Abstract

Life expectancy is the key indicator describing social and health aspects of human development. International comparative studies indicate that there is still no systematic evidence about reduction of longevity disparities between socioeconomic groups and areas and that they exist even in countries with strong social policies. The majority of the ongoing research focuses on mortality disparities by socio-economic status. However, the spatial dimensions of mortality changes are considered as equally important as they build the bridge between survival and socio-economic or cultural contexts. Yet measuring longevity disparities in small territorial units is a methodological challenge. Estimations of life tables for small areas might be based on direct data and standard demographic methods, statistical modelling, or combination of these two approaches. Prior studies show that very often conventional life table estimation methods may return implausible results. Therefore, the standard computation of life tables is usually complemented by some (often arbitrary) adjustments, smoothing or modelling parts of mortality curves. During the last decades the field of small area life table estimations has been increasingly relying on advanced statistical modelling approaches, including Bayes modelling. These approaches allow to obtain age-specific mortality estimates by using estimated parameters from a standard mortality schedule (e.g. national or higher rank regional unit) or by borrowing information for mortality estimation from either neighboring or similar areas or from areas with better quality data. As a result, one gets much more stable and statistically robust results, allowing to assess the magnitude and directions of changes of inequalities. The limitations of modelling include arbitrary choice of parameters, risk of overlooking specifics of mortality patterns in some areas. This study aims at testing three life table estimation strategies (direct data and standard estimation; TOPALS linear spline modelling; Bayesian modelling) using varying size municipality-level data for a small country, Lithuania. We found that application of the direct data approach is in certain situations impossible and/or requires substantial mechanical adjustments such as the aggregation over age groups and/or time intervals. In addition, some modelling is still required in order to get more plausible age-specific mortality patterns. Our results also suggest that even smoothing and modelling at old ages cannot solve all problems related to multiple zero cases and random fluctuations. In case of Lithuania, this approach also leads to implausible life expectancy values for smaller municipalities and wide confidence limits. Therefore, in many cases, the identification of direction of temporal changes in municipality-specific life expectancies is highly problematic. The modelling approaches such as TOPALS generally provide more stable and statistically robust estimates. For example, the modified TOPALS approach based on estimated parameters from standard (national) mortality schedule is efficient for producing more plausible age-specific mortality profiles for the smallest municipalities with multiple zero counts and random noise due to small numbers. Although the modelling approaches provide more stable, statistically robust, and realistic life expectancy estimates, they may lead to overlooking specifics of mortality in certain areas. More exhaustive analysis is needed to combine the advantages of both direct-data and modelling approaches.