Influenza Vaccine Effectiveness in the Elderly Based on Administrative Databases: Change in Immunization Habit as a Marker for Bias

Background: Administrative databases provide efficient methods to estimate influenza vaccine effectiveness (IVE) against severe outcomes in the elderly but are prone to intractable bias. This study returns to one of the linked population databases by which IVE against hospitalization and death in the elderly was first assessed. We explore IVE across six more recent influenza seasons, including periods before, during, and after peak activity to identify potential markers for bias. Methods and Findings: Acute respiratory hospitalization and all-cause mortality were compared between immunized/nonimmunized community-dwelling seniors >= 65years through administrative databases in Manitoba, Canada between 2000-01 and 2005-06. IVE was compared during pre-season/influenza/post-season periods through logistic regression with multivariable adjustment (age/sex/income/residence/prior influenza or pneumococcal immunization/medical visits/comorbidity), stratification based on prior influenza immunization history, and propensity scores. Analysis during pre-season periods assessed baseline differences between immunized and unimmunized groups. The study population included,140,000 seniors, of whom 50-60% were immunized annually. Adjustment for key covariates and use of propensity scores consistently increased IVE. Estimates were paradoxically higher pre-season and for all-cause mortality vs. acute respiratory hospitalization. Stratified analysis showed that those twice consecutively and currently immunized were always at significantly lower hospitalization/mortality risk with odds ratios (OR) of 0.60 [95% CI0.48-0.75] and 0.58 [0.53-0.64] pre-season and 0.77 [0.69-0.86] and 0.71 [0.66-0.77] during influenza circulation, relative to the consistently unimmunized. Conversely, those forgoing immunization when twice previously immunized were always at significantly higher hospitalization/mortality risk with OR of 1.41 [1.14-1.73] and 2.45 [2.21-2.72] pre-season and 1.21 [1.03-1.43] and 1.78 [1.61-1.96] during influenza circulation. Conclusions: The most pronounced IVE estimates were paradoxically observed pre-season, indicating bias tending to overestimate vaccine protection. Change in immunization habit from that of the prior two years may be a marker for this bias in administrative data sets; however, no analytic technique explored could adjust for its influence. Improved methods to achieve valid interpretation of protection in the elderly are needed.