4 Results of Loglinear analysis
In order to test for differences in mortality, loglinear models were formulated (using the
software package GLIM 4.0 (Francis et al., 1993)). Loglinear analysis of demographic
processes is a way to test hypotheses on connections between categorical variables in
demographic processes. In the case of mortality numbers broken down by age (A),
period (P), cohort (C), economic activity (SEC), economic main structure (EMS), and
number of employed (EMPL) it is possible to test several associations.
As explained in section 2, this type of analysis yields a test criterium, the likelihood
ratio, or deviation. Though this quantity does not follow a known distribution, and a
formal statitical test is therefore impossible, it does give an indication of the relative
importance of each of the variables in explaining the variation in mortality numbers. On
the basis of this quantity one may decide whether mortality is for example sector-
specific or not. Results of these analyses are shown in table 2.
Table 2 Results of Ioglinear analysis
|
Model |
Scaled |
Residual |
% |
|
∏ |
17,112 |
2,855 |
0.00 |
|
2 EMS+EMPL+SEC |
14,139 |
2,849 |
17.37 |
|
3 EMS+EMPL+SEC+A |
13,849 |
2,835 |
19.07 |
|
4 EMS+EMPL+SEC+C |
9,150 |
2,835 |
46.53 |
|
5 EMS+EMPL+SEC+P |
9,230 |
2,836 |
46.06 |
|
6 EMS+EMPL+SEC+A+C |
8,135 |
2,821 |
52.46 |
|
7 EMS+EMPL+SEC+A+P |
7,774 |
2,822 |
54.57 |
|
8 EMS+EMPL+SEC+C+P |
7,501 |
2,822 |
56.17 |
|
9 EMS+EMPL+SEC+A+P+C |
7,234 |
2,809 |
57.73 |
|
10 EMS+EMPL+SEC+A+P+C+A*EMS |
7,175 |
2,795 |
58.07 |
|
11 EMS+EMPL+SEC+A+P+C+A*EMPL |
6,721 |
2,781 |
60.72 |
|
12 EMS+EMPL+SEC+A+P+C+A*SEC |
7,208 |
2,767 |
57.88 |
|
13 EMS+EMPL+SEC+A+P+C+P*EMS |
6,555 |
2,796 |
61.69 |
|
14 EMS+EMPL+SEC+A+P+C+P*EMPL |
6,384 |
2,783 |
62.69 |
|
15 EMS+EMPL+SEC+A+P+C+P*SEC |
7,176 |
2,770 |
58.07 |
|
16 EMS+EMPL+SEC+A+P+C+C*EMS |
6,931 |
2,795 |
59.50 |
|
17 EMS+EMPL+SEC+A+P+C+C*EMPL |
5,877 |
2,781 |
65.66 |
|
18 EMS+EMPL+SEC+A+P+C+C*SEC |
7,188 |
2,767 |
57.99 |
Within loglinear analysis it is also possible to test for higher order interactions (for
example A*P*C, but also interactions between each of the time dimesnions on the one
hand and a factor on the other). We did test for such interactions, and results are shown
in table 2, but the results were not satisfactory. Gains in scaled deviance were small,
standard errors became too large and parameter values became uninterpretable. As an
example figure 7 shows the parameter estimates for age by number of employed
according to model 11. Values larger than zero indicate higher than average mortality
chances and values smaller than zero lower than average chances of dying. Though the
hypothesis that the more employees a firm has, the lower the chances of dying are, is
confirmed by this model, mortality hardly decreases with age for firms with 1 or more
employees. The decreasing age effect seems to be true (in this model) for zero-
employees-firms only. This is clearly not what we expected.
14