Contents lists available at SciVerse ScienceDirect
Vaccine
j our nal homep ag e: www.elsev ier .com/ locate /vacc ine
rief report
ariable definitions of the influenza season and their impact onaccine effectiveness estimates
heena G. Sullivana,∗, Ee Laine Tayb,c, Heath Kellyb,c
WHO Collaborating Centre for Reference and Research on Influenza, Melbourne, AustraliaVictorian Infectious Diseases Reference Laboratory, Melbourne, AustraliaAustralian National University, Canberra, Australia
a r t i c l e i n f o
rticle history:eceived 12 March 2013eceived in revised form 20 June 2013ccepted 27 June 2013vailable online 9 July 2013
a b s t r a c t
Vaccine effectiveness (VE) studies are often made for a “season” which may refer to different analysisperiods in different systems. We examined whether the use of four different definitions of season wouldmaterially affect estimates of influenza VE using data from the Victorian general practice sentinel surveil-lance network for 2007–2012. In general, the choice of analysis period had little effect on VE estimates(≤five percentage points) when there was a statistically significant protective effect of vaccination (2007,
2010 and 2012). In contrast, for years when the analysis period varied widely depending on the methodused and when VE estimates were imprecise, the change in VE estimate was as much as 43 percentagepoints (2008). Studies of influenza VE should clearly define the analysis period used and, where possible,provide sensitivity analyses to align this definition with other VE studies.
Studies of influenza vaccine effectiveness (VE) are typicallyeported for a “season.” However, the analysis period used variescross studies and multiple methods are used to define a season. Inermany, for example, data are analysed for the season defined byeeks when >10% of sentinel specimens were influenza-positive
1] whereas the US Flu VE Network uses a more complicatedlgorithm (when the percentage of positive specimens “began toncrease in the community or on the week of 17 January 2011,
hichever came first, and ended after either 12 weeks of surveil-ance or 2 weeks without cases”) [2]. In Canada, VE estimates are
ade for the influenza surveillance period used by local healthuthorities with dates given for that period [3,4]. Conversely,ooled analyses from Europe limit the analysis to the intervaletween the week of symptom onset of the first case or 14 days afterhe commencement of the vaccine campaign and the week after theast case of influenza, separately for each country and type/subtype
5].
The definition of season may be important because an estimatef influenza VE should be made during periods when the virus is
∗ Corresponding author at: WHO Collaborating Centre for Reference and Researchn Influenza, 10 Wreckyn Street, North Melbourne, VIC 3051, Australia.el.: +61 3 9342 3917; fax: +61 3 9342 3939.
circulating and the vaccine can potentially provide protection [6].Indeed, estimation of VE outside the designated influenza seasonhas been used to demonstrate residual confounding in previousestimates of influenza VE against all-cause mortality in the elderly[7]. We therefore aimed to determine whether the use of differentdefinitions of season would materially affect estimates of influenzaVE from an established sentinel surveillance network.
2. Methods
We used data collected as part of the Victorian general practice(GP) sentinel surveillance network for the years 2007–2012 to eval-uate how VE estimates changed when the “season” was definedusing different methods. This network has been described in detailelsewhere [8]. Briefly, recruitment follows the case test-negativedesign [9]: patients seeing their GP for influenza-like-illness (ILI;combination fever, cough and fatigue [10]) during the influenzasurveillance period (roughly May–October) are recruited at theGP’s discretion, swabbed and tested for influenza by real-timeRT-PCR. Those testing positive are cases and those testing neg-ative are noncases. The GPs collect demographic data (age, sex),symptom onset date, vaccination date and, since 2011, the pres-ence of conditions predisposing the patient to severe influenza
illness. Surveillance begins in epidemiological week (epiweek [11])18 and ends in epiweek 42, except in 2009 when surveillancestarted a week early and was extended until the end of theyear.
Fig. 1. Number of ILI consultations per week by case status for the Victorian GP sentinel surveillance network, 2007–2012. Only patients from whom a swab was taken areincluded. Lines indicate the analysis period using four different definitions: (1) the entire surveillance period; (2) weeks for which both a case and a noncase presented (solidl ic perd
acul(piostbscfodt
wdtni
ines); (3) the epidemic period based on ILI reports (dashed lines); (4) the epidemifferent y-axis scale used for 2009.
