---------------------------------------------------------------------------------
log: C:\AAA Miker Files\newer web pages\soc_388_notes\soc_388_2007\cla
> ss_18_log.log
log type: text
opened on: 29 Nov 2007, 10:56:32
. table color labor, contents(sum uwcount sum wncount mean weight) by(sex)
----------------------------------------------
sex and | labor
color | unemployed part-time other
----------+-----------------------------------
male |
white | 3511 4227 31467
| 3530 4183 31131
| 1431 1408 1408
|
black | 604 356 2245
| 815 462 2783
| 1921 1849 1764
|
other | 165 157 924
| 119 124 797
| 1029 1124 1228
----------+-----------------------------------
female |
white | 2281 7833 18945
| 2234 7559 18704
| 1394 1373 1405
|
black | 545 563 2132
| 653 644 2498
| 1705 1627 1668
|
other | 89 216 725
| 64 162 574
| 1029 1070 1127
----------------------------------------------
. table color labor, contents(sum uwcount sum wncount mean weight) by(sex)
----------------------------------------------
sex and | labor
color | unemployed part-time other
----------+-----------------------------------
male |
white | 3511 4227 31467
| 3530 4183 31131
| 1431 1408 1408
|
black | 604 356 2245
| 815 462 2783
| 1921 1849 1764
|
other | 165 157 924
| 119 124 797
| 1029 1124 1228
----------+-----------------------------------
female |
white | 2281 7833 18945
| 2234 7559 18704
| 1394 1373 1405
|
black | 545 563 2132
| 653 644 2498
| 1705 1627 1668
|
other | 89 216 725
| 64 162 574
| 1029 1070 1127
----------------------------------------------
. *I am going to run some loglinear models to immitate what Clogg and Eliason do in their Table 6
. desmat: poisson uwcount labor*sex=dev(3)*dev(2) color*labor=dev(3)*dev(3) color*sex=dev(3)*dev(2)
--------------------------------------------------------------------------------
Poisson regression
--------------------------------------------------------------------------------
Dependent variable uwcount
Optimization: ml
Number of observations: 18
Initial log likelihood: -81627.074
Log likelihood: -123.390
LR chi square: 163007.367
Model degrees of freedom: 13
Pseudo R-squared: 0.998
Prob: 0.000
--------------------------------------------------------------------------------
nr Effect Coeff s.e.
--------------------------------------------------------------------------------
uwcount
labor
1 unemployed -0.677** 0.018
2 part-time -0.433** 0.017
sex
3 male -0.018* 0.009
labor.sex
4 unemployed.male 0.161** 0.009
5 part-time.male -0.347** 0.007
color
6 white 1.852** 0.011
7 black -0.364** 0.014
color.labor
8 white.unemployed -0.282** 0.018
9 white.part-time 0.193** 0.017
10 black.unemployed 0.340** 0.022
11 black.part-time -0.229** 0.022
color.sex
12 white.male 0.080** 0.009
13 black.male -0.097** 0.011
14 _cons 7.053** 0.011
--------------------------------------------------------------------------------
* p < .05
** p < .01
. poisgof
Goodness-of-fit chi2 = 86.53056
Prob > chi2(4) = 0.0000
. poisgof, pearson
Goodness-of-fit chi2 = 89.79915
Prob > chi2(4) = 0.0000
. *This is clogg and Eliason's model 1, ignoring the weights.
. desrep, z
option z not allowed
r(198);
. desrep, zscore
option zscore not allowed
r(198);
. desrep, zval
--------------------------------------------------------------------------------
Poisson regression
--------------------------------------------------------------------------------
Dependent variable uwcount
Optimization: ml
Number of observations: 18
Initial log likelihood: -81627.074
Log likelihood: -123.390
LR chi square: 163007.367
Model degrees of freedom: 13
Pseudo R-squared: 0.998
Prob: 0.000
--------------------------------------------------------------------------------
nr Effect Coeff s.e. z
--------------------------------------------------------------------------------
uwcount
labor
1 unemployed -0.677** 0.018 -38.575
2 part-time -0.433** 0.017 -26.205
sex
3 male -0.018* 0.009 -2.004
labor.sex
4 unemployed.male 0.161** 0.009 18.482
5 part-time.male -0.347** 0.007 -46.843
color
6 white 1.852** 0.011 162.348
7 black -0.364** 0.014 -25.659
color.labor
8 white.unemployed -0.282** 0.018 -15.386
9 white.part-time 0.193** 0.017 11.300
10 black.unemployed 0.340** 0.022 15.425
11 black.part-time -0.229** 0.022 -10.465
color.sex
12 white.male 0.080** 0.009 9.162
13 black.male -0.097** 0.011 -8.679
14 _cons 7.053** 0.011 640.281
--------------------------------------------------------------------------------
* p < .05
** p < .01
. *Now the crazy one.
. gen crazy_wt_count= uwcount*weight
. desmat: poisson crazy_wt_count labor*sex=dev(3)*dev(2) color*labor=dev(3)*dev(3) color*sex=dev(3)*dev(2)
--------------------------------------------------------------------------------
Poisson regression
--------------------------------------------------------------------------------
Dependent variable crazy_wt_count
Optimization: ml
Number of observations: 18
Initial log likelihood: -1.137e+08
Log likelihood: -71806.645
LR chi square: 2.273e+08
Model degrees of freedom: 13
Pseudo R-squared: 0.999
Prob: 0.000
--------------------------------------------------------------------------------
nr Effect Coeff s.e.
