- 408 - 
where x is a vector of covariates, T, is the random variab- 
le "waiting time" until a change to state k, f(t) is the 
probability density distribution of waiting times referring 
to state j and G.(t) is the probability of survival (survi- 
val function) in state j. 
As mentioned in the introduction, parameters are estimated 
on data of the cumulative General Social Survey 1980 to 1984 
(Allbus 80-84). The analysis of "unmarried episodes" utili- 
zes only the 1982 and 1984 sample (about 6000 households) 
because there was the danger of a systematic bias by a dif- 
ferent operationalization of "age" in the 1980 survey. Esti¬ 
mation of effects on divorce likelihood refers to the subset 
of first marriages from all three surveys. The sample size 
is further reduced by missing data, particularly if income 
is included in the equations, or by the necessity to draw a 
random sample in case of Cox-regression because of limited 
computer space. 
4. Explanations of Non-monotonic Risk Functions 
Life-table-estimates of the hazardrate for first marriage 
follow a non-monotonic, bell-shaped pattern (Figure 2). This 
"marriage bell" corresponds with a typical S-shaped cumula- 
tive distribution of age at marriage. Survival-functions for 
women and men are of the same shape, delayed by two to three 
years with a crossover point at age 41 (Figure 3). The pro¬ 
portion of never married women at age 50 (6.9 8) is a little 
higher than the proportion of never married men (5.3 8). 
S-shaped cumulative distributions and non-monotonic bell- 
shaped hazardrate functions are reported by many authors 
studying entry into marriage (e.g. Sørensen and Sørensen, 
1985; Espenshade, 1982). Hence, the "marriage bell" seems to 
be a cross-cultural invariant law. How can this regularity 
be explained?