Revised estimates of influenza-associated excess mortality, United States, 1995 through 2005
© Foppa and Hossain; licensee BioMed Central Ltd. 2008
Received: 05 March 2008
Accepted: 30 December 2008
Published: 30 December 2008
Excess mortality due to seasonal influenza is thought to be substantial. However, influenza may often not be recognized as cause of death. Imputation methods are therefore required to assess the public health impact of influenza. The purpose of this study was to obtain estimates of monthly excess mortality due to influenza that are based on an epidemiologically meaningful model.
Methods and Results
U.S. monthly all-cause mortality, 1995 through 2005, was hierarchically modeled as Poisson variable with a mean that linearly depends both on seasonal covariates and on influenza-certified mortality. It also allowed for overdispersion to account for extra variation that is not captured by the Poisson error. The coefficient associated with influenza-certified mortality was interpreted as ratio of total influenza mortality to influenza-certified mortality. Separate models were fitted for four age categories (<18, 18–49, 50–64, 65+). Bayesian parameter estimation was performed using Markov Chain Monte Carlo methods. For the eleven year study period, a total of 260,814 (95% CI: 201,011–290,556) deaths was attributed to influenza, corresponding to an annual average of 23,710, or 0.91% of all deaths.
Annual estimates for influenza mortality were highly variable from year to year, but they were systematically lower than previously published estimates. The excellent fit of our model with the data suggest validity of our estimates.
To estimate excess mortality due to influenza, two fundamental approaches have previously been used. The most popular one is based on Serfling's seasonal regression method  and has resulted in numerous estimates of excess mortality due influenza [3–8, 12, 14, 15, 19]. This periodical regression approach is based on parametric estimation of a sinusoidal "baseline" function that represents mortality in absence of influenza. The difference between the baseline function and the observed numbers of deaths is then interpreted as the number of excess deaths due to influenza. Typically, the baseline mortality function is fitted to weekly or monthly mortality rates or numbers during non-influenza months, using two or more Fourier terms . This approach is intuitively appealing as it captures the strong seasonal periodicity of mortality. However, the particular choice of a parametric baseline function lacks epidemiological justification: Why should the baseline function be sinusoidal rather than of any other periodic form? Depending on the shape of the "true" baseline function, under- or overestimation of excess mortality due to influenza might result. If, for example, the true baseline function is "higher" (i.e. the definite integral of the true function is larger) than the assumed sinusoidal function, then overestimation would result and vice versa. Another, potentially more important shortcoming of the periodical regression approach lies in the fact that seasonally correlated causes of mortality, including influenza, are not controlled for, which might lead to confounded estimates of excess mortality.
To avoid this difficulty, one could gauge all-cause mortality with some independent measure of influenza transmission (or mortality). Following this rationale, Thompson et al.  estimated excess mortality due to both influenza and respiratory syncytial virus (RSV). They used a generalized linear model (GLM) with a Poisson distribution and a logarithmic link function to model the weekly number of deaths. They also used two Fourier terms in their model, but, in addition, used indicators of influenza and RSV transmission. These indicator variables were defined by the proportions of specimens testing positive for influenza A(H1N1), influenza A(H3N2), influenza B and RSV. Several potential shortcomings of this methods are, however, apparent. First, this model also makes a priori assumptions about the baseline mortality function–in this case an exponentiated sinusoidal function. Although this might conceivably be true, there is little empirical evidence to support this assumption. Second, the multiplicative form of the model implies that excess mortality, given a certain amount of influenza activity, depends on the current level of all-cause mortality. Again, this does not appear to be a well-founded assumption. Finally, the proportion of test positive specimens is likely to be a poor measure of excess mortality. While a high proportion of test positive specimens is compatible with high levels of influenza transmission (and excess mortality), this is not necessarily true. The model, however, implies that five hundred influenza positives, obtained from a thousand tests, are associated with less excess mortality than two influenza positives, obtained from three tests. This appears to be an unrealistic assumption. The seasonally changing frequency of influenza testing  is, at least partly, due to the seasonally changing incidence of other agents causing influenza-like illness (ILL).
