Although there was a low proportion of influenza cases in this study for which strain typing results were available, the statistically significant estimate of 59% effectiveness of influenza vaccine against all influenza in 2007 was generally consistent with Victorian state-wide strain typing data which indicated a partial match of circulating strains to those contained within the vaccine. These data showed A/H3N2 to be the predominant circulating subtype in 2007 (accounting for 56% of the characterised isolates) of which 42% were the A/Wisconsin/67/2005-like (vaccine) strain and the other 58% were the A/Brisbane/10/2007-like strain . When stratified by subtype, the 2007 vaccine was 68% effective (95% CI, 32 to 85%) against A/H3N2 infection and although the A/Brisbane/10/2007-like strain appeared to be the most dominant A/H3N2 strain, the relatively high VE estimate is likely to be explained by the antigenic similarity between the A/Brisbane/10/2007-like and A/Wisconsin/67/2005-like strains . However, stratified analysis did not indicate a significant protective effect of the vaccine against type A/H1N1 or type B infection in 2007, a finding which is supported by apparent mismatch of circulating strains to vaccine strains: 96% of the characterised A/H1N1 isolates were the (non-vaccine) A/Solomon Islands/3/2006-like strain whilst the characterised type B isolates were split between B/Florida/4/2006-like (41%), B/Shanghai/361/2002-like (35%) and B/Malaysia/2506/2004-like (24%) .
With a non-significant point estimate of 9%, the adjusted effectiveness of influenza vaccine against all influenza in 2008 was considerably lower than in 2007. The 2008 influenza season in both Victoria and across Australia was of lower magnitude than 2007 and characterised by a higher proportion of cases from influenza type B virus [19, 20]. This contrasts with a Western Australian study of the 2008 influenza season, which like Victoria was dominated by type B influenza virus with a late peak, that found a much higher and statistically significant VE point estimate of 58% (95% CI, 9 to 81%) against all influenza . Although this study was restricted to children aged 6-59 months, for whom there is a funded vaccination program in Western Australia, the reason for such a large difference is unclear. Both the sentinel general practice surveillance and other state-wide subtyping data indicated an approximately equal predominance of type A/H3N2 and type B viruses in 2008 , although few cases from the sentinel surveillance were able to be strain typed. Crude analysis suggested that the vaccine was 68% effective at a statistically significant level against type B infection, although after adjustment was 49% and not significant. This finding is generally consistent with strain typing data for isolates from across Victoria in which 42% were the vaccine B/Florida/4/2006-like strain and 58% were B/Malaysia/2506/2004-like, between which there was little antigenic similarity given their different lineages (B/Yamagata/16/88 and B/Victoria/2/87 respectively) [19, 20]. Strain typing of isolates sourced from elsewhere in Victoria indicated that circulating A/H3N2 was exclusively the A/Brisbane/10/2007-like strain and there was very little circulation of any A/H1N1 strains.
This study demonstrates the importance of conducting type- and subtype-specific assessment of influenza VE given the considerable variation that cannot be differentiated from a measure of VE against all influenza, despite what strain typing of circulating isolates may suggest about vaccine match/mismatch. A Canadian study that measured influenza VE at the trivalent component level during the 2006-2007 northern hemisphere season also observed wide variation between type- and subtype-specific adjusted VE point estimates from 12% to 92% , whilst two other observational studies in Wisconsin, United States of America (USA) and Canada also found type-specific variation of VE point estimates from -35% to 58% and 58% to 70% respectively [23, 24]. However, stratification of cases to assess type- and subtype-specific influenza VE compromises power as evidenced in this and the Canadian and USA studies. Insufficient power also compromised the ability of our study to generate more precise age group-specific estimates of VE, particularly in 2008 despite the collapse of five age groups into three. This was especially evident for those aged ≥ 65 years (the main risk group eligible for vaccination) in which a protective - but not statistically significant - effect against A/H3N2 influenza was demonstrated in 2007 but had too few cases to generate any VE estimates in 2008, highlighting a previously recognised limitation that the system is best suited to estimating VE amongst working age adults who comprise the majority of the surveillance population . Thus, whilst the program functions well as a representative surveillance system in assessing magnitude and duration of influenza seasons, further recruitment of sentinel GPs may be required to sufficiently power VE calculations, particularly during seasons of low magnitude or a dominant subtype.
A further limitation of this study is that the analysis has not controlled for the potential confounding effect of chronic or co-morbid conditions that are indicated for influenza vaccination. Several Canadian observational studies for which the specific confounding effect of co-morbid conditions was reported resulted in variations of the adjusted type- and subtype-specific VE estimates against seasonal influenza about the crude estimate of -23% to 7% [22, 24], and an increase of 15% on the crude seasonal VE against pandemic (H1N1) 2009 influenza . Whilst the confounding effect of co-existing chronic medical conditions on VE estimates may be modest and variable, these data will be included in the patient questionnaire and analysis in future seasons as a single variable. Pooling of confounders has been shown as unlikely to result in residual confounding .
Although clinical trials are the ideal method to assess vaccine efficacy, ethical, practical and financial considerations have lead to the emergence of observational studies - in particular case control studies such as this one - to routinely assess influenza VE [22, 24, 27–29]. However, inherent in observational study designs are biases that should be considered when interpreting and generalising the results to other populations. This study used test-negative control subjects which modelling, assuming no bias, has shown generally slightly underestimates the true VE under most conditions of sensitivity, specificity and influenza to non-influenza ILI attack rates  but was higher than traditional control subjects when assessed over three consecutive seasons . Another consideration is the sampling frame of attendees of general practices, for which a high proportion are working-age adults probably representing the mid-range of the clinical spectrum of influenza. More severe presentations (particularly among children and the elderly) are more likely to present to hospitals whilst asymptomatic or mild infections, estimated to be 34% , will not present to any medical facility. It is difficult to speculate how exclusion of cases from the peripheries of the clinical spectrum might affect the VE estimates, but highlights the importance of interpreting these results in the context of medically attended ILI in the general practice setting.
Ascertainment bias of influenza status within the study has been minimised by laboratory testing of all study participants with an assay that is at least 90% sensitive and 100% specific for influenza , and censoring of observations for which there was greater than four days between onset and specimen collection. Other factors, such as consistency of respiratory specimen collection, are difficult to quantify but may influence VE estimates. Furthermore, participants' illness and vaccination status are only known for the current season and don't account for cross-protection or prior immunity provided by previous vaccination or influenza infection.