The other day we observed that through week 51 of 2006 (two weeks ago) flu activity in the US looked about "normal" except for the dominance among subtyped isolates of H1 influenza instead of the more usual H3. I began to wonder how common or uncommon this was and have done a little digging, but haven't found all the information I am looking for yet concerning the relative dominance of H3 versus H1 in the years since 1977, which is when H1 made its reappearance. Data on CDC's flu activity website regarding subtypes only goes back to 2000 - 2001 (at last that I could find), and that is also the first year before this one where H1 dominated. I have a paper by Simonsen et al. from Archives of Internal Medicine two years ago that has a table (Table 1, p. 267) indicating which years between 1977 and 2001 had either H3 dominance or one of either H1 or influenza B dominance. This is almost what I want, except that I can't differentiate H1 dominance from influenza B dominance from this table. From 1976 to 2006 H3 has dominated 19 times and either H1 or B 12 times. Since 1990, it has been an H3 year every year except 1990-91, 1992-93, 1995-96 and the aforementioned 2000-01, 12 out of 17 years. Some of the five years may be influenza B, except I know that the last of them, 2000-01, was an H1 year.
The Simonsen et al. paper is interesting for a number of reasons. The group around Simonsen at NIAID and Viboud at Fogarty International Center (both part of NIH) are among the top experts in examining influenza over time and making inferences about influenza-related excess mortality. One of their findings has been that H3 years are almost three times worse in terms of pneumonia and influenza mortality than H1 years. If this pattern holds, the fact that this appears to be an H1 year is good news. I'll give a very brief explanation of their general methodology because it is the one CDC uses to estimate excess mortality from influenza (the widely quoted 36,000 death figure). This doesn't bear directly on the H1 and H3 question, but will be of interest to readers here.
The researchers first extracted the numbers of all deaths that had pneumonia and influenza (P&I) listed as one of the underlying causes of death in national death data for the years 1968 to 2001. they then tabulated them by month and in five year age groups between 65 years old and 85+. This is more difficult than it might sound because the International Classification of Disease coding scheme changed a number of times in this interval (we are now in ICD-10). Thus adjustment had to be made over different data sets and is tricky. But that's what they get paid the big bucks for.
Now they have month by month P&I deaths for the period -- 408 data points covering 33 influenza seasons. The "baseline" force of mortality in the US changed over this time, so they "de-trended" the data by removing changes in the summer months (when influenza is least active). The assumption here is that the summer number represents the floor above which influenza fluctuates, so they canceled out the changing height of the floor. That's not exactly what they did, but that's the idea behind their de-trending. Now they have P&I deaths throughout 33 years, all with the same baseline.
The first thing you always see with monthly influenza data is its strong seasonality. There is something that makes influenza a seasonal disease and no one can agree exactly what it is. The two broad category of explanations are that there is some seasonal driving force (temperature or humidity or some behavior change like "back to school"); or that the dynamics of the viral infection (how it interacts with its host) leads to periodic or cyclical behavior that is roughly the same as the seasons. Most people are in the former camp, but the latter is a possibility. Anyway, you see these troughs and peaks, with the peaks in the winter and the troughs in the summer. The waves are not completely regular, however, because when you get to "flu season" you see irregularly timed and sized spikes poking through a more regular wave-like form. Those spikes are "flu outbreaks" and it is the difference between those spikes and the more regular wave-like pattern that represents the "excess P&I mortality." From this point on all you have to do is subtract the regular wave-like pattern from the monthly mortality data and you have your excess mortality number. But where do you get the regular wave-like pattern that you are going to subtract? You have to estimate it with a statistical model.
Simonsen et al. use a variation of something called Serfling's method, first published for this purpose in 1963 and still in use. Many people are familiar with the statistical method called linear regression, where a best fitting line is calculated for a scattering of points. Linear regressions are used to indicate trends in data or to fashion formulas that will allow you to predict outcomes given the kind of information that was used to produce the regression line. Serfling's method is much the same, except that instead of fitting a straight line it fits a wave-shaped curve (a sinusoid). The difficulty is, however, that you can't use the flu months to fit the wave because those are the months that have outbreaks, i.e., the non-wavelike months. So instead what is done is to use the data from all the months except December through April to estimate the sinusoid curve in the hopes that it will give you what would have happened if there were no outbreak, i.e., the usual flu background. In essence, they are using the data in the months from May through November to predict what should be flu in December through April if there were no outbreaks. Once they have an estimate for that, they can subtract that predicted number from the number that actually occurs in those months to give excess deaths.
