Non-response subgroup-tailored weighting: The choice of variables and the set of respondents used to estimate the weighting model

The use of the logistic regression model to predict the probability of response and create non-response weights is classic. In most cases, the model is estimated using socio-demographic variables and all units in the selected sample. However, substantive analyses are often restricted to a sub-group of the sample. This paper investigates whether weights are more effective if they are designed using variables correlated with the response propensity in the sub-group in question and sample units in the selected sub-group using data from the British Household Panel Survey (BHPS). The findings demonstrate that, for some estimates, the tailored weights results in significantly different results than the usual weights.