Alternatively, one could postulate that mortality directly attributed to influenza (influenza-certified mortality) represents a certain proportion of all mortality attributable to influenza. This assumption implies that the coefficient associated with influenza-certified mortality represents the ratio of total influenza mortality to influenza certified mortality [17, 21]. Here we use a method for the estimation of influenza excess mortality which is similar to the one recently presented by Schanzer and colleagues : we adopt the proportionality assumption and avoid specific parametric assumptions about the baseline function. In addition, and deviating from the Schanzer model, we allow for random variability of influenza-certified mortality by adopting a hierarchical modeling approach. We present the resulting estimates of U.S. excess mortality due to influenza for the years 1995 to 2005. We compare these to estimates obtained from a Thompson-like model , as well as to previously published estimates of influenza-associated excess mortality.
We used Multiple Cause-of-Death Data for the years 1995 to 2005 (Multiple Cause-of-Death Microdata, 1995–2005, National Center for Health Statistics, Hyattsville, Maryland). This dataset is in the public domain and can be electronically downloaded from the web site of the National Bureau of Economic Research http://www.nber.org/data/vital-statistics-mortality-data-multiple-cause-of-death.html. We defined deaths as influenza-certified if influenza was given as underlying cause of death. The corresponding diagnostic code for ICD-9 (1995 to 1998) was 487 and and for ICD-10 (1999 onwards) the code range was J10–J12. Influenza years were defined as lasting from July 1 of one year to June 30 of the following year. We defined four age categories: < 18, 18–49, 50–64 and 65+. Observations with missing age (N = 4,490) were not included in this analysis.
To estimate excess mortality due to influenza, is calculated using expression 5, with posterior estimates of γ i and ϕ. As total influenza mortality cannot be lower than influenza-certified mortality, the minimum value for the range in the prior distribution for ϕ was set to one (see additional file 1).
To calculate estimated excess mortality due to influenza, all parameters in 8 are replaced by their posterior estimates
The parameters for this hierarchical model were estimated using a Markov chain Monte Carlo (MCMC) algorithm implemented in WinBUGS, version 1.4.1 (Imperial College and Medical Research Council, UK) . Uninformative prior distributions were used (additional file 1). To ensure positivity of all θ i , the normal priors of this parameter were truncated at non-positive values (additional file 1). The empirical posterior distributions of the parameters were obtained from MCMC samples of 30,000, resulting from three chains with 200,000 burn-in iterations and 10,000 samples each. Posterior means and 95% credible intervals (CIs) were calculated for all parameters of interest after ensuring convergence of all model parameters.
The parameters of the TL model could easily be estimated using a GLM procedure in any standard statistical software package. However, to allow for direct comparison of the model fit we used the same estimation procedure as for the current model. The fit of the two age-specific models was compared using the deviance information criterion (DIC) . DIC penalizes the model goodness-of-fit for additional complexity. The complexity is measured by the effective number of parameters.
where E(·) and V(·) are the operators for the posterior mean and empirical variance, respectively and e i = Y i - μ i . The empirical variance of e is computed for each iteration.
Age category-specific estimates for the detection ratio ϕ.
Posterior Mean (95% CI)
Estimated Numbers of Deaths Attributable to Influenza, United States, 1995–2005, according the the current model.
Posterior Mean (95% CI)
Comparison of the age-specific fit (DIC) of the current with the TL model.
Our estimates of excess mortality due to influenza are substantial, especially for the influenza years 1997/98, 1999/2000 and 2003/04, during which influenza A(H3N2) predominated. The lowest estimates were obtained for the years 2000/01 and 2002/03, when influenza A(H1N1) and B viruses predominated. Nevertheless, our estimates are markedly lower than previous estimates. For the year 1995/96, for example, we attributed 12,067 excess deaths to influenza. For the same period, Simonsen et al.  estimated that 25,071 deaths were attributable to influenza in ages 65+ alone. Thompson et al.  estimated the number of excess deaths during that influenza year at 36,280–more than three times our estimate. The obvious question arises, which of these estimates are closest to the true excess mortality? As pointed out above, the method of Simonsen et al.  is problematic for two reasons. First, it does not account for temporal correlation between baseline mortality and influenza excess mortality. The resulting estimates of influenza excess mortality may therefore be confounded. Second, their model makes a priori assumptions about the parametric shape of the baseline function; these assumptions may or may not be true. They should, in any event, be validated. The Thompson model , which superficially resembles a hybrid between the Simonsen model and the model proposed by Schanzer et al.  (or the current model), addresses the issue of temporal confounding by controlling for the proportion of influenza test positives. As pointed out in the Background section, the use of that specific variable to control for influenza mortality may not be appropriate. We compared estimates from the TL model with estimates from the current model. The TL model is based on the Thompson model, but influenza-certified mortality is substituted for proportion positives. Although the resulting estimates were about a sixth higher than our estimates, the seasonal pattern was highly consistent with the pattern seen with the current model. This consistency implies relative robustness of excess deaths estimates to the choice of a specific baseline function. The vast difference between our and Thompson's estimates  can therefore not be explained by differences in model structure, nor in the way the baseline function is modeled. They may rather be due to the use of proportion of specimens testing positive to control for influenza mortality.
