289 - 
set of co-variates characterizing individual i. A particu- 
lar attractive specification of the function for p, (t) is: 
p. (t) = 1 - exp - exp(A, (t) + BX,)) 
(2) 
because it will provide estimates that approximate those 
that would have been obtained had the proportional hazards 
or the Cox model been estimated for the underlying 
continuous time model. In particular, these estimates will 
not be dependent on the length of the time interval. This 
specification can be estimated using the complementary 
log-log "link" specification in GLIM. 
If now A,(t) is further specified as: 
A(t) - a + alogt, 
(3) 
we obtain an approximation to the Weibull model. 
Interactions may here be tested using interaction terms 
involving logt and the co-variates. 
The discrete time approximation demands that the data with 
the unit of analysis being spells be converted to 
observations on each time unit for each individual. This 
produces an enormous amount of information -- here about 
10,000 observations. Since the event is quite rare and it 
was necessary to reduce the sample size considerably because 
of hardware limitations, the time unit was redefined as a 
two week period and a random 308 sample from these observa- 
tions were selected. This gave a sample of 2452 observa¬ 
tions for the present analysis.