Seasonal Influenza Vaccine Effectiveness (SIVE): an observational retrospective cohort study - exploitation of a unique community-based national-linked database to determine the effectiveness of the seasonal trivalent influenza vaccine

Article history

The research reported in this issue of the journal was funded by the HS&DR programme or one of its proceeding programmes as project number 09/2000/37. The contractual start date was in April 2011. The final report began editorial review in December 2012 and was accepted for publication in June 2013. The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The HS&DR editors and production house have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the final report document. However, they do not accept liability for damages or losses arising from material published in this report.

Copyright statement

© Queen's Printer and Controller of HMSO 2013. This work was produced by Simpson et al. under the terms of a commissioning contract issued by the Secretary of State for Health. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

Objectives

The aim of the Seasonal Influenza Vaccine Effectiveness (SIVE) study was to examine the effectiveness of the seasonal influenza vaccination in individuals registered with a national sample of general practices in Scotland. Our three specific objectives were to evaluate:

1. uptake of the influenza vaccine by the relevant at-risk populations, i.e. patients with relevant comorbidities and those aged ≥ 65 years, as well as by the general population

2. the reduction in the expected incidence of influenza-related morbidity and mortality in these at-risk groups, as this is the major rationale behind current immunisation policies

3. the effectiveness of the influenza vaccine in the population as a whole.

Study design

Almost all individuals resident in Scotland are registered with a general practice, which provides a comprehensive array of health-care services, including the issuing of prescriptions for medications. Specialist hospital care services are typically accessed through referral from primary care or, in emergency situations, through patients attending accident and emergency. General practitioners (GPs) also provide and co-ordinate much of the care of patients discharged back into the community after a hospital admission. The Practice Team Information network practices included in this study are a representative sample constituting 5% of Scottish practices. These practices receive an annual financial incentive to electronically record all face-to-face contacts with patients. 13 Data from practices within Scotland are of high quality (the completeness of capture of contacts and accuracy of clinical event coding in primary care has been found to be > 90%14) and their value for epidemiological research has been repeatedly demonstrated. 15 Using each patient's unique Community Health Index (CHI) number, general practice patient-level data were extracted and then linked to the Scottish Morbidity Record (SMR) catalogue, which has information on all inpatient hospital admissions within Scotland as well as information on death certification linked from the General Register Office for Scotland (GROS)16 (Figure 1). Hospital data from 1981 onwards have been shown to be reliable, with completeness and accuracy rates > 90%. 17 Additionally, we used the Health Protection Scotland (HPS) virology data set, which records all laboratory-confirmed cases of influenza in Scotland.

We established the following key characteristics of each identified patient in the cohort: sex; age (0–4, 5–14, 15–44, 45–64, 65–74 or ≥ 75 years); socioeconomic status [Scottish Index of Multiple Deprivation scores18 expressed as quintiles: 1 (most affluent) to 5 (most deprived)]; smoking status (current smoker, ex-smoker, non-smoker or not recorded); urban/rural location (one large urban and six remote rural locations were included in the study); whether or not the patient belonged to any clinical at-risk groups (i.e. those suffering from chronic respiratory, heart, kidney, liver or neurological disease, immunosuppression or diabetes); pregnancy status; Charlson comorbidity index;19 previous pneumococcal and influenza vaccination; and number of previous primary care consultations, prescribed drugs and hospital admissions (e.g. in the year prior to 1 September 2000). No direct measure of functional status existed within the database; to allow this to be taken into account, home consultations by a GP only (vs. practice attendance by the patient) and nursing home residence (including social care support) were also included in our propensity score.

FIGURE 1.

Flow diagram for the SIVE study. ISD, Information Services Division.

