Methods of Weighting for Unit Non-Response

Abstract

Despite the best efforts of survey practitioners, non-response is a persistent problem for survey researchers as a potential source of bias for survey estimates. This paper is concerned with unit non-response and the various methods of weighting which are used in an attempt to reduce non-response bias. The methods may be categorized as follows: (i) using auxiliary population information; (ii) using auxiliary information for the intended respondents; (iii) using no auxiliary information. The methods are placed in a single framework defined by weighting classes and their properties compared. Considerable importance is rightly attached to efficient, cost-effective survey design. An important aspect of this is the careful design of samples, based on random selection with known probabilities of the individuals in a population. This provides a target sample of intended respondents, each of whom may provide answers to a set of survey questions and hence yield a vector of responses. Non-response occurs when some of these intended responses are missing. We distinguish between unit non-response, when the entire vector of required responses is missing for some sampled units and item non-response when responses are obtained to some questions and not to others. This paper provides a review of weighting methods that can be used at the analysis stage to compensate for unit non- response. The fundamental problem arising from unit non-response is that the sample design breaks down and, in particular, the selection probabilities are changed in an unknown way. In the usual sample survey framework it is the sample design that provides the link between sample and population and is the basis of statistical inference. To the extent that non-respondents are not missing at random, the result is that any survey estimators will be potentially biased. Causes of unit non-response are many and varied and are related to the organisation and procedures used in the field-work and to the subject of the inquiry, as well as to the characteristics of the selected units. The main types are, of course, non-contacts and refusals, both of which are susceptible to differences in field methods. The best approach to the problem of unit non-response in surveys is to make strenuous efforts to minimise it in the first place. To minimise non-contact, survey organisations instruct interviewers to call back on selected households or individuals on a number of occasions, at different times and days of the week and where possible to make appointments if an interview cannot be obtained immediately. The acceptance of proxy interviews in appropriate cases will also reduce non-contacts. To allow sufficient time for recalls, the length of time allowed for field-work in an area must be set realistically high if non-contact rates are to be kept low. Much can also be done to minimise refusals. One of the most important approaches is through the quality control and motivation of interviewers. The training of interviewers should emphasise the importance and expectation of good response rates and it is essential