VE estimates from the sentinel network have been reported forll years 2007–2011 [12–16]. The present analysis focussed on thehanges to VE estimates seen when the analysis was restrictedsing four different definitions for season: (1) the entire surveil-
ance period; (2) consecutive weeks with both a case and a noncase;3) the epidemic period based on ILI reports; and (4) the epidemiceriod based on the product of ILI and laboratory reports (compos-
te variable). The last two definitions are based on an adaptationf a recently described WHO method for defining the influenzaeason [17]. The method aligns several years of historical data onhe median week of peak activity and assigns a seasonal thresholdased on visual inspection of the aligned data. The start of a sea-on is defined as the two consecutive weeks where the baseline isrossed. The composite variable was added as an improved methodor defining a season, given that it may be a better proxy indicatorf influenza activity because it accounts for both ILI surveillanceata and the proportion of ILI episodes that are laboratory proveno be due to influenza [18,19].
VE was estimated as 1-OR using logistic regression. Patientsere considered vaccinated if they had received the vaccine ≥14
ays prior to the onset of symptoms and excluded if vaccinationook place <14 days before symptom onset or if the date of vacci-ation or symptom onset was unknown. Patients were considered
nfluenza positive if they tested positive to any of A(H1), A(H3)
iod based on the product of ILI reports and laboratory reports (dotted lines). Note
or B viruses by real-time RT-PCR, but no separate analyses wereconducted for type or subtype. Patients presenting more than 8days after symptom onset were excluded to reduce the possibil-ity of false negative results. Models were adjusted for age group(<20, 20–64, 65+), week of onset, days between symptom onsetand consultation and, for 2011 and 2012, presence of a predisposingcondition. Week of onset was included as a continuous variable toavoid dropping any weeks without both a case and a control fromthe regression models. The only change in the analytical approachwas to vary the analysis period used.
3. Results
The data available for each year are shown in Fig. 1 with linesindicating the start and end of each analysis period using thefour definitions. Predictably, using the entire surveillance period(method 1) resulted in the longest analysis period (up to 22 weeksin 2009) and largest sample sizes, while the composite definition(method 4) generally defined the shortest analysis period (as fewas 7 weeks in 2010), resulting in the smallest sample size and the
highest ratio of cases to noncases (see Table 1).
VE estimates are shown in Table 1. In 2007, 2010 and 2012,point estimates changed by no more than five percentage pointsand nearly all VE estimates indicated a statistically significant
† All adjusted models include age group (<20, 20–64, 65+), week of onset, time betondition.‡ For 2010, 2011 and 2012 surveillance continued until epiweek 44 but no sampl
rotective effect of vaccination. Similarly, estimates for 2009, whenost influenza was due to the new A(H1N1)pdm09 virus, fluc-
uated in a narrow range from 0% (method 4) to −8% (method) but suggested no protection. In contrast, estimates for 2011,lthough positive, were not statistically significant and showedreater variation from 31% (method 4) to 49% (methods 1 & 3).inally, 2008 estimates showed the greatest instability rangingrom −5% (method 3) to −49% (method 4) and all had wide, non-ignificant confidence intervals.
. Discussion
The data presented here suggest that the choice of analysiseriod has little effect on VE estimates when confidence inter-als are narrow, but can have a dramatic effect when estimatesre imprecise, as seen in 2008 and 2011. These effects may be lessarked in estimates derived from very large surveillance systems.VE estimates for the same season often vary by country or
urveillance system due to differences in the types of vaccine used,opulation structure, age-specific vaccine coverage, the health careystem, the ILI definition used and the statistical model used. Theestrictions placed on the analysis period chosen may also sub-ly influence such variations. For example, if the season is definedy ILI activity only (method 3), elevated non-influenza ILI prior tohe season may lead to a longer analysis period than would haveeen chosen using the composite definition which also considers
aboratory-confirmed cases (method 4). In our data, this was seenn 2011 and led to substantially different analysis periods, sampleizes and VE estimates. Put into context for the 2010–11 north-rn hemisphere season, a German study [1] reported a VE estimate
f 70% (95% CI: 40, 85) based on consecutive weeks in which 10%f sentinel specimens were positive for influenza while a Canadiantudy reported VE = 37% (95% CI: 32, 15) [3] using the entire surveil-ance period. While these variations in VE will not be entirely due
onset of symptoms and consultation and, for 2011–2012, presence of a predisposing
re submitted by sentinel GPs.
