Investigating sources of non-response bias in a population-based seroprevalence study of vaccine-preventable diseases in the Netherlands

Abstract

Background: PIENTER 3 (P3), conducted in 2016/17, is the most recent of three nationwide serological surveys in the Netherlands. The surveys aim to monitor the effects of the National Immunisation Programme (NIP) by assessing population seroprevalence of included vaccine preventable diseases (VPDs). The response rate to the main sample was 15.7% (n = 4,983), following a decreasing trend in response compared to the previous two PIENTER studies (P1, 55.0%; 1995/1996 [n = 8,356] and P2, 33.0%; 2006/2007 [n = 5,834]). Non-responders to the main P3 survey were followed-up to complete a "non-response" questionnaire, an abridged 9-question version of the main survey covering demographics, health, and vaccination status. We assess P3 representativeness and potential sources of non-response bias, and trends in decreasing participation rates across all PIENTER studies.

Methods: P3 invitees were classified into survey response types: Full Participants (FP), Questionnaire Only (QO), Non-Response Questionnaire (NRQ) and Absolute Non-Responders (ANR). FP demographic and health indicator data were compared with Dutch national statistics, and then the response types were compared to each other. Random forest algorithms were used to predict response type. Finally, FPs from all three PIENTERs were compared to investigate the profile of survey participants through time.

Results: P3 FPs were in general healthier, younger and higher educated than the Dutch population. Random forest was not able to differentiate between FPs and ANRs, but when predicting FPs from NRQs we found evidence of healthy-responder bias. Participants of the three PIENTERs were found to be similar and are therefore comparable through time, but in line with national trends we found P3 participants were less inclined to vaccinate than previous cohorts.

Discussion: The PIENTER biobank is a powerful tool to monitor population-level protection against VPDs across 30 years in The Netherlands. However, future PIENTER studies should continue to focus on improving recruitment from under-represented groups, potentially by considering alternative and mixed survey modes to improve both overall and subgroup-specific response. Whilst non-responder bias is unlikely to affect seroprevalence estimates of high-coverage vaccines, the primary aim of the PIENTER biobank, other studies with varied vaccination/disease exposures should consider the influence of bias carefully.

Keywords: National immunization programme; Non-response bias; Seroprevalence; Survey.

Fig. 1

Flowchart of response type allocation based on participation behaviour for PIENTER 3

Fig. 2

Ordered variable importance for predicting an Absolute Non-Response from a Full Participant in P3

Fig. 3

Panel A. Ordered variable importance for predicting an NRQ from a FP using all data, in P3 (ANR n = 14,043. FP n = 5,553). Panel B– As panel A, but excluding participants who had missing data for all three of the NRQ-derived variables “Religion,” “Self-Reported Health Condition,” and “NIP Participation” (ANR n complete data = 9,451, FP n complete data = 5,546)

Fig. 4

Ordered variable importance for predicting a QO from a FP in P3