4. ESTIMATING BIAS IN MULTIVARIATE RELATIONSHIPS 
TABLE 17: Heckman Sample Selection Model Using Panel and Nonresponse Study 
- Women of Cohort 1960 - 
Onty 
Panel with 
Panel Persons t 
Panel Persons 
Heckman Correction 
Nonrespondents 
 
 
 
 
 
coeme 
 
(coefficient 
independent 
coeffident 
dependen 
ent 
 
 
 
 
 
 
 
 
a- 
 
 
 
ad a 
varlables 
ad. error 
varjadle 
tad. eror) Pa 
 
 
P. 
nphh96 
292,04 
0,003 
0015 
274,68 
0,035 
277,43 
131.38 
98.67 
112.03 
hhinc96 
marrie 96 
0,000 
0,000 
1267,27 
0,000 
1302,08 
1300,91 
292.70 
276,12 
252.02 
915,77 
-845,30 
alo96 
0,001 
0,001 
917,24 
0,001 
265.68 
269.47 
252.74 
0,040 
-400,75 
—400,41 
0,051 
capi 
392,25 
0,028 
204.97 
193.63 
177.16 
254,18 
0,000 
0,000 
0,000 
255,87 
254,39 
humkap 
56.46 
51.52 
45.91 
-1077,99 
-1194,14 
-1058,97 
0,154 
0,080 
0,212 
constant 
739,97 
678.56 
864.40 
183 
valid observations 
213 
211 
panel observations 
183 
nonrespondents 
30 
F 
21,78 
24,67 
Prob 2 F 
0,000 
0,000 
2 
0,3809 
0,376 
Wald chi2 
96,63 
Prob 2 chi2 
0,000 
0,027 
9 
32,183 
 
Bold coefficients mark 5%-significance-level. 
hhinc96 = 
Variables are: 
household income 1996 
nphh96 - 
number of persons in household (nphh96), 
marrie96 - 
being married 1996 (1 if married; O otherwise), 
alo96 - 
being unemployed 1996 (1 if yes; 0 otherwise), 
capi = 
CAPI interview 1996 (1 if yes; 0 otherwise), 
humkap - 
cumulated education in school and vocational training in years 
In the left column we find the estimates for the survey data (only panel persons), in the middle 
the Heckman correction and in the right column the coefficients for the complete sample 
containing panel persons and nonrespondents. The last one was also the validation sample OS 
in the bias example. The same results are listed again. 
In general, there are no great differences between panel sample and 
panel-nonresponse. The Heckman corrections are not strong, although each"" of the 
coefficients, including the constant, goes in the right direction of the validation sample. Also 
the level of significance of the estimates does not change substantially (capi being near to 
5%). We recognise with p-0,027 that the selection equation is not very closely related to the 
Considering the estimate for capi as equal. 
76