Procedures

Clinical conditions (consultations for influenza and other acute respiratory infections), risk group information and prescribing data (e.g. seasonal influenza vaccination and pneumococcal polysaccharide status) were extracted from primary care records. Influenza primarily infects the respiratory tract, which can often result in pneumonia and influenza being grouped together in data recording. 20 We therefore extracted from the SMR and the GROS details on the primary diagnosis of hospitalisation and cause of death from influenza grouped with pneumonia. Furthermore, we also analysed chronic obstructive pulmonary disease (COPD) grouped with influenza and pneumonia, as influenza can account for a significant proportion of COPD exacerbations resulting in hospitalisation and death. 21

A number of secondary analyses were undertaken using other outcomes. The effect of vaccination status on hospital admissions and deaths relating to cardiovascular and cerebrovascular events as a composite outcome was analysed. In addition, exploratory analyses were undertaken to assess the effect of vaccination status on outcomes for which it would not be expected to have an effect, for example appendicitis or trauma. This approach of using an alternative outcome as a negative control has been shown to be a useful method for detecting residual bias. 22

Data from 1 September 2000 to 31 August 2009 were used. This allowed for the analysis of nine influenza seasons (2000/1 to 2008/9), yielding a total of 1,767,919 person-seasons for analysis. The influenza season was defined as the period from the date of the first influenza isolate reported by HPS for each year, in or post week 40 of that year, until the date of the last influenza isolate, pre or in week 20 of the subsequent year (i.e. the period of peak influenza) (Figure 2).

FIGURE 2.

Relationship of first influenza season (2000/1) to pre-, post- and non-influenza season periods. Influenza season was defined as the period from the date of the first isolate ≥ week 40 until that of the last isolate ≤ week 20.

Vaccination was used to define exposure status if it was given at a time point between the start of the pre-influenza season (e.g. 1 September) and the end of the influenza season. An individual was defined as vaccinated 14 days after the seasonal influenza vaccine was administered. The time period from the first day of the influenza season to day 14 post-vaccination was defined as ‘unexposed’ and the period from day 14 post vaccination until the end of the influenza season was defined as ‘exposed’. A detailed description of our methods has previously been published. 23

The earliest influenza season began on 26 September and the latest began on 25 November (Table 1). All seasons finished in May.

General practitioners in this study were also involved in the HPS sentinel swabbing scheme, whereby practices are encouraged to obtain nasal/throat swabs from patients of all ages who have symptoms suggestive of influenza. Each general practice is requested to submit five swab samples per week to the West of Scotland Specialist Virology Centre in Glasgow, a World Health Organization-accredited National Influenza Centre which participates in a quality assurance programme to maintain this status. Here, RT-PCR testing for a range of respiratory pathogens is carried out. Swabs should be obtained from any patient presenting for consultation with influenza symptoms, across all age groups and regardless of whether the patient has or has not been vaccinated. We also included results for the patients in our study from swabbing carried out for routine diagnostic purposes in primary and secondary care outside the sentinel scheme.

Vaccine effectiveness is expressed as a percentage, and represents a reduction in risk provided by the vaccine for a given outcome. To calculate VE, each patient's swab data were linked using his or her CHI number, allowing patient characteristics such as vaccination status to be established from general practice and hospital admission data.

Framework for detecting residual confounding

We undertook additional analyses to identify the presence of residual confounding. This has been recommended as part of an analytical framework when reporting VE using observational study designs. 9 We assessed the variation in VE using the following criteria:

1. Seasonality. Stratification by season is more important when VE is measured using non-specific outcomes. Each year of observation was partitioned into four periods: the non-influenza period, pre-influenza season, influenza season (when influenza virus is circulating) and post-influenza season (see Figure 2). Maximal VE should be seen during the influenza season. The vaccine should have no effect on outcome in the pre-influenza and non-influenza seasons. The non-influenza season used vaccination status from the previous influenza season. This was to minimise the bias that might occur when vaccination status is applied retrospectively. This retrospective application of vaccine status included patients who died during the preceding non-influenza season as unvaccinated, despite the fact that they would not have survived long enough to be eligible for vaccination.

2. Vaccine match. VE should be lower in years during which the influenza vaccine was a poor match for the circulating virus.

