause trouble.

 Note that a missing value can be treated as being a very large number,
for variable generation purposes. Watch out for things like
replace var1 =4 if var2>6, because if var2 is missing, that command will
make var1=4;


* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;
  INSHEET;

 This is to get spreadsheet data into STATA:

* Save the spreadsheet as a tab file. Then say
INSHEET USING MYNAME.TXT;

OUTSHEET:
 This is to get spreadsheet data out of STATA.

Outsheet using myname.csv, comma, replace

That uses commas to separate the variables.


* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

  CONVERTING STATA 9 DATASET TO STATA 4.

 Stata has set things up so later versions are incompatible with
earlier ones. Stata 9 on Steel will read my co-authors' files, but not
my Stata 7 on my PC. Here's how to solve that:

Upload the file smith.dta to Stata 9.

Use smith.dta
outsheet using smith.txt

Download the smith.txt to Stata 7. On the PC:

set mem 50m   *for a big dataset
 insheet using smith.txt
save smith1.dta






* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;
* IF STATEMENTS AND STRING VARIABLES;

 replace appointed =1    if  state =="NJ" ;
replace appointed =1    if  state =="AK" ;

LOGICAL OPERATORS:
 "not equal to " is ~=
 Use & instead of "and" in logical statements.


* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;
* SUMMARY STATISTICS;

Do not use the SUMMARIZE or SUMMARIZE, DETAILS command. Instead, use
tabstat, like this:

tabstat  budget felclosed felconv, stats(min  p25  median  mean  p75
max)   f(%7.2f) columns(statistics)  ;

 INSPECT is a good command too. It gives you little histograms for
each variable.

  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

* HOW TO HAVE ROBUST STANDARD ERRORS IN TOBIT IN STATA;

 The documentation at

http://www.stata.com/support/faqs/stat/tobit.html

 is cryptic and seems to be missing some lines in the middle, so I am
writing up these instructions. I still don't understand what is going
on, but this at least makes the command work.

Suppose you have regressed winrate using tobit  on prosrate  budpros
term crimerate  pop. You used tobit because winrate is censored at 0
and 100-- you have the observations at the extremes, but the values
cannot be less than 0 or greater than 100.  You used this command:

tobit winrate prosrate  budpros term   pop, ll(0)  ul(100)
;

  Another way to get exactly the same results is to use INTREG, the
interval regression command. This is designed for when some of your
data is intervals, as when for example you know that somebody's
income is between 1 and 3 million, but not exactly where in between.
Here, we use it differently.


 First, define two  new variables  winrate1 and winrate2.   Each
observation is in an interval [winrate1, winrate2].  Winrate1 will be
winrate, or will be missing (representing negative infinity) if
winrate takes its  lower bound of 0.   Winrate2 will be winrate, or
will be missing (representing   infinity) if winrate takes its   upper
bound of 100.  Here are the Stata commands to generate those
variables:


gen winrate1=winrate;

replace winrate1  =. if  winrate  <= 0;

gen winrate2=winrate;

replace winrate2   =. if  winrate  >= 100;

 Suppose our data was like this: (12, 0, 50, 100, 23)

Then winrate 1 would be (12,. , 50, 100, 23)

Then winrate 2 would be (12, 0, 50, ,, 23)

 The intervals in the regression would be ([12,12], [-infinity, 0],
[50,50], [100, infinity], [23,23]).

Then use the INTREG command. Its first two variables, winrate1 and
winrate2, are the bounds that make up the intervals for each
observation of the dependent variable, and the rest are independent
variables.

intreg winrate  winrate2  prosrate  budpros term      pop;

The regression above  should give  exactly the same results as the
tobit command specified earlier. It doesn't quite, for me, but I put
that down to different maximization algorithms for the two commands.
Now, though, we can have robust standard errors as an option to
correct for heteroskedasticity:

intreg winrate  winrate2  prosrate  budpros term      pop, robust;

 This last command has the exact same coefficient estimates, but
different, and consistent, standard errors.


 What if you only left-censor, so winrate can't be less than 0, but it
can be greater than 100? Then generate the bounds like this:

gen winrate1=winrate;

replace winrate1  =. if  winrate  <= 0;

gen winrate2=winrate;

What if you only right-censor, so winrate can  be less than 0, but it
can't be greater than 100? Then generate the bounds like this:

gen winrate1=winrate;

gen winrate2=winrate;

replace winrate2  =. if  winrate  >= 100;



* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

*MARGINAL EFFECTS IN TOBIT;

See if dtobit, dlogit2, dprobit work. They report marginal effects
instead of
coefficients.  Use the command dprobit, at(xxxx) if you want it
calculated at
xxx- but xxx has to be a matrix,  and  I don't know how to get a
median matrix.
Also, dtobit has to be installed from the web, epsecially.


.    Otherwise, use the mfx command below.

mfx is the marginal effects command.

