Joe Cardinale and Larry W. Taylor 341autonomous-shift dummy variables or lagged term. For with a single intercept,
there is a constant hazard unless time-varying covariates actually change in value.
A single-intercept probit model was chosen by Estrella and Mishkin (1998) to fore-
cast turning points in US recessions, and by Chin, Geweke and Miller (2000) to
forecast turning points in US unemployment.
Our empirical results also appear quite robust to dropping each spell, in turn,
from the respective samples of either upswings or downswings. For example, we
estimate the logit model with the full sample of ten downswings, as well as with
the sample of nine downswings that excludes, say, the third spell. At the 5% signif-
icance level, our sensitivity analysis again indicates that the interest rate spread is
statistically significant for downswings and that capacity utilization is statistically
significant for upswings.
7.9 Conclusion
Duration analysis has many uses, both in academe and in industry. Consider that
in a recent issue ofBusiness Week(January 22, 2007) there were two duration appli-
cations in separate fields. Hugh Moore of Guerite Advisors (Business Week,p.13)
notes that the average amplitude in the fall of housing starts is 51% from peak to
trough, and the average amplitude in the fall of housing expenditures as a per-
centage of GDP is 28% from peak to trough. Housing corrections, or recessions,
last an average of 27 months. In the same issue (Business Week, p. 62), the global
macro-group for Barclays Global Investors (BGI) reports devising a set of signals,
or leading indicators, that predict turning points from recession to expansion in
various countries. Profits are realized by buying stock and shorting bonds before
the recovery is generally recognized. Using leading indicators is similar to using
covariates in a duration analysis.
In this chapter we have emphasized classical, nonparametric methods in the
duration analysis of unemployment cycles. The nonparametric turning point
algorithm from Harding and Pagan (2002), or BBQ, is derived from the classical
graphical approch of Burns and Mitchell (1946). The life table analysis follows
directly from Cutler and Ederer (1958), and the logit model derives from the sem-
inal work of Cox (1972). These classical techniques are just as relevant today as
when first introduced, and have both micro- and macroeconometric applications.
Consider the many economic studies of individual job histories. Adamchik
(1999), for example, uses the nonparametric Kaplan–Meier estimator and the
semiparametric proportional-hazards model in her study of the effect of unemploy-
ment benefits on re-employment in Poland. In a related study, Bover, Arellano and
Bentolila (2002) examine not only unemployment benefits, but also the relation-
ship between the unemployment duration of Spanish men and the business cycle.
In the latter study, they employ a logit model with autonomous shift dummies
that is closely related to Cox’s famous proportional-hazards model. The approach
is very flexible since Cox’s model is no longer proportional when explanators that
vary with time are included in the model. Following the early frontier work of