3. Severity of influenza season. VE should be greater in years during which the circulating virus caused a large excess mortality during the influenza season.

4. Age. It is thought that influenza vaccine is less effective in the oldest age groups owing to immune senescence. 10 If this assumption is correct, VE should be lowest in the oldest subgroup. A stratified analysis of age groups was undertaken to assess for this effect.

5. Specificity of outcome measure. VE should be greatest for the most specific outcome (laboratory-confirmed influenza infection) and lowest for the least specific outcome (all-cause mortality). In addition to the primary analysis, in the three non-laboratory databases (primary care, acute hospital discharge and death register) we undertook analyses using the more influenza-specific outcomes (influenza-coded deaths, hospital admissions and primary care attendances) and less specific outcomes (any emergency hospital admissions, all-cause deaths and any primary care attendances).

Statistical methods

We constructed a propensity score model and calculated adjusted odds ratios (ORs) to assess differences in vaccine uptake by patients with each of the characteristics outlined previously. The model was non-parsimonious in order to include a wide range of factors that influence propensity to be vaccinated. 24 Using this model, a probability of receipt of the influenza vaccination was assigned to each individual in the cohort. Unadjusted illness and mortality rate ratios (RRs) were calculated as the ratios of the rate of emergency admission to hospital or death in vaccinated patients to the rate of emergency admission to hospital or death among those who did not receive the vaccine. The unadjusted estimate of VE was (1 − RR) × 100. Adjusted illness and mortality hazard ratios (HRs) were estimated from a Cox proportional hazards model adjusted for all patient characteristics, and the propensity score (as deciles) and VE were calculated as 1 − HR where HR was that of the measured outcome in vaccinated compared with unvaccinated individuals.

For our analysis we pooled data from all nine seasons. This a priori approach gave a more powerful analysis and increased precision, particularly for the less common outcomes (with some patients represented for multiple seasons). In this pooled analysis, we accounted for the within-person correlation resulting from repeated measures on the same individual in different influenza seasons by making an adjustment for clustering using robust standard errors in the model. The effect of time was implicitly considered in the model by using the Cox proportional hazards approach and the inclusion of year as a fixed effect allowed for variation between seasons. In our propensity score matched analysis, only patients with similarly matched propensity to be vaccinated were included. To test the validity of the pooled analysis, heterogeneity between years was determined by testing for interaction between vaccine status and year for the outcomes. Where significant heterogeneity occurred, the pooled analysis was restricted.

For estimates of VE derived from linked virological swab data, we carried out a nested case–control study design. A generalised additive logistic regression model25 was fitted adjusting for the effects of week during the study period, age, sex, deprivation, previous number of primary care consultations and being in a clinical at-risk group. Some of these patients did not receive the influenza vaccine; some received the vaccine, but after they were tested; and others received the vaccine before they were tested. We therefore measured VE by comparing swabs taken after vaccination from individuals who were vaccinated with swabs taken from all those who were not vaccinated at the time the swab was taken (people who were unvaccinated at the time of the swab and who were then subsequently vaccinated counted as unvaccinated in our analysis, as did people who were never vaccinated). We assessed only the first dose when two doses were given. We also stratified our analysis based on patients aged ≥ 65 years and those aged < 65 years who were classified as at risk. Using the ‘epitools’ library in R version 2.14.1 (the R Foundation for Statistical Computing, Vienna, Austria), we calculated 95% confidence intervals (CIs) for the OR and RR by median-unbiased estimation using the ‘odds ratio’ and ‘rateratio’ functions. 26 Tests of the differences were performed using the ‘rate2by2.test’ function.

Further statistical methods to adjust for confounding

Vaccine uptake

In total, during the 1,767,705 person-seasons of observation over nine influenza seasons, 274,071 seasonal influenza vaccinations were administered to the whole population, of which most (93.6%; n = 256,474) were distributed to at-risk patients targeted for vaccination. There was 69.3% uptake of the vaccine among those aged ≥ 65 years (178,754 vaccinations during 258,100 person-seasons). For at-risk patients aged < 65 years there was a 26.2% uptake (77,264 vaccinations during 295,116 person-seasons).

