Accounting for informative item nonresponse in biomarkers collected in longitudinal surveys (WP3)
A growing number of longitudinal surveys now incorporate the collection of biological data such as blood, saliva, and anthropometric measurements. While these types of biomarkers offer enormous potential for addressing important public health and social policy questions, many survey respondents decline to provide this kind of data for a variety of different reasons. This type of item non-response is likely to be ‘informative’, in the sense that measurements are missing for reasons that are related to the unobserved values. For instance, anxious individuals may decline a request to provide a blood pressure measurement because they are worried about their health. Older adults who are experiencing age related memory loss may refuse to do tests of cognitive ability in case it provokes anxiety. Existing methods to take account of missing data include propensity weighting, multiple imputation, selection models, pattern mixture models and sensitivity analysis.
In addition, longitudinal surveys also suffer from other forms of non-response at various stages of investigation which may have different non-response mechanisms associated to them. For instance, respondents may decline to be part of the survey at the first waves, in subsequent waves (called attrition) or at the nurse visit. Methodological research and development in this work package will aim take into account the different stages and mechanisms that can lead to missing biomarkers.
To compensate for missing data under different non-response mechanisms and evaluate the effectiveness of the methods, there is a need for auxiliary information and external criterion against which they can be compared. Auxiliary information, such as administrative data, gives the possibility of improving the models dealing with missing data. Nevertheless linking to administrative databases requires permission from respondents and is therefore an additional (possibly correlated) source of bias. Utilising linked administrative data to account for missingness in social surveys with biomarkers poses a number of methodological challenges such as the need to compensate for linkage error.
