323 Coherence Consideration in Binary Time Series Analysis
14.4 Discussion
In this chapter, we explored the use of residual coherence as a graphical tool for identi-
fying potentially useful interactions in logistic regression for binary time series. In many
cases, in practice, the inclusion of interaction terms leads to improved models and, as was
illustrated, an interaction covariate could be more signicant than its factors. There are sit-
uations, however, when the difference between S
2
(λ; u) and S
1
(λ) is relatively large for a
frequency band due to a local minimum in the linear coherence S
1
(λ) giving rise to a rel-
atively large residual coherence RS(u) for some u without a substantial contribution from
X
u
(t). This is another reason why we need to couple the residual coherence with some
further evidence as to the importance of an identied interaction covariate. This can be
done, for example, by hypothesis testing and model selection criteria and by some sort of
residual analysis.
Clearly, instead of prescreening the variables using residual coherence, we could sim-
ply include interaction terms in the model and apply model selection. The advantage of
prescreening using residual coherence is that the search for useful interaction terms is facil-
itated using spectral information which accommodates model selection and hypothesis
testing. Moreover, exploring potential general relationships among time series is an impor-
tant time series problem where coherence measures, including residual coherence, as well
as other measures have an important role before tting parametric models.
References
Hinich, M. (1979). Estimating the lag structure of a nonlinear time series model. Journal of the American
Statistical Association, 74(366):449–452.
Kedem, B. (1980). Binary Time Series. Dekker, New York.
Kedem, B. and Fokianos, K. (2002). Regression Models for Time Series Analysis. Wiley, Hoboken, NJ.
Kedem-Kimelfeld, B. (1975). Estimating the lags of lag processes. Journal of the American Statistical
Association, 70(351):603–605.
Khan, D., Katzoff, M., and Kedem, B. (2014). Coherence structure and its application in mortality
forecasting. Journal of Statistcal Theory and Practice, 8(4):578–590.
Kimelfeld, B. (1972). An orthogonal representation of nonlinear systems. Technical Report No. 66,
Department of Statistics, Carnegie-Mellon University, Pittsburgh, PA.
Kimelfeld, B. (1974). Estimating the kernels of nonlinear orthogonal polynomial functionals. Annals
of Statistics, 2(2):353–358.
Koopmans, L. (1974). The Spectral Analysis of Time Series. Academic Press, New York.
Nelson, J. and Van Ness, J. (1973a). Choosing a nonlinear predictor. Technometrics, 15(2):219–231.
Nelson, J. and Van Ness, J. (1973b). Formulation of a nonlinear predictor. Technometrics, 15(1):1–12.
Priestley, M. (1988). Non-Linear and Non-Stationary Time Series Analysis. Academic Press,
San Diego, CA.
Shumway, R., Azari, A., and Pawitan, Y. (1988). Modeling mortality uctuations in Los Angeles as
functions of pollution and weather effects. Environmental Research, 45(2):224–241.
Slud, E. and Kedem, B. (1994). Partial likelihood analysis of logistic regression and autoregression.
Statistica Sinica, 4(1):89–106.
Tick, L. (1961). The estimation of transfer functions of quadratic systems. Technometrics, 3(4):563–577.
Wiener, N. (1958). Nonlinear Problems in Random Theory. M.I.T. Press, Cambridge, MA.