This is a method that sounds like it shouldn't work, but it seems to work pretty well, at least as measured by how remarkably well it "fits" the non outbreak monthly data. There are clearly some assumptions that have gone into this but none that are outlandish and there is every reason to believe it gives a pretty good estimate of excess P&I mortality. That's not to say there aren't problems with it or reason to try to improve on it. Right now, however, it's the best method in regular use.
To sum up, we seem to be having an H1 year in seasonal influenza, which is different than most recent years which have been H3 years, and also has a better prognosis in terms of overall mortality from pneumonia and influenza. The big question, then, is whether it will be joined by H5N1, which would be, to understate it a bit, a Very Bad Thing.
- Log in to post comments
From the peanut gallery. . . Would H5N1 have a preference to "combo up" with either H1 or H3 to make it easier to "go global" ???
The peaks and the valleys Revere... Summer against winter. Why if its infectious do we have those. Is it because the UV is higher, people are spread out more during summer? I read somewhere and thats a lot of reading in the last year that a virus that was high path wouldnt discriminate and likely give a reverse baseline in summer against the peaks of winter and from that a conclusion could be drawn about number of cases ...pneumnonias in particular. A centerline. Interesting stuff to say the least. Also on ProMed today a blurb stating that sustainability in wild fowl for H5 isnt likely and that its the poultry that are infecting the WF. ???? Uh and thats based on what science?
I dont know that I buy that as there to me seems to be an endless loop of bird infections. If they are endemically infected wouldnt that mean that it will mutate in and out of high and low pathogenics. Birds dont live that long anyway so it has to stay infectious to transmit itself. One mutation in China now circles the globe or most of it. Its merging via reassortment or recombination with other viruses. The best question that I can get to here is/are there two or three really true strains or just one with a clade or two of difference?
Reason I ask is that the way I understand it was that we infected pigs with mild H1N1, then their little mixing bowls did something else with it and then reinfected humans with a high path version of the same stuff. That might not be correct but thats the way I read it. No one is sure of course but thats what I understand.
nice to see you highlighting this fine work by NIH group.
re. what an H1N1 season might mean, especially wrt emergence of H5N1.
it may be a good thing: first, just because last 3 decades have seen H3N2 seasons which have been more deadly than H1N1 seasons (and a less deadly interpandemic season is just a good thing!), and second, because a mild H_N1 infection (or H_N1 vax) could be good thing in event of a subsequent H5N1 epidemic.
see slide #28 in Rob Webster's DoD Geis talk at IOM on Dec 19 2006 at
http://www.iom.edu/CMS/3783/38416/38509.aspx
a little neuraminidase immune response could go a long way...
oops. I mean slide #24
As the CDC site says in its report for Dec. 29, "It is too early in the influenza season to determine which influenza viruses will predominate or how well the vaccine and circulating strains will match."
Influenza B has already closed some schools in S. Carolina (in Nov. 2006). Alabama has had localized A and B outbreaks.
We should know more in the next few months.
Thanks for the good summary of how flu deaths are calculated!
I'd like to see the figure or even better have the
datapoints in computer readable form.
Is it possible ?
I just counted the number of available HA-sequences at genbank:
1918 004,000
1919 001,000
1933 004,000
1934 008,000
1935 001,000
1940 001,000
1942 001,000
1943 005,000
1946 001,000
1947 005,000
1948 001,000
1950 001,000
1951 006,000
1952 001,000
1954 002,000
1955 001,000
1956 001,000
1957 005,000
1968 000,015
1969 000,005
1970 000,003
1971 000,007
1972 000,022
1973 000,009
1974 000,010
1975 000,006
1976 000,013
1977 007,012
1978 009,004
1979 001,003
1980 001,010
1981 000,006
1982 003,010
1983 033,014
1984 005,008
1985 001,039
1986 008,011
1987 006,023
1988 016,021
1989 007,049
1990 016,022
1991 024,045
1992 004,084
1993 000,186
1994 003,126
1995 041,138
1996 045,182
1997 014,132
1998 027,182
1999 038,260
2000 117,189
2001 161,071
2002 059,334
2003 053,414
2004 010,313
2005 050,206
2006 033,023
Revere, are you referring to the paper by Reichert et al (2004) in Am J Epidemiol?
