On the use of data collection process information for the treatment of unit nonresponse through weight adjustment - ARCHIVED

Nonresponse weight adjustment is commonly used to compensate for unit nonresponse in surveys. Often, a nonresponse model is postulated and design weights are adjusted by the inverse of estimated response probabilities. Typical nonresponse models are conditional on a vector of fixed auxiliary variables that are observed for every sample unit, such as variables used to construct the sampling design. In this note, we consider using data collection process variables as potential auxiliary variables. An example is the number of attempts to contact a sample unit. In our treatment, these auxiliary variables are taken to be random, even after conditioning on the selected sample, since they could change if the data collection process were repeated for a given sample. We show that this randomness introduces no bias and no additional variance component in the estimates of population totals when the nonresponse model is properly specified. Moreover, when nonresponse depends on the variables of interest, we argue that the use of data collection process variables is likely to reduce the nonresponse bias if they provide information about the variables of interest not already included in the nonresponse model and if they are associated with nonresponse. As a result, data collection process variables may well be beneficial to handle unit nonresponse. This is briefly illustrated using the Canadian Labour Force Survey.