Residents in the provinces of Cremona, Bergamo, Mantova, Lecco, and Pavia in the northern Italian region of Lombardy, who were at least 65 years of age and sought influenza vaccination at local health authorities’ district offices or participating general practitioners were eligible for enrollment into this prospective cohort study during each of the 3 vaccination seasons, 2006–2007, 2007–2008, and 2008–2009. We excluded residents who were in the hospital, nursing homes, or rehabilitation centers in the 30 days preceding immunization, as well as those receiving home care or who were allergic to influenza vaccines.
Eligible subjects were informed by the vaccinators about the study and asked for their consent to participate; all those who accepted were administered a brief questionnaire to record basic demographic data and information on potential confounders, including smoking status, conditions potentially affecting immune response, functional status (as assessed through self-reported answers to questions about physical capabilities), presence of children in the household, and receipt of an influenza vaccine the previous year. Then they were administered either ATIV or the conventional nonadjuvanted trivalent subunit vaccine (Agrippal; Novartis Vaccines and Diagnostics), according to local, regional, and national influenza vaccination policy recommendations. There was no attempt at random assignment of vaccines. Information on the type of administered vaccine was recorded for each participant by the vaccinators, along with the previously collected information.
Both vaccines contained the recommended virus strains for the respective influenza season in the Northern Hemisphere. Vaccine doses were donated by the manufacturer and delivered to the local health authorities, who then distributed them to their district offices or participating general practitioners.
For each participant, residence status was confirmed through record linkage with administrative databases; all linkage failures were excluded from the study. Additionally, the presence of chronic disease or other relevant routinely collected medical history information was ascertained through record linkage with databases containing data on hospitalizations (discharge diagnoses), outpatient drug prescriptions (active ingredient and estimated duration of treatment), receipt of ambulatory care with specialist, and certified exemption from copayment of health-care costs.
The primary outcome was defined as a hospitalization for influenza or pneumonia occurring during a defined period in the influenza season (and in any case at least 3 weeks after vaccination), recorded in the hospital administrative database (discharge diagnosis with International Classification of Diseases
, Ninth Revision, Clinical Modification, codes 480–487). Influenza season was defined on the basis of a nationwide surveillance network (“Influnet”) that monitors virologically confirmed influenza occurrence in Italy annually. The network includes 1,000 general practitioners and family physicians, and it provides weekly incidence data, stratified by age and region (14
). We pooled data over the 3 influenza seasons, such that our elementary data record was a “person-season” at risk. As many people were included for more than one of the 3 years of observation, we used generalized estimating equations (15
) to take account of the correlation induced by measuring the experience of the same people for more than one influenza season.
Our case definition did not include positive laboratory confirmation of influenza virus. Therefore, in order to increase the specificity of the identification of cases hospitalized for influenza-related conditions, we defined 3 different time windows during the influenza season in which hospitalizations were counted. The broadest time window corresponded to the entire influenza season, as determined from Influnet. The narrowest time window corresponded to the period of adjacent weeks, around the peak influenza occurrence, having an influenza rate that exceeded 1 case per 1,000 person-weeks (17
). An intermediate time window was defined in the same way but with a threshold of 0.5 case per 1,000 person-weeks. Although the broader windows capture more cases, they are less specific for influenza-related cases. Accordingly, we defined a priori our primary analysis to be based upon the narrowest window as it provided the greatest specificity and hence the least bias. Also, in order to estimate the amount of potential misclassification of the discharge diagnosis, a sample of hospital discharge records was validated, and the diagnosis was compared with the actual hospital discharge diagnoses.
Since adjuvanted vaccine was preferentially recommended for high-risk frail individuals at many sites, it was known a priori that analysis of study outcomes would have to take this source of potential bias into account. To assess and control for confounding, we used stratification coupled with Mantel-Haenszel summary estimates of a pooled effect measure. Variables assessed as potential confounders included age, gender, influenza season, local health authorities and vaccine provider, functional status, smoking, recent infectious disease, transfusion, intestinal disorder, self-reported flu symptoms, cumulative length of stay in the hospital and cumulative number of drug prescriptions (both in the 5 years preceding the vaccination), infectious disease, and chronic conditions, such as chronic obstructive pulmonary disease (COPD), kidney disease, diabetes, cardiovascular disease, peripheral vascular disease, cancer, and history of hospitalization for pneumonia, influenza, or emphysema. Because of the number of potential confounders, we were not able to control for all confounders simultaneously in a single stratified analysis; we therefore also conducted a multivariate analysis that used a propensity score as a summary confounding score. For the propensity score model, we used a logistic regression to estimate the probability of receiving ATIV versus TIV. Variables that were included in this model included age, sex, influenza season, community and provider, cumulative length of stay in the hospital, and cumulative number of drug prescriptions in the 5 years preceding the vaccination, physical impairment, smoking, presence of children in the home, recent transfusion, recent intestinal disorder, recent self-reported flu symptoms, recent infectious disease, history of hospitalization for pneumonia, influenza or emphysema, COPD, diabetes, cardiovascular disease, chronic kidney disease, peripheral vascular disease, and cancer.
To avoid including in the propensity score model nonconfounding predictors of exposure, which would not reduce confounding but would decrease precision, we fit a preliminary logistic model predicting hospitalization with influenza-like illness that included all these covariates, along with study vaccine, to determine the strength of relation of each variable with the study outcome. As a second stage, we then created the propensity score model using those predictors from the preliminary outcome model that had a relative risk of at least 1.4: age, sex, influenza season, community and provider, physical impairment, cumulative length of stay in the hospital and cumulative number of drug prescriptions in the 5 years preceding the vaccination, history of hospitalization for pneumonia, influenza or emphysema, COPD, chronic kidney disease, diabetes, recent infectious disease, and recent transfusion (18
). From this model, we computed the propensity to receive ATIV for each person-season of observation and added that to the data as an additional, derived variable. To improve comparability of the 2 vaccine groups, we controlled for the propensity score in multivariate models, but first we excluded all outlier observations (“trimming”), defined as those below the lower 2.5% of the tail of the TIV observations (3,355 observations) and above the upper 2.5% tail of the ATIV observations (2,894 observations). These tails are outside the primary area of overlap of the propensity scores, and they increase residual confounding in any type of analysis (19
In the multivariate analyses, we used generalized estimating equations to account for the inclusion of people in more than one season. Our final multivariate analysis was based on doubly robust estimation, in which the strongest confounders and the propensity score based on all confounders were included in the logistic model; this model in principle should provide the best control of confounding achievable with these data (20
We conducted an additional analysis to compare the risk of hospitalization before each influenza season in the 2 vaccine groups. We therefore identified the events that occurred in the period May–September in people subsequently vaccinated and enrolled in the study. This was done under the assumption that, if our model in the primary analysis had completely adjusted for confounding, with the application of the same doubly robust technique outside the influenza season (i.e., without influenza activity and independently of the vaccine effect), we would estimate no risk difference between the 2 cohorts; any increased risk would represent the magnitude of the residual bias.
Statistical analysis was performed by using SAS, version 9.1, software (SAS Institute, Inc., Cary, North Carolina). We used multiple (i.e., 5) imputations (Proc MI and Proc MIANALYZE) to handle missing values (21
The study protocol was submitted to and approved by the ethics committees of the participating local health authorities.