High vaccine uptake was found among the oldest age group (≥ 75 years), care home residents, those previously vaccinated (Table 2), those with chronic diseases (except for chronic respiratory disease) and those being issued many prescriptions (Table 3). After adjustment, the OR for uptake was lower for individuals who smoked and those with no recorded smoking status (see Table 2) and for those with five or more comorbidities when compared with those with no comorbidities (see Table 3). There were similar findings for at-risk patients and those aged ≥ 65 years, although for the latter, an OR (with 95% CI) < 1 was also found among people whose contacts with primary care were by home consultation only (unadjusted OR 0.53; 95% CI 0.50 to 0.56) and, after adjustment for covariates, among those admitted multiple times to hospital when compared with those not admitted (> 10 hospitalisations: adjusted OR 0.71; 95% CI 0.53 to 0.95).

For our propensity score, our prediction model showed a high discriminatory power to identify patients who did and did not receive influenza vaccine (see Appendix 1) [all patients: receiver operating characteristic – area under the curve (ROC_AUC) = 0.97; patients aged ≥ 65 years: ROC_AUC = 0.88; at-risk patients aged < 65 years: ROC_AUC = 0.90].

Laboratory-confirmed influenza

A total of 3323 swabs were taken from 3016 patients over the nine seasons and then tested with RT-PCR for evidence of influenza infection. Although all subgroups were represented, patients from whom swabs were taken were more likely to be younger (aged < 75 years), female and relatively affluent (Table 4). During our study, male patients and the socioeconomically affluent were more likely to test positive for influenza. Of all swabs taken, 13.9% tested positive for RT-PCR-confirmed influenza. This included one-quarter of swabs from school-aged children.

Vaccine effectiveness for the trivalent inactivated influenza vaccine was 57.1% (95% CI 31.3% to 73.3%) (Table 5). VE was 59.6% (95% CI 22.0% to 79.1%) for at-risk patients aged < 65 years and 18.8% (95% CI – 103.7% to 67.6%) for patients aged ≥ 65 years. Peak VE was found in the 2007/8 influenza season.

Clinical outcomes

In total, during the nine influenza seasons, 5591 of 162,698 (3.4%) primary care acute respiratory consultations (acute respiratory infections including influenza-like illnesses) were for influenza-like illnesses, 3096 of 86,531 (3.6%) emergency hospitalisations (all cause) were due to influenza and pneumonia and 2096 of 10,121 (20.7%) deaths (all cause) were due to influenza and pneumonia. One-fifth of hospitalisations due to influenza or pneumonia that occurred during the nine influenza seasons were in at-risk patients aged < 65 years (n = 502/2835 hospitalisations; 17.7%) and most occurred in older patients aged ≥ 65 years (n = 1700/2835; 60.0%). Almost all deaths from influenza or pneumonia occurred in those groups targeted for vaccination (n = 2001/2096; 95.5%); most occurred in people aged ≥ 65 years (n = 1860/2096; 88.7%).

In Tables 6 and 7, we report the incidence rate of clinical outcomes (and unadjusted rates) for unvaccinated versus vaccinated individuals as well as VEs for the whole cohort. In Table 8 we report results for one subgroup, clinically at-risk patients aged < 65 years; Table 9 presents the results for people who are aged ≥ 65 years. In Tables 6 and 7, where results for the whole population are reported, it should be noted that unadjusted rates among the vaccinated are far higher than among the unvaccinated. This is largely due to the majority of the vaccinated being older people (aged ≥ 65 years) who are more likely to have clinical outcomes (particularly hospitalisation and death) than younger, unvaccinated individuals.

