Estimating mode effects from a sequential mixed-mode experiment using structural moment models

Until recently, the survey mode of the household panel study Understanding Society was mainly face-to-face interview, but it has now adopted a mixed-mode design where individuals can self complete the questionnaire via the web. As mode is known to affect survey data, a randomized mixed-mode experiment was implemented during the first year of the two-year Wave 8 fieldwork period to assess the impact of this change. The experiment involved a sequential design that permits the identification of mode effects in the presence of nonignorable nonrandom mode selection. While previous studies have used instrumental variables regression to estimate the effects of mode on the means of the survey variables, we set up a more general framework based on novel structural moment models to characterize the effect of mode on the distribution of the survey variables by its effect on the moments of the joint distribution. We adapt our estimation procedure to account for nonresponse and complex sampling designs,and to include suitable auxiliary data to improve inferences and relax key assumptions. Finally, we demonstrate how to estimate the effects of mode on the parameter estimates from generalized linear and other exponential family models when both outcomes and predictors are subject to mode effects. This framework is used to investigate the impact of the move to web mode on Wave 8 of Understanding Society.