### Full text

1. INTRODUCTION TO THE NONRESPONSE PROBLEM 12 The second reason is the uncertainty about the consequences of nonresponse. There is generally practically no information about the typology of the sampled but not interviewed persons. Brehm“ (1993) calls them "the phantom respondents" (which is also the title of his book). Naming the problem is trivial: we do not know much about these nonrespondents because they did not answer our survey questions. Assessing the impact of this loss of information is anything but trivial, however. Groves/Couper" (1998:49) argue that the "biggest drawback in attempting to study nonresponse" is the fact that the people we are interested in are exactly the nonrespondents. Above all the question we are interested in is: Do the nonrespondents really differ from the respondents? If they appear quite similar, then the problem is negligible. Often we do not know this, however. I, therefore, call the nonresponse problem "the dark chapter". 1.4 Consequences of the Nonresponse Problem There are two important consequences of considering the impact of nonresponse: 1) Bias in univariate statistics Point estimates as means or proportions calculated on the basis of the respondents might be over- or underestimated. The bias problem could be ignored if the total nonresponse rate is low and negligible and the nonrespondents share the same characteristics as the participants (i.e. one can assume them to be missing at random). The bias problem is larger if one of these conditions is destroyed. The nonresponse bias is at its highest if we have both: a high nonresponse rate and nonrespondents differing from survey people. Groves/Couper (1998) demonstrate several examples of bias* of means. Brehm (1993:93-96) explains the bias contribution in the formula for the population mean. Estimates of variance based on the respondents' sample will generally underestimate the population variance which also means that statistical inferences might be incorrect. The mathematical formulas are given by Brehm (1993:97-100). 2) Bias in multivariate relationships In multivariate relationships, there is a danger of biased regression coefficients and underestimation of confidence intervals. The following graph shows a simple regression 2° Brehm, John (1993) The Phantom Respondents: Opinion Surveys and Political Representation. 21 Groves, R.M./Couper, M.P. (1998) Nonresponse in Household Interview Surveys. Lower and higher bias situations are visualised in Groves/Couper (1998), pp.4-5.