Pär Karlsson: Measurement errors in panel surveys, evaluation of Markov Quasi-Simplex and Markov Latent Class models

You might think that you need almost perfect replications in order to quantify measurement errors, but there are methods that are capable of quantifying measurement errors in longitudinal studies, e.g. panel surveys, even when the objects under study might have changed over time. Markov Quasi-Simplex models are useful for continuous data and Markov Latent Class models are useful for categorical data. The separation of measurement errors from the natural change is possible if certain key assumptions are met: the change should be in some sense stable (follow an autoregressive process or Markov process), and the measurement errors should be independent. Only the collected data within the survey is used by these methods, thus the costs for applying these methods are negligible. Measurement errors have been modelled in three surveys conducted by Statistics Sweden: the categorical classification of Employment status in the Labour Force Survey (LFS), the invoice values of arrivals and dispatches within the European Union (Intrastat), and the number of employees in the Short-Term Employment Survey (STES). The sampling units in LFS are people, while enterprises are sampled in Intrastat and STES. As these surveys have been ongoing for many years, and they publish results every month or quarter, time series of the quantified measurement errors (or more precisely estimated reliability coefficients in the quasi-simplex models and classification probabilities in the latent class models) have been constructed, covering more than 5 years. Except for the total number of employees in STES, the behaviour of the time series indicates deviations from the basic first order Markov assumption. Explanations could be extra within-object variability either between time points or between seasons. In the case of extra within-object variability, it will be impossible to separate the measurement error from the extra variability, so the estimated reliability coefficient will be lower than the true reliability. For the total number of employees in STES the assumptions seemed to work. Moreover it was possible to verify that the estimated reliability was higher for data that had undergone the normal editing process compared to the raw uncorrected data.