to the different analysis periods used, it would be instructive toknow by how much the choice of analysis period influenced theestimates. In our study, the analysis period was the only variablechanged and differences in VE results should thus be due to changesin this variable only. Discrepancies among the methods may be dueto statistical effects, such as changes in sample size, but may alsobe due to surveillance practices, such as altered testing practisesas the season progresses. Consistency among studies may help usto better understand variations in VE estimates, at least in thoseregions of the world where there is an influenza season; in tropicalregions these effects may be more difficult to explore.
The definition of season should be considered a priori in VE stud-ies. If interim estimates are planned, the definition used could havedramatic effects on the data used and interpretation of the results.For example, in our dataset an early estimate for 2008 using data forthe first half of the season defined by the entire surveillance period(method 1; weeks 18–31) would exclude most weeks during theepidemic period defined by ILI and laboratory reports (method 4;weeks 31–41). In this case, the early estimate would suggest pro-tection, while the later estimate would suggest no protection. Inpractice, interim and end of season results were comparable in theUS 2007–2008 season [20] and European 2010–2011 season [5],but not the 2011–2012 European season [21]. As with final esti-mates, interim estimates should be restricted to the period wheninfluenza is circulating [6] and are vulnerable to changes in testingpractices within a season. Interim estimates are increasingly beingreported [22], but should be interpreted with caution.
5. Conclusion
“Season” is an arbitrary term defined differently in differentsurveillance systems. Studies of influenza VE should clearly statethe how the analysis period was defined and, where possible, alignthis definition with other studies reporting VE. Both interim and
ccine 3
ficaios
A
fMom
R
[
[
[
[
[
[
[
[
[
[
[
[
S.G. Sullivan et al. / Va
nal estimates should be restricted to the period when influenza isirculating. The inclusion of a figure identifying the data collectionnd analysis periods can provide a useful description of the tim-ng of case and noncase circulation. When estimates are impreciser when a particularly long or short analysis period is indicated,ensitivity analyses should be reported.
cknowledgements
The General Practitioner Sentinel Surveillance network is partlyunded by the Victorian Government Department of Health. The
elbourne WHO Collaborating Centre for Reference and Researchn Influenza is supported by the Australian Government Depart-ent of Health and Ageing.Conflict of interest statement: None.
eferences
[1] Englund H, Campe H, Hautmann W. Effectiveness of trivalent and monovalentinfluenza vaccines against laboratory-confirmed influenza infection in personswith medically attended influenza-like illness in Bavaria Germany, 2010/2011season. Epidemiol Infect 2012:1–9.
[2] Treanor JJ, Talbot HK, Ohmit SE, Coleman LA, Thompson MG, Cheng PY, et al.Effectiveness of seasonal influenza vaccines in the United States during a seasonwith circulation of all three vaccine strains. Clin Infect Dis 2012;55(7):951–9.
[3] Skowronski DM, Janjua NZ, De Serres G, Winter AL, Dickinson JA, Gardy JL, et al.A sentinel platform to evaluate influenza vaccine effectiveness and new variantcirculation, Canada 2010–11 Season. Clin Infect Dis 2012;55(3):332–42.
[4] Skowronski DM, Masaro C, Kwindt TL, Mak A, Petric M, Li Y, et al. Estimatingvaccine effectiveness against laboratory-confirmed influenza using a sentinelphysician network: results from the 2005–2006 season of dual A and B vaccinemismatch in Canada. Vaccine 2007;25(15):2842–51.
