Ethnic inequalities in mortality rates and life expectancy in England and Wales: Why we should treat experimental statistics with caution

Experimental statistics, published earlier this week by the ONS, suggest that prior to the Covid-19 pandemic ethnic minority people had lower mortality rates and longer life expectancies than White people. Although received with some surprise, these findings were not unexpected. They reflect earlier analysis of mortality rates during the Covid-19 pandemic, which showed the now well-known higher rates of Covid-19 related mortality for ethnic minority people, but lower overall mortality rates.

So how should we interpret such findings? Although some have taken these findings to indicate that the inequalities experienced by ethnic minority people during the pandemic do not reflect wider and longstanding inequalities in health and their determinants, it is important to remember that these are experimental statistics – statistics that should be treated with caution and not considered to be robust and conclusive.

Indeed, even beyond their experimental nature, there is good reason to treat them with caution. It is counterintuitive that Covid-19 related mortality would show such greatly increased risks for ethnic minority people, but overall mortality would show the opposite. Counterintuitive unless, we are comfortable with the argument that the broader social and economic inequalities that shape mortality risk do not shape the increased risk of Covid-19 related mortality faced by ethnic minority people, and instead that we can find the causes of this increase risk in the attributes of ethnic minority groups themselves, rather than in the deep and persisting inequalities they face as a result of structural, institutional and interpersonal racism.

Any concern with the interpretation of the findings published in this ONS report encourages a close look at the analysis behind it. A starting point is to note that estimates of mortality rates require a count of the number of people who died (within a particular time frame) and the size of the population from which they came. The resulting calculation is simple, number of deaths divided by population size.

An estimate of life expectancy is a bit more complex. It uses number of deaths within a particular age group and the size of the population within that age group to estimate survival into the next higher age group. It does this sequentially across age groups, and thereby provides an estimate of how long someone of a particular age would live. It then combines those estimates to provide an overall life expectancy for the population under study. Although more complex, it uses the same data as those used to estimate mortality rates.

However, comparisons of mortality rates and life expectancy across ethnic groups in England and Wales is not quite so straightforward and requires quite innovative work.

First, we do not record ethnicity on death certificates, so to count the number of deaths within an ethnic group we need to match the death certificate to a source of data on the person’s ethnicity, such as a census record.

Second, we don’t know the size of each ethnic group. To calculate this we need to use the estimate provided by the most recent Census (in the current context this is the 2011 Census, so ten years out of date), then take into account ethnic specific rates of emigration out of the country. (In addition, the ONS does two other things to account for the time period between the 2011 Census and the estimate of mortality rates, they exclude from the data i) people who were born after the census, and ii) people who immigrated into England and Wales after the census.)

So, what are the clues that the innovative, but experimental, work done by the ONS to estimate ethnic differences in mortality and life expectancy has used approaches to addressing these problems that might lead to inaccurate conclusions?

First is the extensive, and reliable, evidence that ethnic minority people have much poorer health than White people. And we know that mortality is closely related to morbidity. The only way to ignore this concern would be to claim this connection between morbidity and mortality is weaker for all groups of ethnic minority people than White people.

Second, some of the findings in the data tables published by the ONS as appendices to the main report are frankly bizarre.

For example, they estimate the life expectancy of an 80-84 year old Bangladeshi woman to be 15.5 years, an 85-89 Bangladeshi woman to be 13.5 years, and a Bangladeshi woman aged 90 or older to be 11 years. Equivalent figures for Black African women are 15.7 years, 13.3 years and 11.7 years. While equivalent figures for White women are a much more sensible: 9.9 years, 7.2 years and 5.3 years.

The figures for the Bangladeshi and Black African groups are, frankly, amazing – they suggest that a Bangladeshi or Black African woman who survives to 85 or older will also survive to 100 or older – and they exceed those for countries with the greatest longevity, such as Japan, where in 2014 (the last year of the period covered by the ONS report) life expectancy for a woman aged 80 was 11.7 years, for a woman aged 85 was 8.4 years and for a woman aged 90 was 5.7 years. It must be asked why, when confronted with such figures, the analysts did not look a bit more closely at the assumptions in their models?

So where did those assumptions go wrong?

First, if you overestimate the size of the population, you underestimate the mortality rate and overestimate the survival rate (see the earlier description of how these figures are calculated). The ONS made two very crude assumptions that could have led to such an error.