Schanzer et al. , like us, used a Poisson model with linear (rather than logarithmic) link function, to analyze weekly mortality data from Canada. Modeling weekly mortality has the advantage of giving higher temporal resolution to the analysis. On the other hand, deaths associated with, but not attributed to influenza may occur with some delay and may thus be partially decoupled from influenza-certified mortality. However, Schanzer and colleagues did not find an obvious lag between weekly influenza-certified mortality and mortality due to other causes. Future studies will be needed to determine what level of temporal aggregation results in the best estimates.
To take into account random variability in influenza-certified mortality, we used a hierarchical model. While the point estimates for ϕ (corresponding to β3 in ) obtained from a GLM are very similar to the ones obtained from the hierarchical model (21.35 and 21.16, respectively, for 65+), the confidence limits are much wider for the latter (95% credible interval 18.06, 24.32 vs. Wald 95% confidence interval 20.91, 21.80). This may even be more pronounced for weekly data, where numbers of influenza-certified deaths are often quite small. To the extent that our hierarchical model takes into account random variability of influenza-certified deaths and thus leads to wider confidence limits around the resulting excess mortality estimates, it is more conservative than non-hierarchical GLM models.
Previous estimates of excess mortality due to influenza may be biased and inflated. We propose a two-level hierarchical Poisson model where the baseline mortality varies with time. The goodness-of-fit statistic indicates that this model fits the data very well, explaining well above 90% of the observed variation of all-cause mortality during the eleven years study period. The resulting estimates are therefore likely of high validity. Future attempts to quantify the public health burden of influenza should also explore demographic approaches that take into account life expectancy.
MMH was partly supported by the National Institutes of Health (NIH grant # 1 R03 CA125828-01). We thank anonymous reviewers whose criticism helped to substantially improve this paper. We would also like to thank Drs. Eric Brenner, MD, and Robert T. Ball, MD, MPH for their insight into the accuracy of the death certificate diagnosis "influenza".
- Zambon MC: Epidemiology and pathogenesis of influenza. J Antimicrob Chemother. 1999, 44 (Suppl B): 3-9. 10.1093/jac/44.suppl_2.3View ArticlePubMedGoogle Scholar
- Collins SD: Excess Mortality from Causes other than Influenza and Pneumonia during Influenza Epidemics. Public Health Reports. 1932, 47 (46): 2159-79.View ArticleGoogle Scholar
- Eickhoff TC, Sherman IL, Serfling RE: Observations on excess mortality associated with epidemic influenza. Jama. 1961, 176: 776-82.View ArticlePubMedGoogle Scholar
- Housworth J, Langmuir AD: Excess mortality from epidemic influenza, 1957–1966. Am J Epidemiol. 1974, 100: 40-8.PubMedGoogle Scholar
- Choi K, Thacker SB: An evaluation of influenza mortality surveillance, 1962–1979. II. Percentage of pneumonia and influenza deaths as an indicator of influenza activity. Am J Epidemiol. 1981, 113 (3): 227-35.PubMedGoogle Scholar
- Choi K, Thacker SB: Mortality during influenza epidemics in the United States, 1967–1978. Am J Public Health. 1982, 72 (11): 1280-3. 10.2105/AJPH.72.11.1280PubMed CentralView ArticlePubMedGoogle Scholar
- Lui KJ, Kendal AP: Impact of influenza epidemics on mortality in the United States from October 1972 to May 1985. Am J Public Health. 1987, 77 (6): 712-6. 10.2105/AJPH.77.6.712PubMed CentralView ArticlePubMedGoogle Scholar
