Trying to figure out where the incipient swine flu pandemic is heading and how fast it is heading there is shooting at a moving target, and this one is moving pretty fast. The best we can do at this point is use whatever information we have to make some educated guesses about different scenarios along with how likely various scenarios are. We used to do this on the back of an envelope, Now we use computer programs. I’m not sure we are doing much better (or much worse), but we can make use of more information and the answer looks prettier when displayed.
Expedited publication of such an assessment is now online at the journal Science. Imperial College (UL) modeler Neil Ferguson and an international team used available early data from Mexico and information on international spread to try to figure out how transmissible the swine flu virus is and take a measure of its virulence. Here’s their summary:
Our estimates suggest that 23,000 (range 6,000-32,000) individuals had been infected in Mexico by late April, giving an estimated case fatality ratio (CFR) of 0.4% (range 0.3% to 1.5%) based on confirmed and suspect deaths reported to that time. In a community outbreak in the small community of La Gloria, Veracruz no deaths were attributed to infection, giving an upper 95% bound on CFR of 0.6%. Thus while substantial uncertainty remains, clinical severity appears less than that seen in 1918 but comparable with that seen in 1957. Clinical attack rates in children in La Gloria were twice that in adults (<15 years-of-age: 61%, ≥15: 29%). Three different epidemiological analyses gave R0 estimates in the range 1.4-1.6, while a genetic analysis gave a central estimate of 1.2. This range of values is, consistent with 14 to 73 generations of human-to-human transmission having occurred in Mexico to late April. Transmissibility is therefore substantially higher than seasonal flu, and comparable with lower estimates of R0 obtained from previous influenza pandemics. (Fraser et al., Science)
There have been a number of deaths from this virus (WHO tallies 53 as of this writing), mainly so far in Mexico, but without knowing how many people are infected (including mild and asymptomatic cases) it is difficult to know how virulent this virus is. This model tries to determine the rough size of that denominator. The estimates in this paper suggest the swine flu virus virulence to be comparable to the 1957 pandemic (“Asian flu,” H2N2), somewhere around 0.4%. Seasonal flu has a case fatality ratio (CFR) of roughly 0.1%. As soon as you start to look at the assumptions that were made, however, you begin to see the level of uncertainty here.
The CFR has a numerator (number of deaths among the infected) and a denominator (the total number infected). The deaths may be undercounted in certain age groups (e.g., the elderly) where many competing causes of death may mask flu related deaths. On the other hand, not all deaths put down as flu related in the early phases of this outbreak, modeled in this paper, may be flu related. In addition, the data are what statisticians call “right censored,” that is, the eventual outcome of many cases in the denominator was as yet unknown. At some point after they were counted as an infected case they may die from a flu related cause but wouldn’t have been counted in the numerator. Special techniques, using time to event modeling, are often used to take deal with right censored data like these.
The denominator, the total number of infected, has even more uncertainty. Since many mild infections were probably missed in Mexico, the paper used the prevalence of infection among travelers as an estimate on the assumption there was more intense surveillance. Using this number the number of cases in Mexico was back calculated. But there are some assumptions, here, too:
Key underlying assumptions in this analysis are that the population mixing in Mexico is equally likely between Mexican residents and tourists, and tourists and Mexican residents are at equal risk of infection (despite demographic and other differences). If infections are concentrated away from traveler destinations, the number of people infected in Mexico will be under-estimated, and conversely will be over- estimated if the epidemic has disproportionately affected geographical zones visited by travelers. Under the assumption that reporting of infections in travelers was complete we obtained estimates of the number of infections occurred in Mexico by late April from a model of the interval-censored country case counts which varied between 18000 and 32000, depending on the mean duration of stay of tourists assumed, with perhaps the most credible single value (based on journey duration data) being 23000 [figure references omitted].
Other assumptions are possible, and Fraser et al. discuss some of them. The main point, however, that when viewing results of a paper like this it is best to consider them as highly tentative. As this unfolds further, we may get a better idea of virulence as expressed by the CFR. At the moment it looks similar to the 1957 pandemic virus, no where near as bad as 1918 but worse than seasonal flu. This wouldn’t be surprising. But neither would a variety of alternatives, given the uncertainty.
Similar observations can be made about the measures of transmissibility and the timing of when the outbreak started. The authors consider the virus to be more transmissible than seasonal flu, again a characteristic of pandemic strains. But the uncertainties here are even greater than with the CFR and the assumptions are, too. Some of the more qualitative features are quite robust, however. It is pretty clear that we are seeing many generations of cases and that sustained person to person transmission is occurring.
Since we don’t understand the dynamics of seasonality, we still don’t know how the emergence of this virus in the “off season” affects things like severity and transmissibility (if at all). We’re heading into flu season for the southern hemisphere, so we may learn more in the next few months.
The authors are frank about the uncertainties, as well they should be:
The future evolution of the transmissibility, antigenicity, virulence and antiviral resistance profile of this or any influenza virus is difficult to predict. It is also unclear whether it will displace existing influenza A subtypes from the human population, as occurred in the 3 last pandemics.
In other words, don’t rely too much on the results of this first pass at the data. Unfortunately, despite all the caveats in the world, the fact that this was done “by a computer” is likely to give the numbers a life of their own. This isn’t a slap at the value of modeling, something about which I am on record as favoring.
I just want to remind everyone not to make the models bear more weight than they are built to sustain.