Investigating the Use of Nurse Paradata in Understanding Nonresponse to Biological Data Collection

The recent collection of biological data in large-scale sample surveys has opened up new possibilities for research into the interactions between physical and social mechanisms in the general population. Whilst the possibilities are undoubtedly exciting, these data can create additional challenges from the viewpoints of both collection and analysis. In particular, the extra burden of biological data collection can lead to increased incidences of nonresponse, potentially affecting the quality of the data and the robustness of results from subsequent analysis. Where the two-stage nurse visit survey design is used, such as in Understanding Society (UKHLS) and the English Longitudinal Study of Ageing (ELSA), the possible effects of the assigned nurse on patterns of nonresponse also need to be considered. In order to address nonresponse to biological data collection, researchers can build response propensity models and use these to adapt future data collection and/or produce post-survey adjustments such as weights. However, variables included in the models must be available for both respondents and nonrespondents in the sample. A recent branch of research concerns the use of new forms of data collected during the survey process, known as paradata, for this purpose. Paradata can include variables such as call histories, response timings and interviewer observations, and may provide a cost-effective way to deal with nonresponse in analyses which use survey data. In this paper, we use paradata collected during the nurse visit in UKHLS to investigate nonresponse to biological data collection and also to examine and explain the effects of the nurse. Preliminary results from UKHLS wave 2 have shown that clustering by nurse is present in responses to three conditional stages: participation in the nurse visit, consent to the blood sample and obtaining the blood sample. When quantifying the nurse effects, we estimated cross-classified multilevel models for the likelihood to respond to each of these stages to take account of the effects of geographical area. The models included a comprehensive set of explanatory variables for the respondents and households as well as data from call records. Given the amount of missing paradata, we also used the availability of call record data to estimate nurse performance indicators that could be used in response propensity models for the following wave of biological data collection. Results indicate that the nurse performance indicators derived from the availability of call record paradata are predictive of the likelihood of a sample member to participate in the first stage of the nurse visit in the following wave, given nurse age and experience. This was not the case for the subsequent two stages (consent to and obtaining the blood sample). This suggests that the recording of call histories by nurses in surveys may be an indication of their performance in the interviewer-type tasks involved in biological data collection, such as making contact and gaining cooperation from sample members. It also shows how these new forms of data may help to explain nurse effects in models of response to biological data collection