  Use a dot if you are only left-censored or right-censored.

predict(e(0, 100)) will give the marginal effect  on the expected
value of the dependent variable conditional on being uncensored,
E(y|a<y<b),

predict(ystar(0,100)) will give  the marginal effects on the
unconditional expected value of the dependent variable, E(y*), where
y* = max(0, min(y,100)). This is usually what you want.

predict(ystar(80,.)) will give  the marginal effects on the
unconditional expected value of the dependent variable, E(y*), where
y* = max(80,y).

eyex predict(ystar(0, 100)) will give the elasticity rather than the
simple marginal effect, because of the eyex option.



 EXAMPLES:

 tobit winrate prosrate  budpros term inoffice  crimerate  pop,
   ll(80) ;

mfx compute, predict(ys(80,.))  at(median);

mfx compute, predict(e(80,.));



tobit winrate prosrate  budpros term inoffice  crimerate  pop,
  ll(0) ul(100) ;

mfx compute, predict(ys(0, 100));

 mfx compute, eyex nose  predict(ys(0, 100));

This last line includes an extra option, "nose", which means "no
standard errors". That is good if you have a lot of dummy variables
for fixed effects, because it takes the computer a long time to
compute so many  standard errors.

Remember that in some contexts you do NOT want marginal effects-- with
censored data where the true value is simply unobserved.

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

COMPUTING ELASTICITIES

mfx compute, eyex predict(ystar(0, 100)) at(median);

 will give the elasticity rather than the simple marginal effect of
tobit, because of the eyex option. This eyex option can be used after
OLS too-- I think, though I'm just guessing, as:

 mfx compute, eyex;

Here are how to manually get elasticities.

 The elasticity  is (%change in Y)/(%change in X) =  (dy/dx)*(x/y)


(1) If y = beta*x  then the elasticity is beta*(x/y).


(2)  If  y  = beta* log(x)  then the elasticity is
 (beta/x)*(x/y) = beta/y

(3)  If log(y) = beta* log(x)  then the elasticity is
(beta*y/x)*(x/y) = beta,
  which is a constant elasticity.

(4) If log(y)  = beta*x  then the elasticity is
 (beta*y)(x/y) = beta*x.



* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;


    sort age;
 generate agerank = autocode(age, _N, age[1],
age[_N]);
 tabulate agerank;

program define   myfunction1     /*arguments are
xvar,  a,b   */
 generate yvar =  xvar^a + b
end

program define   myfunction2     /*arguments are
a,b   */
 generate  xvar = 10^a +b;
end

 Make this a do file, and then in your STATA program, type run
myfunction1;


See p. 142 of volume 1 for an easy way to customize statistics.


program define mean  /* one argument, the variable */
   1. quietly  summarize  `1'
  2. di  "The mean equals "  _result(3)
 3. end;


 It is easy to graph predicted and actual values. Use hte WIndows
EDITOR
and SORT features to sort by size of predicted value. Then graph as

 replace obs =_n;
   graph pred1 var1 obs;

 Also, check the data with

 graph var1, box;

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

*SIMPLE REGRESSION;
 regress murder metro black;

 *Weighted regressions.

regress murder metro black [weight=pop];

* White-heteroskedasticity-corrected standard errors.

regress murder metro black, robust;

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

*COMBINING COMMANDS;
regress murder metro black [weight=pop], robust;

summarize murder, detail;

 *OR, BETTER;

 tabstat  murder divorce, stats(mean    sd    median  min max);

tabstat  murder divorce [weight=pop], stats(mean    sd median  min
max);

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

*Putting in squared terms, logs.
* I must use GENERATE to generate the new variables.
* Note that if a squared term comes in positive, then the function is
increasingly steep.

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

sort year;
by year: summarize crime;
*This sorts the data by year and then summarizes the crime variable;

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;

*Dropping outliers.;
 drop if state=="DistrictofColumbia";

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;


ds;
  *This gives a list of the variables

describe, fullnames;

 * : This gives info on thevariables using their full
names, not abbreviations if the are longer than 8 characters.

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;


   sort age;
 generate agerank = autocode(age, _N, age[1],
age[_N]);
 tabulate agerank;


* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;


program define   myfunction1     /*arguments are
xvar,  a,b   */
 generate yvar =  xvar^a + b
end

program define   myfunction2     /*arguments are
a,b   */
 generate  xvar = 10^a +b
end

 Make this a do file, and then in your STATA program, type run
myfunction1;

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;



See p. 142 of volume 1 for an easy way to customize statistics.

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;



program define mean  /* one argument, the variable */
   1. quietly  summarize  `1'
  2. di  "The mean equals "  _result(3)
 3. end;

* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;


 It is easy to graph predicted and actual values. Use hte WIndows
EDITOR
and SORT features to sort by size of predicted value. Then graph as
 replace obs =_n
   graph pred1 var1 obs


* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;
 ROBUST REGRESSION

  rreg is the command. It uses a fancy procedure to downweight
observations.
See: http://www.ats.ucla.edu/stat/stata/dae/rreg.htm


* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;


* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~--;