**   run the commands exhibited in:
**      resample:  A command for randomly resampling a data set
**   but using StataQuest, rather than Stata.

**   NOTE:  In Example 2, StataQuest creates a bootstrap distribution that
**   differs just slightly from that created in Stata Version 5.0 (even
**   with the command -version 4.0- in effect).  Apparently, the command
**   -sqswreg- in StataQuest is not quite identical to the command -sw reg-
**   in Stata 5.0.  The results shown in resample.tex are from Stata 5.0,
**   not from StataQuest.

**  John R. Gleason
**  Syracuse University
**  430 Huntington Hall
**  Syracuse, NY  13244-2340
**  73241.717@compuserve.com
**  10Mar97


* Example 1:
* load the data set law_sch.dta:
more

use law_sch, replace
describe
* correlate lsat and gpa:
more

corr lsat gpa
* generate a random sample with replacement:
more

resample lsat gpa
more
describe
* correlate lsat_ and gpa_:
more

corr lsat_ gpa_
* now, a complete bootstrap simulation.  First, some preliminaries:
more

set seed 970211
set obs 600
gen r_boot = .
* then, the slow part:
more

loop, c(resample lsat gpa in 1/15; corr lsat_ gpa_ in 1/15; replace r_boot = _result\[4\] in I_) i(600)
* here's the bootstrap SE of r:
more

summ r_boot
more


* Example 2:
* load the data set mammal42.dta:
more

use mammal42, replace
describe
*  consider a stepwise regression of y = pdox_sl on 8 carriers (X's):
more

sqswreg pdox_sl lbod lbr lgest swav life is_p expo ind, forward pr(.4) pe(.2)
*  set up for a bootstrap test;  first, the commands to be repeated:
more

global LOOP_CMD "resample pdox_sl in 1/42; sqswreg pdox_sl_ lbod lbr lgest swav life is_p expo ind in 1/42, forward pr(.4) pe(.2); replace r2 = _result(8) in I_; reg pdox_sl_ lbod lbr lgest swav life is_p expo ind in 1/42; replace R2 = _result(8) in I_"
*  and then some more preliminaries:
more

gen r2 = .
gen R2 = .
set seed 970222
set obs 100
*  Finally, the actual simulation:
more

loop, i(100)
*  Here's the adjusted R^2 following stepwise selection:
more

summ r2
*  Stepwise selection eliminates all X's about 1/3 of the time:
more

count if r2 == 0
*  In the other 2/3 of the cases, it's generally larger than the adjusted R^2
*  based on all 8 X's:
more
reg r2 R2 if r2
more


*  Here is Figure 1:

set textsize 150
* left panel of Figure 1:
more

gr R2, bin(50) xlab ylab freq bor gap(4) xline(.64)   /*
      */ b2("Adjusted R-squared without variable selection")   /*
      */       /*    saving(fig1a, replace)  */

* right panel of Figure 1:
more

gr r2, bin(50) xlab(-.2,0,.2,.4,.6) ylab freq bor gap(4) xline(.64)  /*
      */ b2("Adjusted R-squared after stepwise selection")  /*
      */       /*    saving(fig1b, replace)  */
* and, the 2 panels combined:
more

gr using fig1a fig1b, margin(7)     /*   saving(fig1, replace)   */