Life expectancy estimation in small administrative areas with non-uniform population sizes: application to Australian New South Wales local government areas

BMJ Open. 2013 Dec 2;3(12):e003710. doi: 10.1136/bmjopen-2013-003710.


Objective: To determine a practical approach for deriving life expectancy estimates in Australian New South Wales local government areas which display a large diversity in population sizes.

Design: Population-based study utilising mortality and estimated residential population data.

Setting: 153 local government areas in New South Wales, Australia.

Outcome measures: Key performance measures of Chiang II, Silcocks, adjusted Chiang II and Bayesian random effects model methodologies of life expectancy estimation including agreement analysis of life expectancy estimates and comparison of estimate SEs.

Results: Chiang II and Silcocks methods produced almost identical life expectancy estimates across a large range of population sizes but calculation failures and excessively large SEs limited their use in small populations. A population of 25 000 or greater was required to estimate life expectancy with SE of 1 year or less using adjusted Chiang II (a composite of Chiang II and Silcocks methods). Data aggregation offered some remedy for extending the use of adjusted Chiang II in small populations but reduced estimate currency. A recently developed Bayesian random effects model utilising the correlation in mortality rates between genders, age groups and geographical areas markedly improved the precision of life expectancy estimates in small populations.

Conclusions: We propose a hybrid approach for the calculation of life expectancy using the Bayesian random effects model in populations of 25 000 or lower permitting the precise derivation of life expectancy in small populations. In populations above 25 000, we propose the use of adjusted Chiang II to guard against violations of spatial correlation, to benefit from a widely accepted method that is simpler to communicate to local health authorities and where its slight inferior performance compared with the Bayesian approach is of minor practical significance.