Full text: Wehner, Sigrid: Exploring trends and patterns of nonresponse

4. ESTIMATING BIAS IN MULTIVARIATE RELATIONSHIPS 
72 
not all coefficients are simultaneously zero. The model explains about 38% of the variance 
(r**2-0,376). All coefficients are significant at a 5%-significance level. They appear in the 
direction that is theoretically expected, according to the analysis in the exploration chapter, 
and can be interpreted in terms of their financial contribution to the household income 1996 
for women of cohort 1960. 
Having income as the dependent variable, it is easy to interpret the coefficients. We 
see that the income is rising by the number of persons (one more person contributes 292,04 
DM) and the most positive effect on income is given for married women in comparison to 
unmarried (1267,27 DM). Being unemployed decreases the household income. We also see 
that target persons from the capi field have lower household income compared to persons 
from the telephone field. The effect is nearly half of the unemployment effect. We also see 
that longer education contributes positively to the household income. Following the 
constructed definition of this variable, we may interpret one more year of education in school 
or later vocational training as being worth 255,87 DM in income, keeping all other 
coefficients of the model comparably constant. 
Regression Estimates for the Model in the Biased Sample 
The column with the regression estimates for the biased sample shows that the significance 
structure and the overall fit is comparable. Also we do not find severe changes in the sense 
that positive coefficients turned to negative (or the reverse). We see that some estimates differ 
greatly, however, e.g. the coefficient for humkap (368,48, although only 255,87 in the original 
sample). Of course this is the bias impact which was intentionally implanted in this sample. 
(The high education groups were excluded.) As a consequence, the effect of years of 
education on the household income is overestimated in the remaining cases. 
Regression Estimates for the Heckman Sample Selection Model 
We see in the general structure that the variables for number of persons and capi interview are 
no longer significant. There is an overall increase in the standard errors. The significant Wald- 
test also indicates that the model differs from the null hypothesis assuming all coefficients to 
be zero. The value for rho (p-0,449) shows the relatively high correlation between the error 
terms of the selection and the regression equation. Rho sigma (po-468,471) is the value of 
the coefficient of the additional correction regressor (the Inverse Mill's Ratio) and is an 
overall estimate of the magnitude of the selection bias (see: Brehm (1993:122); STATA 6.0 
manual, pp.18-20). po is rather high in our case.
	        
Waiting...

Note to user

Dear user,

In response to current developments in the web technology used by the Goobi viewer, the software no longer supports your browser.

Please use one of the following browsers to display this page correctly.

Thank you.

powered by Goobi viewer