Mortality Associated With Influenza and Respiratory Syncytial Virus in the US, 1999-2018

Key Points Question What was the excess mortality from respiratory syncytial virus (RSV) and influenza in the US from 1999 to 2018? Findings This cross-sectional study estimates a mean of 6549 underlying respiratory deaths associated with RSV each year (range, 5035-7645) and estimates a mean of 10 171 underlying respiratory deaths associated with influenza per year (range, 393 to 23 176), with greater interannual variation for influenza than for RSV. The highest mortality for both viruses was among individuals aged 65 years or older; RSV mortality was 5-fold higher than influenza mortality among children younger than 1 year. Meaning This study suggests that, despite changes in epidemiology, endemic respiratory viruses continue to have a significant death toll in the US, especially among infants and elderly individuals.


eMethods.
Here we provide additional details on the excess mortality regression models used to estimate the burden of influenza and RSV. We fit linear regression models to each death outcome, week, age, and location (national or by region), following:

mr_(t:c,a) = β0 + ns(t) + ∑ 19 =1 1,s*L1(flu)(t) + β2*L2(RSV)(t)
where mr_(t; c,a) represents the five-week moving average of the mortality rate per 100,000 population for cause c, age group a, and week t; ns(t) is a natural cubic spline with 60 degrees of freedom representing a smooth function of time for seasonality in mortality that is not attributed to influenza or RSV. Splines have been used to model seasonally-varying excess mortality as an alternative to more conventional harmonic models as they allow for greater flexibility in baseline mortality. 1,2,3,4 We tested 2 and 3 degrees of freedom per year and used AIC to select the best model. 3,4 Flu(t) and RSV(t) are the weekly influenza and RSV circulation proxies (see main text for details); terms s1-s19 represent the severity of different influenza seasons (akin to the case fatality rate, ie the ratio of influenza cases to influenza deaths which may vary depending on the mix of strains each year). We considered different lags between viral activity and mortality, as detailed below. We also ran multiple sensitivity analyses considering different mortality outcomes and proxies of viral circulation, as described below. To account for autocorrelation in the model we used residual bootstrapping to calculate the confidence intervals. We randomly resampled the residuals from the model and added them to the predicted values using the 'car' package in R. 5 We reran the model on these new y values and repeated this process 1000 times. We used the 2.5 and 97.5 percentiles as our upper and lower confidence intervals. We chose our 5 age groups for comparison with the seminal study by Thompson et al 6 and our geographic aggregations by Health and Human Services region were based on the availability of RSV surveillance data.

Model calibration (lag between mortality and viral circulation)
We tested lags between viral circulation (both RSV and influenza) and mortality, using AIC to select the optimal lag for each age group. Prior work allowed for a lag of 0-3 weeks between viral circulation and mortality. 1,2,6,7 Because we used a 5-week moving average in our mortality data, we allowed for a lag of up to 8 weeks. We did this only for underlying respiratory mortality and used the same lag for the other causes of death. This resulted in 81 models for each age group.
Below we show the lowest 10 AICs for each age group. Adjusting the lag between RSV circulation and mortality had little impact on model fit for those 5-49 years, supporting a low association between mortality and RSV circulation in this age group. There was no consensus as to the optimal lag across age groups, so we allowed each age group to have its own lag between viral circulation and mortality in the final models.

Influenza type/subtype analysis
We obtained weekly regional data on influenza testing and influenza like illness using the 'cdcfluview' package in R. 8 In main analyses of weekly mortality data, we use all influenza subtypes combined as a proxy for influenza incidence, but we also run a seasonal-level analysis that considers subtype-specific circulation. We applied the H1 and H3 proportions from the subtyped data to the unsubtyped influenza A samples and constructed subtype-specific influenza time series at the weekly and seasonal level. To estimate excess influenza mortality by virus subtype we regressed seasonal estimates of excess influenza mortality against the percentage positive for each subtype during that season with the intercept set to zero. 9 Excess mortality for influenza type/subtypes each season was estimated as the type/subtype coefficient multiplied by type/subtype covariate (proportion positive*ILI) for each respiratory season.
Further, to compare our results with other papers, we also ran sensitivity analyses where we model weekly mortality as a function of weekly subtype-specific incidence proxies for influenza.

Sensitivity analyses
We ran several sensitivity analyses to test model assumptions. To examine the impact of increased PCR testing for RSV over time, we ran a version of the final model which only included the proportion positive based on antigen tests from 2010 to 2018. To determine if regional level mortality counts were robust enough to present meaningful estimates, we resampled the national data using a binomial distribution to simulate expected counts in the largest (Region 4, ~20%) and smallest (Regions 8 and 10, ~4%) HHS Regions for the two extreme age groups (<1 and 65+). We fit the model to this simulated data and compared with estimates from the original models. We found that down-sampled estimates for those <1 year were not stable, but estimates in those 65+ years were robust, suggesting regional analyses in those 65+ were valid. We used the Shapiro-Wilk test to assess if our regional-level data were appropriate for ANOVA. We found that our regional data did not violate the assumption of normality but did not have constant variance. As a sensitivity analysis, we ran a non-parametric Kruskal-Wallis test. The Kruskal-Wallis test supported our conclusions that there were no statistically significant differences in influenza mortality between regions, but there were significant differences in RSV mortality.
Similarly, the Kruskal-Wallis test supported our conclusion that excess UR RSV mortality was highest in Region 6 and lowest in Region 10.

Missing data
Where death certificates were missing the exact date of death (0.003%), the fifteenth of the month was used. Death certificates where age of deceased was missing were excluded from analysis (0.01%). Influenza surveillance for Region 10 was missing prior to September 21, 2008. We used neighboring Region 8 which had the highest correlation after 2008 to fill in the missing weeks.

Comparing influenza estimates with other studies
To compare our influenza estimates with other studies we tested the impact of adjustments to our model on the estimated average annual mortality rates. To simplify these comparisons, we aggregated all age groups and converted everything to rates per 100,000 population. When other studies provided estimates by year, we compared averages only for overlapping years. For comparing to studies estimating the impact of the 2009 pandemic, we calculated the rate for the same months covered in those analyses.
Our comparison analyses include the following models: (a) Our final model as described in the main text A model similar to the above but with all ages aggregated rather than summing results from age-specific models (c) A model with sine/cosine terms instead of a spline term to model baseline seasonality (e) A model which has individual terms for each influenza type/subtype We used the same causes of death described in each paper. For estimates produced with the multiplier method we compared with all-cause mortality unless the paper specifically indicated that more specific causes of hospitalization and death were considered. All models use underlying cause of death except model (f) which includes contributing causes of death. These results are described in eTable 7 and summarized in the discussion of the main text.