First, to estimate the size of the population within which deaths occurred (and the number of deaths) the ONS analysts linked census data to patient register data and death certification data. Where there was a mismatch between these data sources the ONS made an estimate of the undercount of the size of the population for a particular ethnic group (the process is described in an appendix to the report), and this undercount was greater for ethnic minority groups. They then provided an adjustment for the undercount, and the subsequent adjustment inflated their estimate of the size of ethnic minority populations to a greater extent than that of the White population, and to a greater extent than the estimate of the number of deaths.

So, an examination of the data that the ONS provide suggest that once adjusting for all other factors the ONS estimate that the 2011 Census undercounted the Bangladeshi population by 6% more than the White population, and the Black African population by a massive 47% more. These numbers undoubtedly lead to a greater reduction in the estimate of mortality rates for ethnic minority groups compared with White people. The size of this effect is not visible in the reports, because the ONS do not provide details nor sensitivity analysis. At a minimum the latter should be provided with experimental statistics

Second, it is likely that the ONS underestimated the number of ethnic minority emigrants when correcting the population size from the counts at the 2011 Census.

This correction is necessary because some of the people recorded at the census will have left England and Wales since then, so will not be in the death statistics (deaths overseas are not recorded in the death certification process). The results of this estimation process are surprising.

For example, for Bangladeshi men aged 65 or older compared with White men aged 65 or older the ONS estimated that there were only 7.9 more emigrants per year per 1,000 in the population, while for Black African people the number is only 4.6 additional people. Why might these figures be wrong?

First, the ONS use a linkage between a 1% sample of the census (the ONS Longitudinal Study) and administrative data on emigration and deregistration from the patient register to count the number of emigrants. It is likely that both emigration records and deregistration data will underestimate the number of emigrants and therefore inflate population size, and, if ethnic minority people are more likely to emigrate (particularly at older ages where most deaths occur), then the ethnic minority population will be overestimated to a greater extent than the White population.

To try and strengthen the analysis, the ONS also use records from the International Passenger Survey to estimate emigration by ethnicity. But the International Passenger Survey does not record ethnicity, so they use the correlation between citizenship and ethnicity to make their estimates and do not discuss the fact that the vast majority of ethnic minority people in the UK are UK citizens, making such estimates of very limited validity.

So, both the approach to census undercount and the approach to address emigration are likely to have led to an overestimate of the ethnic minority population relative to the White population. This then produces a consequent deflation of ethnic minority mortality rates and inflation of their life expectancy relative to White people.

 And, there is a final major problem.

As well as underestimating ethnic minority emigrants relative to White emigrants, as described above, the ONS analyses do not have data on mortality rates for emigrants, so they have statistically removed them from the data. However, mortality rates for emigrants should not be disregarded in analysis of ethnic inequalities; migrants’ experiences and circumstances prior to emigration will shape their mortality risk. There are two relevant and very well documented theories in relation to this.

First is health selection – those who emigrate to a new country are more likely to be healthy than those who stay. So White migrants to a new country will have a lower mortality rate than White people who remain, leading to the risk of an overestimation of mortality rates for White people – admittedly to a very small extent.

Second is salmon bias – those who return to their country of origin are more likely to have poor health compared with those who stay in the country they originally migrated to. So ethnic minority people who return to their country of origin are likely to have a higher mortality rate than those who remain. Removing emigrants from the data will underestimate mortality rates for ethnic minority people.

So, there are important concerns that we should take into account when interpreting the experimental statistics on ethnic differences in life expectancy and mortality rates produced by the ONS. It is likely that the key problems discussed above will have led to an underestimate of mortality rates and an overestimate of life expectancy for ethnic minority people compared with White people.

There are also other concerns about the analysis – for example, combining all White groups into one, when some White minority groups face marked health inequalities compared with White British people (Gypsy, Roma and Traveller people, for example); and the process of age standardisation of data, where the choice of the ‘standard’ population runs the risk of amplifying statistical bias.

With the information supplied by the ONS it is not possible to estimate the size, or significance of the problems with these data. It is unfortunate that these issues were not made clearer and were not, it seems, explored in more detail prior to the release of these experimental statistics. Nonetheless, it is important to remember that these are experimental statistics, and no assertions on ethnic inequalities should be made based solely on them. Indeed, it is important to continue to pay attention to the extensive evidence we have on ethnic inequalities in health, in experiences of health and social care, and among the health and social care workforce.