Answer refused: Exploring how item non-response on domestic abuse questions in a social survey affects analysis and results

We explore the pattern, potential drivers, and implications of item non-response on survey questions about domestic abuse. We draw on a longitudinal representative prospective survey on children and their families in Scotland (N:3646) and use multivariate regression models to look at non-response on domestic violence questions among mothers of young children. By triangulating data from multiple survey sweeps we hypothesise that item non-response may be due to mothers experiencing violence, and we observe that factors which predict experiencing violence also predict item non-response. We compare conservative and generous dependent variables on domestic abuse prevalence and find that both yield similar results in multivariate models, but that the actual social gradient of domestic violence is likely to be steeper than we can see in survey data. We discuss the ethical implications of imputing missing data and argue that sometimes it is unethical to do so.