ABSTRACT
Due to decreasing response rates during the past few decades (Schnell 1997, pp. 36ff.; Rao and Pennington 2013, p. 651), a lot of research has been conducted about the issue of non-response bias, which is of particular concern when the missing data mechanism is Missing Not At Random (MNAR) rather than Missing At Random (MAR) (Schafer and Graham 2002, p. 152).1 When the mechanism is MAR, the cause of nonresponse Z is correlated with the survey variable Y (e.g. attitudes towards poverty), but this correlation can be explained by a vector of characteristics X, which are observable for both respondents and non-respondents (e.g. age and gender of sampled persons). In that case, the non-response bias can be corrected by weighting. When the mechanism is MNAR, the correlation between Y and Z cannot or only partly be explained by X (see Figure 14.1). Thus, correction of non-response bias becomes much more difficult.