Marissa: No, but that is the paper I was looking for in my files and couldn't find, as I think it has the data I wanted (thanks for the link). I read this when it came out but my office is so chaotic I couldn't find my copy and instead used the Archives article by Simonsen et al. because it has the cited table in it and I had it to hand. You will see almost the same figure in the Archives paper as in Fig. 5 of the one you linked to.
Reveres,
As a frequent reader of your blog, I very much appreciate your posts that clearly explain flu science--at least one of you is really, really good at that. And I often need things to be explained.
But this particular paper by Simonsen et al happens to be one I've spent some quality time with. So let me amplify a couple points in your very nice account of it.
1) Seasonal variation. Even if flu went extinct today, we'd still see a rise in baseline mortality during winter. We know that because in years when very little flu circulates, the baseline mortality rate still goes up by about 15% over the midsummer rate. People argue about why that happens--everything from other respiratory viruses to cold weather bringing on heart attacks to a deficit of vitamin D brought on by lack of sunlight is bandied about--but it ain't just flu.
2) Total flu mortality. Simonsen et al analyzed both excess P&I mortality and excess all-cause mortality. P&I is better for figuring out the relative severity of flu epidemics because it jumps by a bigger percentage when flu comes around. But analyzing all deaths is the best way to get the whole picture of flu mortality. Almost no deaths get recorded as being caused by "influenza" in a regular epidemic year. A lot more get recorded as "pneumonia." A hefty chunk of those really are due to flu, but not all--pneumonia is a very common way to go, especially for older people. But flu also causes people to pop off by other mechanisms, including heart attack and stroke. Those get recorded as the cause of death, even though had the person not caught the flu, they�d have lived on for years. So looking at excess all-cause mortality is the best way to get the sum total of the mortality burden.
3) The punch line (let's not leave THAT out!) The authors did their analysis so they could look at the TREND of flu mortality over time, from 1968 to 2001. That's why they separated out the H3N2 seasons (which have become more common in recent years) and why they adjusted for changes in age of the population as a whole (many more older people live in the US now than in 1968). They also focused on people older than 65 years, the group that suffers the vast majority of flu deaths in standard seasonal epidemics (that changes dramatically in pandemic years). They found that the rate fell from 1968 to about 1980. No surprise there; '68 was the first year for H3N2, and slowly the population gained immunity.
The trend from 1980 to 2001 was a little harder to figure. Without any adjustments, the trend was UP--more flu deaths were occurring. After adjustment, it looked basically flat, although there was a lot of noise in that signal. That result was an unpleasant surprise, because between 1980 and 2001 the immunization rate among the elderly increased four fold, from about 15% to about 65%. You'd think that would reduce flu mortality, especially because several routinely cited papers say that flu vaccine reduces mortality among the elderly by a whopping 50%. And that's not just "excess" (flu-related) mortality. That's ALL mortality, period. But the trend didn't show any gain at all despite a vast increase in coverage.
Simonsen et al also pointed out that, leaving the trend finding aside, it's a little crazy to say that flu shots could have such a super-power. Flu causes an average of about 5% (range 0-10%) of all deaths in any winter season. So how could the vaccine prevent 10 times as many deaths as the disease causes? Them was fighting words for some. Anyone interested should see the Archives letters at http://archinte.ama-assn.org/content/vol165/issue17/index.dtl under "Editor�s Correspondence."
Anyway, keep on with the explaining you do. It's very helpful.
Mr. Nobody: First, thanks for the kind words and also for the clarifications. There was quite a lot in that paper I chose not to talk about (as you note it was quite controversial because of the inferences about vaccine efficacy). My main uses of it were for the table of H3 and non-H3 years and as an excuse to discuss Serfling's method.
Regarding the latter, there is a latent assumption that flu drives P&I mortality (probably true) and that therefore without flu there wouldn't be significant P&I seasonality. That is a bit of a wobbly assumption, perhaps, but your statement that in non flu years there is still a rise in P&I mortality doesn't quite answer it because I think they assume flu is always there, just low in the winter and, left to itself, higher in the winter. The method is then used to estimate excess mortality above that "always there" flu. In truth, I don't think we have the data to decide this.
Similarly, the idea that a "hefty chunk" of non-P&I is really flu related is plausible to me (as a physician) but I'm not sure we have data for it (if you have a cite I'd love to have it, though).