After adjustment for confounding factors (including age in the model), recipients of influenza vaccine were less likely to have acute respiratory illnesses, less likely to require hospitalisation for influenza, pneumonia, COPD or cardiovascular disease, and less likely to die from all outcomes studied in the whole population (see Tables 6 and 7). In the at-risk population aged < 65 years, analyses were less precise due to lower event rates for most of the outcomes. However, significant VE was found for hospitalisations and deaths from cardiovascular disease and deaths due to influenza, pneumonia or COPD (see Table 8). Significant VE was found for most hospitalisation and death outcomes in the population aged ≥ 65 years (see Table 9).

Significant heterogeneity was found in 2000/1 and 2001/2 when compared with other years (test for interaction: p < 0.0001) and these years were excluded from the final pooled analysis. After propensity score matching was performed (for seven seasons from 2002/3 to 2008/9) there were 78,048 matched pairs of patients. Influenza vaccine recipients were less likely to be hospitalised or to die from influenza, pneumonia, COPD or cardiovascular disease (see Tables 6 and 7).

In our sensitivity analysis using negative controls, we found significant VE for trauma hospitalisation with the exception of at-risk patients aged < 65 years. However, further sensitivity analyses (see Table 10) revealed that VE for trauma during the post season was greater than that during peak influenza season, and this suggests that it may be a poor choice of negative control. No significant VE was found for appendicitis and hernia hospitalisation.

Sensitivity analyses

Vaccine match and severity of influenza season

Poor precision resulted in difficulty comparing VE by season. Using the laboratory-confirmed influenza data (see Table 5) the most severe season was 2007/8, with 29.15% of positive tests. This year (2007/8) also had the highest VE in our clinical outcomes (Figure 3: deaths from influenza and pneumonia for people aged ≥ 65 years are shown here as an example). During the season 2005/6, where the influenza B component of the vaccine was regarded as poorly matched, VE was the second lowest for laboratory-confirmed influenza in the nine seasons (see Table 5). However, no obvious decline in VE for this poorly matched year was found in our clinical outcomes.

FIGURE 3.

Death from influenza and pneumonia by season in patients aged ≥ 65 years.

Where all-cause death was used as an outcome, VE was similar between seasons (Figure 4).

FIGURE 4.

All-cause death by season in patients aged ≥ 65 years.

Modelling an unmeasured confounder

We used estimates from published data to model the effect of inadequately measured confounders on VE in our matched population.

We found that for one outcome, death due to influenza or pneumonia, a scenario of 20% prevalence of frailty among unvaccinated individuals resulted in the vaccine no longer being effective (VE 13.2%, 95% CI 2.2% to 22.3%). 9 Frailty resulted in a twofold increase in the risk of death and was not present in the vaccinated population. By varying the prevalence of frailty and its risk on outcome, we modelled the change in VE estimates for a number of scenarios (Table 11). We also found that this model can be displayed effectively in graphical form (Figures 5 and 6). In Figures 5 and 6 the VE is estimated given levels of the unmeasured confounder in the vaccinated population (x-axis) and unvaccinated population (each of the lines representing a different prevalence of the unmeasured confounder in the vaccinated population). At points where the prevalence of the unmeasured confounder is the same in both unvaccinated and vaccinated individuals, the baseline result is achieved, as there is no difference in the level of the unmeasured variable in the two groups.

TABLE 11 - Sensitivity analysis to quantify the effects of a hypothetical unmeasured confounder on the matched analysis results

Increase in the risk of outcome on account of the confounder	Prevalence of confounder		Death from pneumonia or influenza: adjusted VE (95% CI)
	In unvaccinated individuals (%)	In vaccinated individuals (%)
–	0	0	27.63 (18.51 to 35.73)
Doubled	5	0	24.01 (14.43 to 32.51)
Doubled	15	5	20.73 (10.74 to 29.61)
Doubled	20	5	17.29 (6.86 to 26.55)
Quadrupled	5	0	16.77 (6.28 to 26.09)
Quadrupled	15	5	8.75 (– 2.75 to 18.96)
Quadrupled	20	5	– 0.69 (– 13.38 to 10.58)

FIGURE 5.

Unmeasured residual confounding and VE (doubling of risk of death from influenza and pneumonia).

