Alternative Time-Series Approaches to Mortality Projection

Ward Kingkade, U.S. Census Bureau

In the field of mortality analysis and projection, the Lee-Carter methodology occupies a central position. This methodology combines the Singular Value Decomposition and modern time series analysis affords data reduction as well as forecasting components. Where data are noisy (e.g. raw death rates by single years of age) and time series are short, this methodology becomes difficult to apply. The proposed paper presents an analysis of US mortality data in ethnic detail, by sex and single-year age, comparing results from the Lee-Carter Methodology to an alternative approach which employs a relational logit model for data reduction and affords similar time-series capabilities. The latter approach is in the vein of earlier work by McNown and Rogers (1989). The model-based approach yields more appealing results under the data availability constraints, while retaining stochastic forecasting capacity, and may be better suited to applications where data are more limited, or which require greater detail.