Bias in control selection associated with the use of rapid tests in influenza vaccine effectiveness studies

Abstract

In test-negative design studies that use rapid tests to estimate influenza vaccine effectiveness (VE) a common concern is case/control misclassification due to imperfect test sensitivity and specificity. However, an imperfect test can also fail to exclude from the control group people that do not represent the source population, including people infected with other influenza types or other vaccine-preventable respiratory viruses for which vaccination status is correlated. We investigated these biases by comparing the effectiveness of seasonal 2023/24 influenza vaccination against influenza A and B based on PCR versus rapid test results, excluding controls who tested positive for SARS-CoV-2 or the other type of influenza. By PCR, VE against influenza A was 49% (95%CI 26-65%) after exclusion of PCR-confirmed influenza B and SARS-CoV-2 controls. Corresponding VE against influenza B was 65% (95%CI 35-81%). VE estimated by adjusting for COVID-19 vaccination status yielded similar estimates to the scenario that excluded SARS-CoV-2-positive controls. When case/control status and exclusions from test-negative controls were determined by rapid test, VE was reduced by 5-15 percentage points. Bias correction methods were able to reduce these discrepancies. When estimating VE from a test-negative study using rapid test results, methods to correct misclassification bias are recommended.

Figure 1.. Directed acyclic graph showing the selection of participants in TND, including (A) or excluding (B) SARS-CoV-2 cases.

HS = Health-seeking behaviour, Vi = Influenza vaccination, Vc = COVID-19 vaccination, AC = Age and chronic conditions, I = Infection due to influenza, C = Infection due to SARS-CoV-2, I* = Test detected influenza in a TND, C* = Test detected SARS-CoV-2, DAi = Diagnostic accuracy for influenza, DAc = Diagnostic accuracy for SARS-CoV-2, MARI = Medically attended acute respiratory infection, TND = Enrolment to a TND. In a TND evaluating VE against influenza, detection of influenza (I*) among the study participants influences case-control status in two ways: 1) other seasonal influenza types should be excluded from the controls due to cross protection of influenza vaccination, and 2) misclassification of true-positive cases as false-negative controls (6). The detection of influenza is dependent on diagnostic accuracy (DA_I) and characteristics of the influenza infection (I). Another source of bias is correlation between the influenza (Vi) and COVID-19 vaccination (Vc) by health-seeking behaviour (HS) (12) which alters the selection of SARS-CoV-2 cases by influenza vaccination status (Vc → C → MARI → TND). Notably, the correlation between V_I and V_C might be influenced by age and chronic conditions (AC) and can be expected to be more pronounced among priority groups, such as those with chronic illnesses and older adults. The correlation can be addressed by adjusting for COVID-19 vaccination (Vc in Figure 1A) or excluding positive SARS-CoV-2 cases (C* in Figure 1B) (12), with the latter depending on detection of the SARS-CoV-2 cases.