FIGURE 6.

Unmeasured residual confounding and VE (quadrupling of risk of death from influenza and pneumonia).

Sensitivity analysis framework

There are two types of confounding which have a role when estimating VE: (i) confounding by indication, or negative confounding, which will underestimate VE and will include people with underlying chronic diseases who are more likely to be vaccinated (and are at greater risk of disease and only moderately elevated risk of death); and (ii) positive confounding, which will overestimate VE and will include healthier individuals who are more readily vaccinated (as they may be more mobile and actively seek the vaccine) and, conversely, very frail people who are less likely to be vaccinated and may die soon thereafter. 43

First, to try and account for residual confounding, we used the most detailed electronic health information available (in particular that from primary care, which uses the granular Read coding system44) to include in our model a score determining propensity to be vaccinated. We found variation in uptake between groups. Among care home residents, for instance, uptake was high and this reflects what is a common practice in the UK of vaccinating residents en masse by visiting nursing home facilities. We also found that with increased use of primary care, uptake increased and this may be due to the increased opportunity to vaccinate. We did, however, find a curvilinear relationship between vaccination and Charlson comorbidity. In particular, for patients aged ≥ 65 years, there were higher ORs for uptake in the presence of at least one Charlson comorbidity or hospitalisation, but lower ORs (than in those with no morbidity) were found among those with a Charlson score of five or more and > 10 hospitalisations. By matching individuals in our cohort on a propensity score we were able to effectively compare people with a similar propensity to be vaccinated. This should have limited the effect of positive confounding, which will overestimate VE. We feel that the use of our propensity score has reduced bias for our estimates of VE. However, our VE for all-cause deaths was still higher than the estimated excess mortality attributed to influenza of 5–10%. 9

We also attempted to explore and account for residual confounding using the framework suggested by Simonsen et al. 8 The influenza vaccine should be not be effective for outcomes unrelated to infection with seasonal influenza and reassuringly no VE was found for hospitalisation due to appendicitis and hernia. A positive VE (which continued in the post-influenza season), however, was found for hospitalisation due to trauma (e.g. for wounds and fractures). However, further sensitivity analyses revealed that VE for trauma during the post-influenza season was greater than during peak influenza seasons, and this suggests that it may be a poor choice of negative control. It is of interest though that the unvaccinated group was more likely to be hospitalised than the vaccinated group for this type of admission. When analysing VEs stratified by the four periods of each year (influenza season and non-, pre- and post-influenza season), maximal VE, determined by use of viral surveillance information, was generally seen for most of our clinical outcomes during the influenza season. VEs outside the influenza season were usually close to zero with wide 95% CIs owing to the relatively small exposure periods outside the main influenza season (at most 2.5 months) and a lack of influenza/pneumonia outcomes. Positive VEs were, however, found for less specific outcomes such as hospitalisations and deaths for any cause, or for outcomes that can be exacerbated or caused by factors other than influenza, for example, acute respiratory disease, COPD hospitalisation or death from cerebrovascular and cardiovascular disease. Our VE for less specific outcomes, however, was found to continue during post- and non-influenza seasons when influenza was not circulating, suggesting some residual bias still exists. Sophisticated seasonality techniques such as case-centred logistic regression, which attempt to account for deaths that take place when influenza is not circulating, have found that without vaccination, excess mortality during influenza season would be 9.8%, and estimated VE 4.6% (half of all deaths being prevented by vaccine). 40

Our unique cohort allowed VE to be derived for a range of outcomes in the same population over multiple influenza seasons. However, although the modest size of our cohort made it feasible to collate centrally almost all cases of influenza-related disease, allowing for completeness of reporting, any analysis by subgroups or by individual season resulted in poorer precision and wide CIs. It was therefore less than straightforward to draw conclusions regarding the validity of our VE estimates using Simonsen's framework. 8 However, we were able to determine that during the most severe season for our cohort (2007/8, as determined from swabs), VE was highest for our outcomes. Although there was no protective effect for the outcome appendicitis and hernia, a protective effect was found for trauma, which is unrelated to influenza. The significant heterogeneity in the matched analysis for the years 2000/1 and 2001/2 may have been caused by accuracy and completeness issues (found in 2000), which were resolved by the Information Services Division (ISD) in subsequent years. 45

