Using comparability ratios for mortality data

Introduction

Mortality statistics are essential tools for monitoring public health, developing public health policy and evaluating the effects of policies and the outcomes of health services. Overall mortality or all cause rates in particular population groups are sufficient for some of these purposes. However, for many purposes we need to know what are the main diseases and injuries leading to death? How do death rates from these causes compare with those of other countries? How do they vary between different areas of the country and between groups of the population? How are they affected by health-related behaviour, exposure to potential hazards and available treatment? Is the burden of disease changing? To understand these issues we need to know whether mortality from particular diseases or injuries is increasing or decreasing over time.

To answer these questions, causes of death have to be classified in a reliable way, which is comparable between places and over time. An agreed system of classification, The International List of Causes of Death, was first developed for this purpose in the late nineteenth century.(World Health Organisation) Much of our understanding of the causes and prevention of disease has depended on comparisons between death rates coded and classified in the same way in countries around the world since then. Needless to say, a classification which was developed over a century ago would be unlikely to reflect current knowledge about the nature and causes of diseases, or the relationships between them. The need for regular updating of the classification has been recognised since the beginning of the twentieth century, and revisions produced almost every decade.

The International Classification of Diseases, Ninth Revision (ICD-9)(World Health Organisation, 1977) was in use between 1979 and 2000 for mortality statistics in England and Wales, so was widely recognised as being out of date when it was replaced by, the Tenth Revision (ICD-10)(World Health Organisation, 1992) in 2001. For example, AIDS and HIV were unknown when it was devised. This revision represented the largest change in the ICD in over 50 years. The first character of each code became alphabetic rather than numeric, enabling expansion of the number of codes to provide for recently recognised conditions and more detail about common diseases. Some diseases and groups of conditions moved from one chapter to another to reflect current ideas of aetiology and pathology at that time.(Ashley, 1990)

However, there is an additional factor in mortality coding which can have a major impact on mortality statistics. This is the application of strict rules for selecting the underlying cause of death from all the conditions written on the death certificate. Changes in these rules usually have much larger effects on the number and proportion of deaths attributed to particular causes than any other change in classification.(Rooney and Devis, 1996) Several of these rules changed substantially in ICD-10 (World Health Organisation, 1992).

As well as the need for an up-to-date classification, we need to be able to look at trends in cause specific mortality over time. In order to do this, we must be able to measure the effect of the change in classification on the proportion of deaths attributed to different causes. This is done using sample ‘bridge coding’ where a sample of deaths are coded to both the old and new classification, for example ICD-9 and ICD-10, and the results compared using statistics known as ‘comparability ratios’. A large bridge coding exercise was undertaken by the Office for National Statistics (ONS) for the change from ICD-9 to ICD-10.(Rooney et al., 2002)

Since 2001, ONS has implemented a number of updates to ICD-10 to keep it up to date ahead of eventually introducing ICD-11. Again bridge coding exercises, of samples rather than a full year of death records, were carried out to ensure that trends in mortality could be analysed consistently over time.

This document describes how comparability ratios are calculated, along with their confidence intervals, and explains how to apply them to analyse mortality data over specific time periods. Details on the specific changes at particular time points since 2001 are also provided. References to documentation about the change from ICD-9 to ICD-10 are also provided for those wishing to analyse longer-term trends.

Calculation of comparability ratios

Comparability ratios are calculated as the number of deaths coded to an underlying cause in the new version divided by the number of deaths coded to the same underlying cause in the old version. Using the bridge coded datasets, users can estimate the impact of the software and coding changes over time, by calculating comparability ratios.

Application of comparability ratios for specific time periods

It is normally recommended that comparability ratios are applied to historic data, to make them comparable to data coded using the most recent version of the ICD. It is only necessary to use comparability ratios if looking at trends in causes of death or grouping years of mortality data together.

In reporting cause-specific trends, ONS have generally taken the position that they do not report adjusted numbers of deaths. Instead, they have tended to report actual numbers of deaths and added warnings to users to note the likely impact of ICD revisions. Where comparability ratios are applied, it should be made clear that these are ‘adjusted’ numbers.

The comparability ratios can also be used to adjust mortality rates. The ratio can either be applied directly to the rate or the numbers of deaths can be adjusted before the adjusted rate is calculated. It is recommended that the latter option is used, so that the adjusted rate and its confidence interval are both based on adjusted numbers of deaths. Since the ratios are different for different age groups and for males and females, applying an overall single comparability ratio to the calculated rate would require this to be calculated taking into account the age/sex distribution of deaths by cause. If the numbers of deaths are adjusted, however, the appropriate ratio can be applied to each age/sex-specific count.

The published ratios for ICD-10 can be applied to mortality data in England and Wales back to 2001, the year that ICD-10 was introduced. If looking at trends across a range of years, Table 1 provides guidance on which version of the comparability ratios should be used for each time period. If analysis needs to go back before 2001, ONS published ratios can only be applied to ICD-9-coded data from 1993 to 2000 since there was a significant change between 1992 and 1993 brought about by the introduction of automated cause coding.

Calculation of confidence intervals

When adjusted counts of deaths are used to calculate rates or ratios, there are two potential ways of calculating confidence intervals for the statistics:

1. the adjusted counts can be used to determine the confidence intervals as if they were actual observed counts
2. the rate or ratio (R) and its confidence interval (\(R_L\), \(R_U\)) can be calculated using the original counts, and the confidence interval applied to the adjusted rate or ratio (\(R^*\)) as follows:

\[ R_L^*=\frac {R_L \times R^*}R , R_U^*=\frac {R_U \times R^*}R \]

While the second method is arguably more accurate, as it uses the actual observed counts to define the confidence interval, it will make little difference if the ratios are fairly close to 1, as they are in the majority of cases. Hence it is recommended, pragmatically, that the adjusted counts are treated as if they were observed counts, and the confidence intervals calculated using the standard methods.

Attributable fractions

Comparability ratios can be used to adjust numbers of deaths for the calculation of mortality indicators which involve the application of attributable fractions, for example to estimate numbers of smoking-related or alcohol-related deaths. In these cases, the comparability ratios should be used to adjust numbers of deaths separately for each cause, or group of causes, within the definition. Attributable fractions can then be applied to these adjusted numbers. Where the attributable fraction is 1 (for example for causes in which all deaths are accepted as alcohol-related) these can be adjusted as a single group of causes.

Details of software in use from 2001 onwards

From January 2001 until December 2010, ONS used the Mortality Medical Data System (MMDS) ICD-10 v2001.2 software provided by the US National Center for Health Statistics (NCHS) to code cause of death. Since then:

• ONS updated to ICD-10 v2010 in January 2011 and incorporated most of the World Health Organisation (WHO) amendments authorised up to 2009

• IRIS version 2013 was introduced in January 2014, which incorporated all updates to ICD-10 authorised by the WHO up to the end of 2013

• In January 2017, ONS updated to IRIS 4.2.3 software, which incorporated all changes to ICD-10 that had been authorised by the WHO up to the end of 2016, but did not produce a bridge coded dataset to compare differences between IRIS 2013 and IRIS 4.2.3

• In January 2020, ONS updated to a successor of IRIS, Multi-causal and Uni-causal Selection Engine (MUSE) 5.5 (IRIS version 5.5), which incorporated all changes to ICD-10 that had been authorised by the WHO up to the end of 2018

• In January 2022, ONS moved to MUSE 5.8, which incorporates ICD-10 changes authorised up to 2019.