Evaluating reliability of combined responses through latent class models

The evaluation of the potential impact of the response errors on the final survey estimates requires ad hoc studies. Often these studies consist in additional reinterview surveys: a subsample of the respondent units at the main survey is interviewed again. In such cases, the evaluation can be done by means of the theory introduced by Hansen et al.(1964) and further investigated in Biemer and Forsman (1992). More recent studies (Biemer, 2004) present an approach based on the fitting of latent class models. These models allow for a more detailed analysis of the impact of response errors on the final survey estimates but, on the other hand, they require some additional assumptions to hold. In this paper, the usage of latent class models is extended to tackle the case of couples of survey questions involved in a questionnaire skip. An application of such models to the data of the control survey on the 2001 Population and Housing Census in presented.