The newswires are humming again with another story of the estimated toll a flu pandemic might exact, if it were as bad as 1918. This time the occasion is a paper just published (.pdf) in the British medical journal, The Lancet, which attempts the most careful estimate yet of the toll of the 1918 epidemic. As often happens, the headlines, and in some cases the contents of the articles are all over the place. Some examples: “Bird Flu Can Kill 62 Million People“, “New flu pandemic could kill 81 million“, and “Study Finds Much Bird Flu Planning is Misplaced.” We took a look at The Lancet article and here’s our take.
The authors used international mortality data from 27 countries and regions where there was a minimum of 80% death registration in the years 1915 to 1923, trying to correlate the wide variation in mortality with community attributes. By restricting their data set to these “registration areas,” they are able to make more comparable comparisons between areas that differed in mortality. They selected two features of the communities that might indicate important differences, per capita income and latitude. The former is a surrogate for many factors that could contribute to differences in mortality, including access to medical care, education, nutrition and many other things that might be related to both income and the risk of dying from influenza. Comparable data are available for most countries in the world today, so any relationship might be used to make an estimate of current mortality variations from a pandemic of the same character as 1918. This assumes that the relationship between per capita income and excess mortality from influenza is of the same nature and effect as it was in 1918. The authors give some reasons to justify this assumption. Latitude is a feature of geography, in particular seasonal variation and level of temperature which some think might be related to influenza outbreak patterns. This is feature that has not changed appreciably since 1918. The correlations they derive are statistical in nature but are derived from actual data. In effect, they attempt to summarize the data in a manner that would allow some insight into why one place had a higher mortality than another. It is a task fraught with difficulty.
Previous studies have used records of “normal” seasonal variation to estimate how much excess mortality is produced by an epidemic by subtracting the normal from the outbreak period; or they have used cause specific mortality to estimate influenza deaths at particular points in time. Coding of cause of death changes over time, however, which produces additional complications. The authors of this paper elected to use mortality from all causes to avoid this problem (the fact of death is more certain than the cause of death). This also would take account of the possibility that influenza might affect other causes of death. Their data did not allow a week by week or month by month analysis, however, and they settled for annual averages. To calculate the excess of total deaths from influenza they averaged the death rates in the three years before the pandemic, 1915 – 1917, and the three years after the pandemic, 1921 – 1923, and compared these average rates to the rates in the years 1918, 1919 and 1920 (the “pandemic years”). By taking the difference in the pandemic years and the non-pandemic years on either side of them they calculated the excess of total deaths in the countries and regions where they felt the data were reliable. In some places they were also able to calculate death rates in specific age categories, as well.
One finding is not new but bears repeating. There was an extremely wide variation in excess mortality from country to country and region to region. Even within the same registration areas, for example, the Census of India, there is a 2.1% excess mortality attributed to the pandemic in Burma compared to a 7.8% in the Indian province of Berar. A 2.1% excess means that 2.1% of the population died of influenza-related causes (by this method of estimation), not 2.1% of the influenza cases. Using excess death percentages like this allows one to estimate the number of deaths that might be caused in a population of a specific size without making an assumption about infection rates and case fatality.
Using this method and some available age and sex specific data from a subset of the countries and regions, the authors were able to confirm the notorious W-shaped mortality curve for age. Here is Figure 1 from the paper:
The next step in the analysis employs a statistical model to see what the effect of the two community characteristics, per capita income and latitude. One surprise is that the single variable of per capita income in a country explains half of the variation in excess mortality. This is a rather large effect of community wealth. What about latitude? The authors say in their text that adding this indicator of distance from the equator produced little effect on the ability of the model to predict excess mortality and that the effect was not statistically significant. [Small technical quibble: Examination of Table 2 in the paper doesn't quite make that case because it presents only beta coefficients, standard errors, p-values and t-statistics from two models, one with per capita income only and one with both per capita income and latitude. What would be useful for a reader paying attention would be an extra line giving the same estimates from a model with latitude only as an independent variable. The reason I say this is because it is quite plausible that these two variables, latitude and per capita income, might be highly correlated, that is, countries and regions that are in the lower latitudes might also be much poorer. This produces a colinearity and would lead to a model that is potentially unstable whose parameter estimates are highly dependent on the particular dataset. These are experienced investigators and the analysis I am suggesting is so easy I feel confident it was taken into account, but the paper is not quite complete without it. It is not the only editing lapse. The units for the latitude model are not given, so the beta coefficient is hard to interpret as an effect measure (yes, that's a real term in epidemiology!). Another complaint is that it isn't easy to see how the variance of the excess mortality was estimated, although there is some mention of the role of unexplained variance in the model. With a paper of this much importance, extra care should be taken to spell out the methodology better for non-specialists. End of technical quibble.]
In any event, the statistical model can then be used to predict excess mortality for a given population size with various age categories and per capita income as of 2004 (the latest year for which such data were available). In this way the authors toted up the excess deaths globally if a strain with the severity of 1918 were to emerge. The numbers are surprisingly high, with a median (50th percentile) of 62 million and a 10th to 90th percentile range of 51 million to 81 million. If these excess deaths were to occur in a single year, it would double the globe’s mortality. Not said, the excess would pile up in a flu season, not spread out over the year. Thus the impact would be considerably greater, leading to perhaps an eight fold mortality excess over a quarter. This is an effect of catastrophic proportions. Nor is the estimated mortality logically or biologically a worst case:
In most discussions of influenza, the 1918?20 pandemic sets the upper limit, in terms of mortality, on what might occur in future pandemics. However, there is no logical or biological reason why that pandemic ?albeit very severe ? should represent the maximum possible mortality in a future pandemic. Random genetic mutation could, in principle, produce a more lethal virus, although pathogens that are too lethal might not survive long enough in the host to effectively transmit to different populations. In addition to this uncertainty about what is genetically possible, future mortality could be larger if the 1918?20 pattern of low older adult mortality were in fact due to some acquired immunity from the pandemics of the mid-19th century. (Murray et al., The Lancet, cites omitted)
But there is another, and more important, surprise. Because of the strong effect of per capita income in the statistical model (NB: a statistical model is not the same as the mathematical models we have discussed elsewhere), the lion’s share of the burden of a pandemic will be felt in the developing world, by their estimate, over 95% of it. Thus less than 5% of the mortality excess will be in the rich countries, although most of the preparation for a pandemic is occurring in those countries. Sub-sharan Africa and east and southeast asia will feel the hammer blow.
Whatever the final toll, these data make a case that a source of the almost 40 fold variation in mortality seen around the world in 1918 was related in some fashion to poverty. If you live in a developed country, this might bring a sigh of relief. “They” will suffer but you won’t. But most developed countries, and the United States is a prime example, is “marbled” with poverty throughout, so pockets of high mortality would also occur throughout our society. The societal “add on” of this will affect everyone in a world where we have become highly interdependent and closely connected in so many ways. Even if this weren’t true, consigning most of the world to death from a pandemic is not something any government should find acceptable. It lets no one off the hook, even if one assumes you wouldn’t suffer (an invalid assumption in this case).
This paper reinforces what we already “know” but often refuse to acknowledge. Whatever the analysis, the bottom line seems to be that the best way to protect ourselves is to have a robust and resilient society with an intact, effective and functioning public health and social services infrastructure. It is highly likely that the variable per capita income is a surrogate for the benefits those things bring to a community’s health.
How many death’s will it take before we know, that too many people have died?