User:Forkandwait/mortality forecasting

Forecasting life expectancy and mortality forms an important subdivision of demography. Future trends in life expectancy have huge implications for old-age support programs like Social Security and pension systems, because the cash flow in these systems depends on the number of recipients still living (along with the rate of return on the investments or the tax rate in PAYGO systems). With longer life expectancies, these systems see increased cash outflow; if these systems underestimate increases in life-expectancies, they won't be prepared for the large payments that will inevitably occur as humans live longer and longer.

Life expectancy forecasting usually is based on two different approaches:


 * Forecasting the life expectancy directly, generally using ARIMA or other time series extrapolation procedures: This approach has the advantage of simplicity, but it cannot account for changes in mortality at specific ages, and the forecasted number cannot be used to derive other life table results.  Analyses and forecasts using this approach can be done with any common statistical/ mathematical software package, like R, SAS, Matlab, or SPSS.


 * Forecasting age specific death rates and computing the life expectancy from the results with life table methods: This approach is usually more complex than simply forecasting life expectancy because the analyst must deal with correlated age specific mortality rates, but it seems to be more robust than simple one dimensional time series approaches.  This approach also yields a set of age specific rates that be used to derive other measures, like survival curves or life expectancies at different ages.  The most important approach within this group is the Lee Carter method, which uses the singular-value decomposition on a set of transformed age-specific mortality rates to reduce their dimensionality to a single time series, forecasts that time series, and then recovers a full set of age-specific mortality rates from that forecasted value.  Software for this approach  include Prof. Hyndeman's R libraries and UC Berkeley's LCFIT system.