Misclassification in multilevel models with application to dental caries research
Misclassificatie in multilevel modellen met toepassing op tandbederf onderzoek
Mutsvari, Timothy; S0195227
The main aim of this thesis was to understand more the misclassification process in detecting the presence or absence of CE while taking into account the multilevel data structure. We suggested possible ways of correcting for misclassification using validation data sets.In Chapter 1 we gave a general introduction of misclassification errors. We focused more on existing literature for adjusting for misclassification errors in statistical models.The statistical approaches explained in this thesis were applied to dental caries research. Hence in Chapter 2 we introduced general information concerning dental caries research, e.g. tooth decay process and diagnosis of CE. In this chapter we also introduced the Signal Tandmobiel study, which motivated us to carry out this research.In Chapter 3 we reviewed the general concepts of frequentist and Bayesian approaches to estimate the model parameters.In Chapter 4 we have presented the multilevel models for SE and SP. We investigated the factors that influence SE and SP as a means of assessing examiners' scoring behavior. In this chapter we also emphasized on the importance of taking the multilevel structure into account. In the absence of a gold standard, SE and SP cannot be estimated. Instead the kappa statistic, is often used as a reliability measure to assess the agreement of examiners. Hence, in Chapter 5 we proposed a hierarchical kappa statistic which is used to assess examiners'agreement when scoring data that have a multilevel structure.In Chapter 6 we focused on the use of external validation data to correct for misclassification in the main data set. Misclassification errors in external validation are often different from those from main data. Hence in Chapter 6 we proposed an approach for using external validation data, in a proper manner. The analysis was done in a multilevel context.In addition to multilevel structure, CE data are spatially correlated, i.e. a carious surface may influence the decay process of the neighboring surfaces. Theautologistic regression model is a popular choice for modeling spatially dependent binary data under the assumption that data are misclassification free. Hence in Chapter 7 we extended the multilevel model to also account for spatially correlated observations while correcting for misclassification.In Chapter 8 we focused on the estimation of prevalence and incidence from longitudinal CE data. We propose a binary Hidden Markov Model (HMM) for the analysis of longitudinal CE data subject to misclassification. The model expresses the prevalence and incidence as a function of covariates while taking into account missingness.Finally, in Chapter 9 we gave general conclusions. We highlighted the contributions of our research to statistical methodology and dental caries research. We ended the chapter by suggesting areas of needs further research.