Cause-of-death time series are severely disrupted by periodical changes in the disease classifications. Existing international databases do not fix this problem. The information provided by the WHO database is fully ruled by successive ICD revisions and reflect their discontinuity. Eurostat data series are more consistent over time, but for quite a short period of time (from 1994 at best) and for too large groups of causes. To reconstruct consistent series, it is necessary to establish transition coefficients between items of two successive classifications, in order to redistribute deaths classified according to the old classification into items of the new classification. When bridge coding (double classification of a sample of deaths simultaneously into the old and new classification) has been performed at a detailed level, transition coefficients can be inferred directly from the results, but there are only two countries in the database where this has been done (and only for the last transition, i.e. from ICD-9 to ICD-10), namely England and Wales and the U.S.A. For the other transitions in the two countries and for all transitions in other countries, coherent time series are reconstructed by producing ex-post double coding. The method developed at INED in the 1980s is used as a guideline, but the work was tailored to each country independently.
For each classification change, the method comprises three steps (Vallin and Meslé, 1988, 1998; Meslé and Vallin, 1996):
- Setting up one correspondence table which lists, for each item of one classification, all the items of the other one that are a priori equivalent in terms of medical content.
- Building fundamental associations of items that identify the smallest possible number of items containing the same medical contents in both classifications and testing the consistency of the associations over time using a statistical test (Barbieri, Chung, and Boe, 2008; Camarda, Pecholdová, and Meslé, 2015).
- Setting up ex-post double-coding according to the structure of fundamental associations, to finally obtain transition coefficients.
The results derived from the medical logic of the classification rules have to be checked statistically, to detect and solve any remaining breaks in the series. Such checks are carried out by age group and sex.
In addition, national statistical offices make occasional changes independent of the official revisions of the classification. To address this problem, the statistical continuity of the series over time is systematically verified and any artificial disruption dealt with appropriately.
Finally country- and time-specific methods are used to deal with ill-defined causes (Ledermann, 1955; Vallin and Meslé, 1988).
- Barbieri, M., Chung, R., & Boe, C. (2008). Automating the redistribution of deaths by cause over ICD changes. Second Human Mortality Database Symposium, Max Planck Institute for Demographic Research, Rostock, Germany, 13-14 June 2008.
- Camarda, C.G., Pechholdová, M. & Meslé, F. (2015). Cause-specific senescence: classifying causes of death according to the rate of aging. 80th Annual Meeting of the Population Association of America. San Diego (USA), May 2015. http://paa2015.princeton.edu/uploads/153074
- Ledermann, S. (1955). La répartition des décès de cause indéterminée. Revue de l’Institut international de statistique, 23 (1–3), 47–55.
- Meslé, F., & Vallin, J. (1996). Reconstructing long-term series of causes of death. Historical Methods, 29 (2), 72–87.
- Vallin, J., & Meslé, F. (1988). Les causes de décès en France de 1925 à 1978 (Travaux et Documents, No.115, 608 p.). Paris: INED/PUF.
- Vallin, J., & Meslé, F. (1998). Comment suivre l’évolution de la mortalité par cause malgré les discontinuités de la statistique. Le cas de la France de 1925 à 1993. In G. Pavillon (Eds.), Enjeux des classifi cations internationales en santé (Questions en santé publique, pp. 113–156, 220 p.). Paris: Éditions INSERM.