[5] Kissling E, Valenciano M, Cohen JM, Oroszi B, Barret AS, Rizzo C, et al. I-MOVEmulti-centre case control study 2010–11: overall and stratified estimates ofinfluenza vaccine effectiveness in Europe. PLoS ONE 2011;6(11):e27622.
[6] Jackson ML, Nelson JC. The test-negative design for estimating influenza vaccine
[7] Eurich DT, Marrie TJ, Johnstone J, Majumdar SR. Mortality reductionwith influenza vaccine in patients with pneumonia outside flu season:pleiotropic benefits or residual confounding? Am J Respir Crit Care Med2008;178(5):527–33.
[
1 (2013) 4280– 4283 4283
[8] Kelly H, Carville K, Grant K, Jacoby P, Tran T, Barr I. Estimation of influenzavaccine effectiveness from routine surveillance data. PLoS ONE 2009;4(3):e5079.
[9] Orenstein EW, De Serres G, Haber MJ, Shay DK, Bridges CB, Gargiullo P,et al. Methodologic issues regarding the use of three observational studydesigns to assess influenza vaccine effectiveness. Int J Epidemiol 2007;36(3):623–31.
10] Thursky K, Cordova SP, Smith D, Kelly H. Working towards a simple case defi-nition for influenza surveillance. J Clin Virol 2003;27(2):170–9.
11] Chu T. EPIWEEK. Stata module to create epidemiological week and equiva-lent epidemiological year; 2010. Available from: http://ideas.repec.org/c/boc/bocode/s4571521.html
12] Fielding J, Grant K, Tran T, Kelly H. Moderate influenza vaccine effectiveness inVictoria, Australia, 2011. Euro Surveill 2012;17(11), pii = 20115.
13] Fielding JE, Grant KA, Garcia K, Kelly HA. Effectiveness of seasonal influenzavaccine against pandemic (H1N1) 2009 virus, Australia, 2010. Emerg Infect Dis2011;17(7):1181–7.
14] Fielding JE, Grant KA, Papadakis G, Kelly HA. Estimation of type- and subtype-specific influenza vaccine effectiveness in Victoria, Australia using a testnegative case control method, 2007–2008. BMC Infect Dis 2011;11:170.
15] Kelly HA, Grant KA, Fielding JE, Carville KS, Looker CO, Tran T, et al. Pandemicinfluenza H1N1 2009 infection in Victoria, Australia: no evidence for harmor benefit following receipt of seasonal influenza vaccine in 2009. Vaccine2011;29(37):6419–26.
16] Kelly HA, Sullivan SG, Grant KA, Fielding JE. Moderate influenza vaccine effec-tiveness with variable effectiveness by match between circulating and vaccinestrains in Australian adults aged 20–64 years, 2007–2011. Influenza OtherRespir Viruses 2012, http://dx.doi.org/10.1111/irv.12018.
17] World Health, Organization. WHO Global Surveillance Standards for Influenza.Geneva: Global Influenza Programme, Surveillance and Monitoring team,World Health Organisation; 2012.
18] Goldstein E, Cobey S, Takahashi S, Miller JC, Lipsitch M. Predicting the epi-demic sizes of influenza A/H1N1, A/H3N2, and B: a statistical method. PLoSMed 2011;8(7):e1001051.
19] Goldstein E, Viboud C, Charu V, Lipsitch M. Improving the estima-tion of influenza-related mortality over a seasonal baseline. Epidemiology2012;23(6):829–38.
20] Belongia EA, Kieke BA, Donahue JG, Coleman LA, Irving SA, Meece JK, et al.Influenza vaccine effectiveness in Wisconsin during the 2007–08 season: com-parison of interim and final results. Vaccine 2011;29(38):6558–63.
21] Kissling E, Valenciano M, Larrauri A, Oroszi B, Cohen J, Nunes B, et al. Low and
decreasing vaccine effectiveness against influenza A(H3) in 2011/12 amongvaccination target groups in Europe: results from the I-MOVE multicentre case-control study. Euro Surveill 2013;18(5), pii = 20390.
22] Kelly H, Steffens I. Complexities in assessing the effectiveness of inactivatedinfluenza vaccines. Euro Surveill 2013;18(7):1–4.
