324 Economic Cycles
converges to Cox’s (1972) continuous-time proportional hazards model. Cox’s
model can be written as:
logh(t)=a(t)+βx, (7.16)whereh(t)is the continuous-time hazard rate, similar toP(t), andxrepresents a
set of covariates that do not vary over time. Like the discrete-time logit model,a(t)
can be any function of time. The termproportional hazardcomes from the fact that,
for anytand any two individualsiandj:
hi(t)/hj(t)=exp(βxi)/exp(βxj)=exp(β(xi−xj)). (7.17)This function does not vary with time because the autonomous time-varying term,
a(t), cancels out. Cox’s model, however, is not just limited to proportional hazards,
since the model is no longer proportional if some of thexvalues vary with time.
Some computer packages allow for time-varying covariates while others do not.
Fortunately, this is not a concern to us since it is always possible to allow for
time-varying covariates in logistic regression.
Cox’s proportional hazards model is semiparametric because the autonomous
time-varying term,a(t), does not have to be specified. Although the estimators of
βare asymptotically unbiased and normally distributed, they are notfullyefficient
because the exact functional form ofa(t)is not specified. However, Efron (1977)
shows that the loss of efficiency is typically so small that it is not of practical
concern. The importance of Cox’s model for duration analysis is well-summarized
by Allison (1984, p. 35):
It is difficult to exaggerate the impact of Cox’s work on the practical analysis
of event history data. In recent years, his 1972 paper has been cited well over
100 times a year in the world scientific literature. In the judgment of many, it
is unequivocally the best all-around method for estimating regression models
with continuous-time data.In practice, time is always measured in discrete intervals, although the intervals
may be irregular for individual histories. Consider also that, if two or more individ-
uals experience events at the same time, that is, if we observe atie, then the model
proposed by Cox (1972) is the logit model. Therefore, although some authors have
argued that continuous-time methods are preferable to discrete-time methods on
theoretical grounds, or have completely ignored discrete-time methods altogether,
the lack of attention to the logit model seems rather unfortunate. The preference
for continuous-time models is largely based on computational grounds, but this is
certainly less of an issue today than it was 25 years ago!
7.5.2 Predetermined variables and unobserved heterogeneity
An important issue is the number and choice of covariates. Consider, for exam-
ple, economic expansions. If the length of the prior expansion and/or contraction
influences the exit probability of the current expansion, then the length of the