What is most interesting to me about all this is how little we really know and the rather slender support of what we do know. These folks are really good at this kind of analysis, but there are limitations to what we can learn in the absence of better data. The disparity between the observational studies on vaccine efficacy (which are individually based studies) and the group based studies in this paper needs to be ironed out. It may be that one or the other is right, that there is some kid of aggrgation (ecologic) bias in the group studies, or that they are just asking different questions, one about individuals and one about groups of individuals and we should recognize that this paper isn't making strong statements about what should be recommended to someone over 65.
I decided not to get in to those things, although you are correct that the purpose in splitting out H3 and H1 (and age stratifying) was to make the data over time comparable to see if a vaccine effect was visible. Note that such times series comparisons are ecologic in nature and that under some circumstances both confounding and effect modification can lead to some severe biases. Adjusting for confounders in a regression doesn't work in that situation, unfortunately, although stratifying does, I believe.
Hope you all are still looking in on comments on aging posts.
Simonsen et al's paper was "controversial" because it questioned the CDC party line that flu shots are hugely effective in preventing flu-related mortality in seniors. I suppose the authors inability to find a downward trend in flu-related deaths concomitant with the huge increase in vaccine coverage could be due to ecologic bias. But we already vaccinate about two-thirds of people over 65 years. At some point you have to wonder whether the vaccine is doing its job.
But I just don't see what is controversial about their other point, that flu shots can't possibly cut ALL winter deaths among seniors by 50% because the burden is nowhere near that big. (For a recent recitation of this notion, see Nichol and Treanor, J Infect Dis. 2006 Nov 1;194 Suppl 2:S111-8).
You all are deep thinkers who aren't too bound up in party-think. So I put it to you: Do you think it's plausible that a flu shot can double the chance that an elderly person will live through the winter?
Mr. Nobody: I wasn't addressing that part of their paper, but I know the authors and think they are excelent scientists whose point is important and worth examining carefully. As someone in the aging population myself, the value of the vaccine to me personally is an important question, but it is a different one than the value of the vaccine to public health. And yes, I agree, a 50% decrease in mortality isn't likely given the numbers, although the extent of potential bias in an ecological design is quite surprising. I was skeptical of this until one of my students did his dissertation on the subject and the theoretical basis for this is not in doubt. It can reverse associations on an individual level under easily conceivable circumstances. The real question is whether such extreme biases actually happen in practice.
I hear you on the "ecologic" bit, and I agree that it's a pitfall to watch out for.
To close out our discussion, here's an idea you may want to think about: there's something really odd about the debate over flu vaccine efficacy, especially when it comes to mortality.
People argue about what proportion of flu-related deaths the vaccine can prevent among the over-65 population. CDC suggests it's up to 80% effective, others say it's less. The Simonsen et al trend analysis suggests it's a lot less. But as you say (and as they said)theirs is not a definitive result. The error bars on the trend were big. And there's the whole "ecologic" criticism. All that seems like a prefectly respectable scientific debate to me. At least it's about something sensible, namely how many flu-related deaths the flu vaccine can prevent.
Then there's this wild claim that the vaccine can prevent fully 50% of ALL winter deaths among relatively health seniors living large in their communities--and even more among institutionalized seniors. That seems crazy to me, implausible on its face. There's just no data to support the idea that flu is linked to anything like that proportion of deaths. And note that no "ecologic" trend arguement has to be invoked. It's a question of the size of the burden vs. the size of the purported effect.
Simonsen and colleagues pointed this out. But the same point could have been made years earlier, including in 2003 when Thompson et al came out with the widely cited "36,000" figure. Or in 1994, when Nichol and others started a long series of healthcare database studies that make the "50% of all deaths" claim. Or in the late 70s and 80s, when nursing home studies were suggesting the same thing.
But that claim is cited over and over. It causes no end of mischief, too: why would anyone need a better vaccine for flu if the old one is such a mircaluous life-saver?
I think a good part of the problem is that people don't distinguish a vaccine effect on "excess all-cause mortality" (that's the flu-related deaths) vs. one on "all-cause mortality" (that's ALL deaths).
Anyway, it's odd that the people can say both that "We need a better vaccine for the elderly, because their immune systems aren't as responsive" and "The current vaccine cuts a senior's chance of dying over the winter by half." Those just don't add up.
Until next time.
paper is here:
http://archinte.ama-assn.org/cgi/content/full/165/3/265
why vaccinate against H1N1 , when H3N2 is so much
more virulent ?
Better support H1N1 to suppress H3N2