--------------------------------------------------------------------------------
crazy_wt_count
labor
1 unemployed -0.689** 0.001
2 part-time -0.435** 0.000
sex
3 male 0.013** 0.000
labor.sex
4 unemployed.male 0.165** 0.000
5 part-time.male -0.340** 0.000
color
6 white 1.858** 0.000
7 black -0.133** 0.000
color.labor
8 white.unemployed -0.263** 0.001
9 white.part-time 0.186** 0.000
10 black.unemployed 0.383** 0.001
11 black.part-time -0.231** 0.001
color.sex
12 white.male 0.059** 0.000
13 black.male -0.083** 0.000
14 _cons 14.294** 0.000
--------------------------------------------------------------------------------
* p < .05
** p < .01
. poisgof
Goodness-of-fit chi2 = 143357.3
Prob > chi2(4) = 0.0000
. poisgof, pearson
Goodness-of-fit chi2 = 148750.5
Prob > chi2(4) = 0.0000
. *OK, crazy model has correct coefficients but unbelievably small SE and wildly large goodness of fit, because we have basically multiplied all counts by 1500
. *let's put that model behind us.
. *The next reasonable approach, is to normalize the weights. Actually, the weighted counts provided by Clogg and Eliason already do this.
. *C+E's weighted counts are already normalized...
. desmat: poisson wncount labor*sex=dev(3)*dev(2) color*labor=dev(3)*dev(3) color*sex=dev(3)*dev(2)
--------------------------------------------------------------------------------
Poisson regression
--------------------------------------------------------------------------------
Dependent variable wncount
Optimization: ml
Number of observations: 18
Initial log likelihood: -79982.751
Log likelihood: -130.057
LR chi square: 159705.387
Model degrees of freedom: 13
Pseudo R-squared: 0.998
Prob: 0.000
--------------------------------------------------------------------------------
nr Effect Coeff s.e.
--------------------------------------------------------------------------------
wncount
labor
1 unemployed -0.690** 0.020
2 part-time -0.435** 0.018
sex
3 male 0.013 0.010
labor.sex
4 unemployed.male 0.165** 0.009
5 part-time.male -0.340** 0.007
color
6 white 1.858** 0.012
7 black -0.133** 0.015
color.labor
8 white.unemployed -0.263** 0.020
9 white.part-time 0.186** 0.018
10 black.unemployed 0.384** 0.023
11 black.part-time -0.232** 0.022
color.sex
12 white.male 0.059** 0.009
13 black.male -0.083** 0.011
14 _cons 7.033** 0.012
--------------------------------------------------------------------------------
* p < .05
** p < .01
. poisgof
Goodness-of-fit chi2 = 100.2395
Prob > chi2(4) = 0.0000
. poisgof, pearson
Goodness-of-fit chi2 = 104.0537
Prob > chi2(4) = 0.0000
. *This is clogg and eliason's model 2, using the normalized weights.
. *changing the dataset by a factor leaves the coefficients the same, but makes the SE and fit statistics reasonable again.
. *The way to do C+E's recommended weighting scheme in Stata, is to use the exposure function, which is also relevant to log-rate models, that is it is supposed to correct for the different exposure of each cell.
. gen inv_weight=1/weight
. desmat: poisson uwcount labor*sex=dev(3)*dev(2) color*labor=dev(3)*dev(3) color*sex=dev(3)*dev(2), exposure( inv_weight)
--------------------------------------------------------------------------------
Poisson regression
--------------------------------------------------------------------------------
Dependent variable uwcount
Optimization: ml
Number of observations: 18
Initial log likelihood: -84619.027
Log likelihood: -124.919
LR chi square: 168988.216
Model degrees of freedom: 13
Pseudo R-squared: 0.999
Prob: 0.000
--------------------------------------------------------------------------------
nr Effect Coeff s.e.
--------------------------------------------------------------------------------
uwcount
labor
1 unemployed -0.688** 0.018
2 part-time -0.440** 0.017
sex
3 male 0.013 0.009
labor.sex
4 unemployed.male 0.166** 0.009
5 part-time.male -0.343** 0.007
color
6 white 1.860** 0.011
7 black -0.136** 0.014
color.labor
8 white.unemployed -0.265** 0.018
9 white.part-time 0.191** 0.017
10 black.unemployed 0.386** 0.022
11 black.part-time -0.238** 0.022
color.sex
12 white.male 0.058** 0.009
13 black.male -0.085** 0.011
14 _cons 14.292** 0.011
ln(inv_weight) (offset)
--------------------------------------------------------------------------------
* p < .05
** p < .01
. poisgof
Goodness-of-fit chi2 = 89.58815
Prob > chi2(4) = 0.0000
. poisgof, pearson
Goodness-of-fit chi2 = 93.54537
Prob > chi2(4) = 0.0000
. *This is C+E's model 3, their correct approved procedure, using the unweighted counts and the inverse weights as exposure.
. *note that the coefficients are the same as all the weighted models.
. *This model, Model 3, also has the same standard errors as M1, which ignored the weights.
. *Take away message: the proper coefficients take the weights into account. The proper SE are unaffected by the weights.
. log close
log: C:\AAA Miker Files\newer web pages\soc_388_notes\soc_388_2007\class
> _18_log.log
log type: text
closed on: 29 Nov 2007, 12:09:07
--------------------------------------------------------------------------------