*
* CHAPTER 11
* Page 387
*
open data ch11_hmda_sw.xls
data(format=xls,org=columns) 1 2380 seq s3 s4 s5 s6 s7 s9 s11 s13 s14 s15 s16 s17 s18 s20 married s24a s25a s26a s27a $
  s30a s30c s31a s31c s32 s33 s34 s35 s39 s40 s41 s42 s43 s44 s45 s46 s47 s48 s50 s51 s52 s53 s55 s56 s57 $
   netw	uria rtdum bd mi old vr school chval dnotown dprop
*
* Compute transformations of the original data. Note that s7==3 evaluates as 1 if
* s7 is 3 and 0 otherwise. You need the == for the test for equality, not just
* one =.
*
set denied  = s7==3
set piratio = s46/100
set black   = s13==3
*
* LPM
*
* Equation 11.1
linreg(robust) denied
# constant piratio
*
* Equation 11.3
linreg(robust) denied
# constant piratio black
*
* Probit
*
* Equation 11.7
ddv denied
# constant piratio
*
* Predicted probabilities. The PRJ instruction computes fitted values based upon
* the most recent estimation. The XVECT option feeds in a particular set of
* values for the "X" in the linear part Xb. PRJ produces no printed output, but
* %prjcdf is the predicted probability for the normal distribution. Note that the
* values are slightly different from those in the text, as these are calculated
* at full precision, while those in the text are rounded.
*
prj(xvect=||1.0,.3||)
compute prob_3=%prjcdf
prj(xvect=||1.0,.4||)
compute prob_4=%prjcdf
prj(xvect=||1.0,.5||)
compute prob_5=%prjcdf
*
disp "Probability at PI=.3" prob_3
disp "Probability at PI=.4" prob_4
disp "Probability at PI=.5" prob_5
*
* Equation 11.8
ddv denied
# constant piratio black
*
prj(xvect=||1.0,.3,0.0||)
compute prob_w=%prjcdf
prj(xvect=||1.0,.3,1.0||)
compute prob_b=%prjcdf
disp "Probability for Non-Black at .3" prob_w
disp "Probability for Black at .3" prob_b
*
* Logit models
*
* Equation 11.10
ddv(dist=logit) denied
# constant piratio black
*
* Computing probabilities off a logit is the same as for the probit, except that
* you need to add in the DIST=LOGIT option.
*
prj(xvect=||1.0,.3,0.0||,dist=logit)
compute prob_w=%prjcdf
prj(xvect=||1.0,.3,1.0||,dist=logit)
compute prob_b=%prjcdf
disp "Probability for Non-Black at .3" prob_w
disp "Probability for Black at .3" prob_b
*
* Compute various transformations off the original data. Note that s53==1
* evaluates as 1 if s53 is 1 and 0 otherwise. Remember to use == for the test for
* equality, not just one =.
*
set hse_inc  = s45/100
set loan_val = s6/s50
set ccred    = s43
set mcred    = s42
set pubrec   = (s44>0)
set denpmi   = (s53==1)
set selfemp  = (s27a==1)
set single   = 1-married
set hischl   = (school>=12)
set probunmp = uria
set condo    = (s51==1)
set ltv_med  = loan_val>=0.80.and.loan_val<=0.95
set ltv_high = loan_val>0.95
set blk_pi   = black*piratio
set blk_hse  = black*hse_inc
set ccred3   = ccred==3
set ccred4   = ccred==4
set ccred5   = ccred==5
set ccred6   = ccred==6
set mcred3   = mcred==3
set mcred4   = mcred==4
*
* Expanded model estimated with all three techniques
*
* Table 11.2
* Column 1
*
linreg(robust) denied
# black piratio hse_inc ltv_med ltv_high ccred mcred pubrec denpmi selfemp constant
disp "Difference in Predicted Probability of Denial, Black vs White" %beta(1)
*
* Column 2
*
ddv(robust,dist=logit) denied
# black piratio hse_inc ltv_med ltv_high ccred mcred pubrec denpmi selfemp constant
*
* Calculation of predicted probability of denial for black vs white. This is
* computed (in all cases) with all other explanatory variables set to their means.
* The PRJ instruction is used extensively. Its first use is to get the mean vector
* itself. PRJ(ATMEAN) evaluates xb at the mean of the explanatory variables; we're
* doing this only because, as a side effect, it produces the vector of means.
* We'll then use PRJ to compute the required probabilities after patching over
* the first element of a copy of the mean vector with 1 (for black) or 0 (for
* white).
*
prj(atmean)
compute testv=%meanv
compute testv(1)=1
prj(xvect=testv,dist=logit)
compute prob_b=%prjcdf
compute testv(1)=0
prj(xvect=testv,dist=logit)
compute prob_w=%prjcdf
disp "Difference in Predicted Probability of Denial, Black vs White" prob_b-prob_w
*
* Column 3
*
ddv(robust) denied
# black piratio hse_inc ltv_med ltv_high ccred mcred pubrec denpmi selfemp constant
prj(atmean)
compute testv=%meanv
compute testv(1)=1
prj(xvect=testv)
compute prob_b=%prjcdf
compute testv(1)=0
prj(xvect=testv)
compute prob_w=%prjcdf
disp "Difference in Predicted Probability of Denial, Black vs White" prob_b-prob_w
*
* Probit only for the remaining models. Each is followed by the tests given in
* table 11.2 on page 403.
*
* Column 4
*
ddv(robust) denied
# black piratio hse_inc ltv_med ltv_high ccred mcred pubrec denpmi selfemp single hischl uria constant
exclude
# single hischl uria
prj(atmean)
compute testv=%meanv
compute testv(1)=1
prj(xvect=testv)
compute prob_b=%prjcdf
compute testv(1)=0
prj(xvect=testv)
compute prob_w=%prjcdf
disp "Difference in Predicted Probability of Denial, Black vs White" prob_b-prob_w
*
* Column 5
*
ddv(robust) denied
# black piratio hse_inc ltv_med ltv_high ccred mcred pubrec denpmi selfemp $
   single hischl uria condo mcred3 mcred4 ccred3 ccred4 ccred5 ccred6 constant
exclude
# single hischl uria
exclude
# mcred3 mcred4 ccred3 ccred4 ccred5 ccred6
prj(atmean)
compute testv=%meanv
compute testv(1)=1
prj(xvect=testv)
compute prob_b=%prjcdf
compute testv(1)=0
prj(xvect=testv)
compute prob_w=%prjcdf
disp "Difference in Predicted Probability of Denial, Black vs White" prob_b-prob_w
*
* Column 6
*
ddv(robust) denied
# black piratio hse_inc ltv_med ltv_high ccred mcred pubrec denpmi selfemp $
  single hischl uria blk_pi blk_hse constant
exclude
# single hischl uria
exclude
# black blk_pi blk_hse
exclude
# blk_pi blk_hse
*
* The calculation of the black vs white probabilities here are a bit more
* complicated because of the interaction terms. We're going to want the test
* values for the blk_pi and blk_hse to be zero for white, but the sample average
* for piratio and hse_inc for black. blk_pi is the 14 slot and blk_hse is 15. We
* can pull the sample averages from the 2 and 3 slots respectively
*
prj(atmean)
compute testv=%meanv
compute testv(1)=1,testv(14)=testv(2),testv(15)=testv(3)
prj(xvect=testv)
compute prob_b=%prjcdf
compute testv(1)=testv(14)=testv(15)=0
prj(xvect=testv)
compute prob_w=%prjcdf
disp "Difference in Predicted Probability of Denial, Black vs White" prob_b-prob_w