Most studies of non-sampling errors in surveys employ interview/reinterview data to estimate and analyze response variances. Because the respondent is most often the focus of attention, these studies restrict the samples to those cases where, what O’Muircheartaigh (1986) terms, 'the essential survey conditions' are the same for both the interview and reinterview observations. The variances analyzed are, therefore, conditional on these essential survey conditions. One such condition is that the respondent actually is asked the question in both trials. In fact, however, a potentially important source of unreliability of survey data is that the procedures used in determining who gets asked which questions are themselves subject to error. Thus, the unconditional variances are also relevant in assessing the quality of survey data.

In this paper we assess the relative importance of response and procedural error variance in the Survey of Income and Program Participation (SIPP). We concentrate on the SIPP not because it is any more prone to error than other surveys, but rather because it has a good reinterview program and the staff has had the curiosity to encourage external scrutiny of these data. We conduct our assessment in two ways. First, (Section 3) we examine each of a series of questionnaire items from the SIPP Reinterview Program and present separate estimates for the response variance, the procedural variance, and the overall variance. Second, we investigate the correlates of response and overall error variance by modeling the entire interview/reinterview outcome as single experiment. A very general counting distribution, the Poisson-Pascal, is taken as a point of departure (Section 4) in analyzing the number of interview/reinterview discrepancies. This distribution subsumes a large family of more specific distributions (see Katti and Gurland, 1961) some which allow for heterogeneity and/or contagion (e.g. the Negative-Binomial or Pascal and the Neyman Type A) and some which do not (e.g. the Poisson). The object  of this preliminary analysis is to identify the most parsimonious distribution which adequately describes the data. Once this is accomplished, the model's parameters are treated as functions of the characteristics of the respondent and the interview situation and the effects of these characteristics on data quality are estimated.