Natural attack rate of influenza in unvaccinated children and adults: a meta-regression analysis

Infectious diseases are a major cause of morbidity and mortality worldwide [1]. The main focus for research on infectious diseases is to improve disease control and ultimately eradicate the infection from the population [2]. Mathematical models of infectious diseases are important for studying the impact of various interventions on disease outcome. Such investigations allow the implementation of targeted control measures, optimizing the use of limited resources [2].

The validity of a model is highly dependent on the validity of the input parameters. These models require information on parameters describing the population under study, disease transmission as well as parameters describing the health outcomes of infection. Lack of reliable data complicates the development of models and new data needs to be collected or analyzed or the model itself needs to be fitted to available data to estimate input parameters [3].

Influenza is an infectious disease that poses a major health concern all over the world. The total economic burden of influenza in the United States has been estimated as $87.1 billion [4]. Developing mathematical models for influenza can be used to determine how to mitigate disease. The attack rate of the influenza virus is crucial for these models and a difficult parameter to obtain.

Generally, influenza epidemic infection rates are obtained from influenza-like illness (ILI), surveillance data and serological surveys [5]. In a static (e.g. decision tree) model, the attack rate is often used directly as an input parameter, whereas in a dynamic transmission model, it is used to calibrate, or calculate directly, the force of infection. Whether static or dynamic, it is most preferable to use the “natural attack rate”, that is, the attack rate within the unvaccinated portion of the population. This is a more “universal” quantity, since it is not as strongly dependent on the population vaccine uptake, though it is still coupled through herd immunity.

Turner et al. used age-specific attack rates of influenza in an economic evaluation based on rates derived from placebo-controlled trials of influenza vaccines and antivirals [6]. The incidence of symptomatic laboratory-confirmed influenza cases in each placebo arm was used to estimate the corresponding attack rates for each subgroup using a random effects model. Overall, the attack rate was 6.55% (range: 0 to 20%) for adults and 19.21% (range: 10 to 35%) for children. All trials except one were in the US and spanned eight influenza epidemics during the period 1984 through 1997-1998. These attack rates were not adjusted for confounders and also were not stratified by strain of influenza.

The objective of this study was to estimate the attack rate of influenza stratified by both age and influenza strain (A and B) for unvaccinated individuals measured by laboratory-confirmed PCR (Polymerase Chain Reaction) and/or culture.

This study (GSK study identifier: 117102) is a post-hoc analysis of a published systematic review [7]. The full methods have been published [7],[8] and the protocol was registered in the PROSPERO database (CRD42012001926).

The systematic review informing this analysis yielded 1356 citations [7]. A total of 34 RCTs and 1 companion report fulfilled the inclusion criteria (Figure 1). These 34 trials contributed to 47 influenza seasons from 1970 to 2009 including 94,821 participants. Reasons for exclusion, study characteristics, and patient characteristics are reported in detail elsewhere [7]. The trials were conducted in North America (22/34), Europe (4/34), Australia (1/34) and a mixture of multi-site trials in North America, Europe, South America, Asia (7/34).

Meta-regression analysis was conducted to reduce the heterogeneity associated with the attack rate calculations. Age was considered an important covariate in natural attack rate. In order to confirm this with our data, a meta-regression analysis was conducted with age stratifications. With the available data from the studies from the systematic review, we were only able to separate ages into two categories: adults and children. Limited number of trials reported on the elderly population alone and most of them did not report the primary outcome. It was difficult to use geography as a covariate since many trials were multi-national in nature. Due to this limitation, we did not include geography as a covariate in the meta-regression.

The forest plot for influenza A infections in children and adults for the primary outcome can be seen in Figure 2A and B respectively. The forest plot for influenza B for children and adults can be seen in Figure 3A and B respectively. Generally, no time trends were observed in the attack rate for adults or children for each type of influenza.

This study estimated the natural attack rate of influenza for unvaccinated children and adults using a post-hoc analysis of a previously conducted systematic review. This is the best method to obtain such attack rates since currently existing methods to obtain this data have a number of limitations. First, estimating attack rates from surveillance data provides attack rates as an aggregate of vaccinated and unvaccinated individuals and not attack rates that are specific to unvaccinated individuals that are presented here. Second, extrapolating from surveillance data also necessitates estimating rates of primary care consultations among influenza cases [5]. Third, even if there is an active surveillance system looking for influenza, surveillance systems will be limited by under-reporting bias [11]. Finally, it’s difficult to determine the size of the population under surveillance which makes estimating rates of infection difficult as there is uncertainty around the denominator of the ratio [11].

Although yearly influenza epidemics can seriously affect all age groups, children younger than two years of age and adults over 65 exhibit higher rates of disease [12]. Further, contact pattern studies suggest that children are more vulnerable to infection due to their relatively more frequent and longer intimate contacts with members of their age group [13]. Therefore, age is believed to be important in the natural attack rate of influenza. Geography dictates contact patterns and seasonality which is known to be a driver in severity of influenza [14]. Outbreaks of seasonal influenza follow largely predictable seasonal patterns and in temperate climate zones seasonal influenza epidemic lasts from six to ten weeks and in tropical zones seasonal patterns are less pronounced with more than one peak of infection [15]. Thus, geography is also an important consideration with regards to natural attack rate. With respect to the two main types of influenza in circulation, influenza A contributes to greater annual epidemics and infrequent yet more-devastating pandemics than influenza B [16]. Therefore, it’s important to distinguish between these two types of influenza types. As the sensitivity of various laboratory methods for detecting influenza can yield different results, this was also included as an important covariate. Since yearly differences in influenza are seen from season to season, the year of study conduct was also considered an important covariate. Therefore, age group, geographical region, type of laboratory-confirmed influenza detected, type of test used to confirm detection and the year of influenza season were all considered important covariates of natural attack rate in our study.