- Simonsen L, Clarke MJ, Williamson GD, Stroup DF, Arden NH, Schonberger LB: The impact of influenza epidemics on mortality: introducing a severity index. Am J Public Health. 1997, 87 (12): 1944-50. 10.2105/AJPH.87.12.1944PubMed CentralView ArticlePubMedGoogle Scholar
- Fleming DM: The contribution of influenza to combined acute respiratory infections, hospital admissions, and deaths in winter. Commun Dis Public Health. 2000, 3: 32-8.PubMedGoogle Scholar
- Donaldson GC, Keatinge WR: Excess winter mortality: influenza or cold stress? Observational study. Bmj. 2002, 324 (7329): 89-90.PubMed CentralView ArticlePubMedGoogle Scholar
- Thompson WW, Shay DK, Weintraub E, Brammer L, Cox N, Anderson LJ, Fukuda K: Mortality associated with influenza and respiratory syncytial virus in the United States. Jama. 2003, 289 (2): 179-86. 10.1001/jama.289.2.179View ArticlePubMedGoogle Scholar
- Reichert TA, Simonsen L, Sharma A, Pardo SA, Fedson DS, Miller MA: Influenza and the winter increase in mortality in the United States, 1959–1999. Am J Epidemiol. 2004, 160 (5): 492-502. 10.1093/aje/kwh227View ArticlePubMedGoogle Scholar
- Uphoff H, Stilianakis NI: Influenza-associated excess mortality from monthly total mortality data for Germany from 1947 to 2000. Methods Inf Med. 2004, 43 (5): 486-92.PubMedGoogle Scholar
- Simonsen L, Reichert TA, Viboud C, Blackwelder WC, Taylor RJ, Miller MA: Impact of influenza vaccination on seasonal mortality in the US elderly population. Arch Intern Med. 2005, 165 (3): 265-72. 10.1001/archinte.165.3.265View ArticlePubMedGoogle Scholar
- Zucs P, Buchholz U, Haas W, Uphoff H: Influenza associated excess mortality in Germany, 1985–2001. Emerg Themes Epidemiol. 2005, 2: 6. 10.1186/1742-7622-2-6PubMed CentralView ArticlePubMedGoogle Scholar
- Dushoff J, Plotkin JB, Viboud C, Earn DJ, Simonsen L: Mortality due to influenza in the United States-an annualized regression approach using multiple-cause mortality data. Am J Epidemiol. 2006, 163 (2): 181-7. 10.1093/aje/kwj024View ArticlePubMedGoogle Scholar
- Schanzer DL, Tam TW, Langley JM, Winchester BT: Influenza-attributable deaths, Canada 1990–1999. Epidemiol Infect. 2007, 135 (7): 1109-16. 10.1017/S0950268807007923PubMed CentralView ArticlePubMedGoogle Scholar
- Serfling RE: Methods For Current Statistical-Analysis of Excess Pneumonia-Influenza Deaths. Public Health Reports. 1963, 78 (6): 494-506.PubMed CentralView ArticlePubMedGoogle Scholar
- Stroup DF, Thacker SB, Herndon JL: Application of multiple time series analysis to the estimation of pneumonia and influenza mortality by age 1962–1983. Stat Med. 1988, 7 (10): 1045-59. 10.1002/sim.4780071006View ArticlePubMedGoogle Scholar
- Anonymous: Seasonal Flu – Flu Activity & Surveillance. 2008.Google Scholar
- Gay NJ, Andrews NJ, Trotter CL, Edmunds WJ: Estimating deaths due to influenza and respiratory syncytial virus. Jama. 2003, 289 (19): 2499-author reply 2500-2 10.1001/jama.289.19.2499-aPubMedGoogle Scholar
- Spiegelhalter D, Thomas A, Best N, Lunn D: WinBUGS User Manual. 2003Google Scholar
- Spiegelhalter D, Best N, Carlin B, Linde A: Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society, Series B. 2002, 64: 583-639. 10.1111/1467-9868.00353. 10.1111/1467-9868.00353View ArticleGoogle Scholar
- Gelman A, Pardoe L: Bayesian measures of explained variance and pooling in multilevel (hierarchical) models. Technometrics. 2006, 48 (2): 241-251. 10.1198/004017005000000517. [044CW Times Cited:1 Cited References Count:34]. 10.1198/004017005000000517View ArticleGoogle Scholar
- Michaud CM, McKenna MT, Begg S, Tomijima N, Majmudar M, Bulzacchelli MT, Ebrahim S, Ezzati M, Salomon JA, Kreiser JG, Hogan M, Murray CJ: The burden of disease and injury in the United States 1996. Popul Health Metr. 2006, 4: 11. 10.1186/1478-7954-4-11PubMed CentralView ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.