We attempted to use further methods to assess the degree of bias in our VE estimates. We were unable to use instrumental variables for this purpose as none were found to be valid. However, we modelled the effect of an unmeasured confounder, such as frailty, on our VE estimates. We found that VE for our primary outcome, death due to influenza or pneumonia, was robust to modelling a high prevalence of unmeasured frailty. This is despite the fact that we included a number of variables in our models which would have captured some degree of frailty, such as residence in a nursing home, provision of social care, number of previous hospital admissions, number of GP consultations and number of repeat prescriptions.

Implications for practice

Using VE estimates for our most specific outcome, RT-PCR-confirmed influenza over a 9-year period, the seasonal influenza programme was found to be effective, particularly in preventing influenza in younger, clinically at-risk groups of patients.

Research recommendations

Although the modest size of our cohort made it feasible to collate centrally almost all cases of influenza-related disease, allowing for completeness of reporting, the analysis of subgroups (in particular, older age groups) or analysis by individual season resulted in poorer precision and wide CIs. Any future work should therefore aim to address this issue by ensuring adequate power to test VE in these subgroups of patients, while minimising the effect of bias, such as health-seeking behaviour. While work is being undertaken to produce better vaccines, continued monitoring and a strong international evidence base for the effectiveness of seasonal influenza vaccination programmes is necessary.

Contributions of authors

Dr Colin Simpson (Reader in Population Health Sciences) and Dr Nazir Lone (Clinical Fellow, Intensive Care) were principal investigators and led the writing of this report.

Professor Lewis Ritchie (Professor of Primary Care), Professor Aziz Sheikh (Professor of Primary Care Research & Development) and Dr Jim McMenamin (Consultant Epidemiologist) helped design the study and commented on drafts of the paper.

Professor Chris Robertson (Professor of Statistics) and Dr Kim Kavanagh (Research Fellow, Statistics) helped to design the study, carry out the analyses and write the paper.

All people

Cohort spilt into two to create model (n = 1,179,201) and test predictions (n = 588,718). From the model output a propensity score based on the model coefficients is created and a score is assigned to each individual in the cohort.

Model output

Model output (PDF download)

FIGURE 7.

Propensity to be vaccinated score in deciles. (a) Vaccinated; and (b) unvaccinated.

FIGURE 8.

Receiver operating characteristic curve for all individuals. AUC, area under the curve.

At-risk patients aged under 65 years

At-risk patients aged under 65 years (PDF download)

FIGURE 9.

Propensity to be vaccinated score in deciles for age < 65 years. (a) Vaccinated; and (b) unvaccinated.

FIGURE 10.

Receiver operating characteristic curve for at-risk patients aged < 65 years. AUC, area under the curve.

Patients aged 65 years and over

Patients aged 65 years and over (PDF download)

FIGURE 11.

Propensity to be vaccinated score in deciles for age ≥ 65 years. (a) Vaccinated; and (b) unvaccinated.

FIGURE 12.

Receiver operating characteristic curve for all individuals aged ≥ 65 years. AUC, area under the curve.

1. PROJECT TITLE

Seasonal Influenza Vaccine Effectiveness (SIVE): exploitation of a unique community-based national linked dataset

2. CHANGES SINCE THE PROTOCOL WAS SUBMITTED

We are now in a position to undertake analysis on individual patient virological swab data, which will be linked to the primary-secondary care dataset. These data will be made available by Health Protection Scotland (HPS) for all eight influenza seasons for which data are potentially available (i.e. 2000-2008). Access to these individual patient data should provide more robust estimates of vaccine effectiveness (VE).

3. PLANNED INVESTIGATION

4. KEY REFERENCES