The blind spot: Studying the association between survey non-response and adherence to COVID-19 governmental regulations in a population-based German web-survey

Currently, a multitude of research deals with adherence to COVID-19 regulations. Although selective non-response might question the validity of generalising research findings, the issue has, as yet, received only little attention. Presumably, choosing to participate in a COVID-19 study is based on a similar decision-making process as that concerning adherence to COVID-19 regulations. Certain characteristics might predict both outcomes which would result in overestimated mean levels and a biased predictor structure of adherence to COVID-19 regu-lations. We used a random sample of adolescents (born 2001-2003) from the German family panel study pairfam who were first interviewed (face-to-face) in winter 2018/19 and were in-vited to participate in a (web-based) follow-up COVID-19-interview in spring 2020. Using a simple weighting procedure and Heckman selection models, we found an overestimated mean of adherence to COVID-19 regulations, with the association with gender being overestimated and that with education and migration background underestimated. Other than expected, the extent of bias was less severe and fewer variables were affected. We suggest including a set of additional variables into the models estimating adherence to tackle the bias in the predictor structure and to address mean level bias by using weights accounting for population charac-teristics. Although COVID-19 studies indeed appear to provide biased results, being able to reduce this bias is generally good news for high-quality COVID-19 studies.

Automated summary

This study investigates the adherence to COVID-19 regulations and highlights the potential bias caused by selective non-response. By using weighting procedures and selection models, the study finds an overestimation of adherence levels, an overestimation of the association with gender, and an underestimation of the association with education and migration background. The authors suggest including additional variables and using weights to address the bias in predictor structure and mean levels.