Meta-regression analyses conducted here showed that age group and type of influenza are important covariates in natural attack rates. For adults, an attack rate of 3.50% (2.30%, 4.60%) was observed for all influenza strains. When stratified by type, an attack rate of 2.32% (1.47%, 3.17%) was observed for influenza A and a rate of 0.59% (0.28%, 0.91%) was observed for strain B. For children, the overall attack rate was 15.20% (11.40%, 18.90%) and when stratified, a rate of 12.27% (8.56%, 15.97%) was seen for strain A and 5.50% (3.49%, 7.51%) was seen for strain B. The weighted sum of the stratified rates may not add up to the aggregate rates as some studies reported results for both strains of influenza. Including HAI testing increased the attack rate for all age groups and all types of influenza. No general trends with time were observed for children and adults for the primary outcome.

This study provides new, up to-date estimates of the attack rate of influenza in children and adults separated by type of influenza. Generally, our overall attack rates for adults and children were lower than the attack rates observed by Turner et al. [6]. The Turner analysis included trials covering only 8 influenza seasons from 1984 to 1998 while our analysis covered 47 influenza seasons from 1970 to 2009. The primary outcome of our studies was PCR and/or culture but the studies used in the Turner analysis did not include PCR as a method of detection. PCR is considered the most reliable diagnostic test for clinical practice with a higher accuracy of detection [17]. Using culture as a method of detection may miss influenza cases and the use of HAI testing may lead to biased results [17]. This may explain why the attack rates observed in the Turner analysis are higher than the results we have obtained (6.55% for adults versus 3.57% and 19.21% for children versus 14.22%). Our secondary outcome, which included HAI testing, also showed an increase in attack rates for all age groups and types of influenza which supports the notion that higher attack rates are observed when HAI testing is used as a method of detection. Furthermore, the definition of children used by our study and by the Turner study is quite different. In our study, children were classified as <18 years while the Turner classified them as <12 years. This may also attribute to the differences seen in attack rates when comparing between studies. Influenza B is most prominent among older children and young adults [18]. In line with this, our attack rates for children exhibited a higher B:A ratio than for adults. The B:A attack rate ratio for children was calculated to be 0.45 while the ratio for adults was 0.25. This further exemplifies that using this method to calculate the natural attack rate does in fact reflect what is observed in real life settings for seasonal influenza [18]. The strengths of our analysis over the Turner analysis include the use of many trials covering more influenza seasons, using a more accurate method of detection (RT-PCR (Reverse Transcription Polymerase Chain Reaction) and/or culture), adjusting for confounders that affect the natural attack rate and providing strain-specific natural attack rates for influenza.

Like any meta-analysis, our study has several limitations. First, although we identified three RCTs reporting data in the senior population, we were not able to derive an attack rate for this population. Two out of the three studies did not report on our primary outcome of interest and the third study reported an attack rate of 0%. Second, we faced an inability to calculate geography-specific attack rates. Although it was decided a priori that geography is an important covariate, the data did not support such an analysis. Although most of the trials were conducted in North America, newer trials were multi-centre in nature spanning over more than one continent, making a geography-specific analysis difficult. Third, the attack rates that are calculated here are attenuated by community herd effects through vaccinated people in the population. For countries with a low background vaccination rate, natural attack rates might be higher than what’s reported here. As most of the studies in this analysis were conducted in North America, the attack rates obtained here would reflect the effects of herd immunity observe through targeted vaccine coverage programs that are used in this region. Furthermore, the attack rates observed in this study are laboratory confirmed. A study by Glatman-Freedman et al. have shown that in general, laboratory confirmed attack rates are higher than clinical attack rates which is an important consideration when using these rates as model inputs [19]. Another limitation of this study is that H1N1 and H3N2 strains were agglomerated as strain A for analysis. The attack rates for these two types for strain A are different from each other and thus lead to different health outcomes. Also, the severity of the influenza strains are known to be in the order of H3N2 > B > H1N1 [18]. Therefore, one must be cautious when applying the agglomerated attack rate presented here for strain A to reflect H3N2 or H1N1 specifically. Although we identified one study that focused on the H1N1 pandemic flu, Talaat et al. [20], it reported on adults and elderly population together and therefore was not included in the meta-regression of attack rate for adults. Lastly, as this study is based on the results of an already conducted systematic review, it is inherently limited by the results of the systematic review. The systematic review only focused on RCTs of vaccines and not antivirals (as was done for the Turner et al. analysis). If the focus of the systematic review was expanded to include antivirals, then a large sample size would have been available for the attack rate calculations and perhaps lead to reduced uncertainty of the estimates that were derived from the analysis.

Influenza models are important tools in public health decision-making, and natural attack rate of influenza is a vital ingredient for such models, as either a direct input parameter or a calibration target. It is imperative that robust data are used to inform these models so that they can better predict disease transmission and also resource allocation that is required to reduce the burden of illness. The sophistication of mathematical models of disease transmission has increased in the last decade [21],[22]; accordingly, the need for valid, precise and accurate model inputs, such as natural attack rate, is greater than ever. The over- or under-estimation of these parameters can have significant impact on estimating the burden of illness and can introduce uncertainty in public health decision making.

In this study, we were able to determine important covariates for natural attack rate of influenza and calculate age-specific and strain-specific attack rates that would serve as essential input parameters to an infectious disease model for influenza. We believe that the use of robust and accurate input parameters are key to reducing uncertainty associated with infectious disease models and therefore would make them more appealing in public health decision making settings.