@phdthesis{bergrab2026_113953,
  title = {Selection of models and variables : an automated Bayesian view for handling missing values},
  author = {Bergrab, Michael},
  year = {2026},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-113953},
  urn = {urn:nbn:de:bvb:473-irb-113953x},
  url = {https://fis.uni-bamberg.de/handle/uniba/113953},
  language = {eng},
  abstract = {This dissertation addresses the challenges of model and variable selection in statistical analysis, with a particular focus on datasets with missing values. Adopting a Bayesian framework, it introduces an automated approach that explicitly accounts for uncertainty due to missing data while enabling robust selection of relevant variables. The core methodology integrates data augmentation and Markov Chain Monte Carlo (MCMC) techniques to jointly estimate the posterior distributions of parameters and missing values.
The work bridges theoretical foundations of Bayesian inference with practical applications, including the use of spike-and-slab priors for variable selection and the analysis of binary regression models with missing covariates. Empirical studies and simulations demonstrate the superiority of this approach over classical methods such as multiple imputation or shrinkage estimators (e.g., Lasso, Ridge, Elastic Net). The results indicate that the Bayesian approach not only enhances estimation accuracy but also improves model interpretability; especially in high-dimensional and high-missingness scenarios.
A key contribution of this work is its systematic approach to model selection, which relies on Bayes factors, marginal likelihoods, and information criteria (e.g., BIC) to compare competing models, even in the presence of incomplete data. These methods identify the model that strikes the optimal balance between fit and complexity, while model averaging and hierarchical priors ensure stability across different missingness mechanisms (MCAR, MAR, MNAR).
This dissertation contributes to statistical methodology by providing a coherent and flexible solution for the simultaneous handling of missing values, variable selection, and model comparison. It emphasizes the importance of accounting for uncertainty in data analysis and offers practical guidelines for implementation in empirical research.},
  keywords = {Bayesian estimation, Handling missing values, Variable selection, Model selection},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{bahrami2021_49487,
  title = {Missing by Design Patterns for Optimizing Survey Response by Efficient and Consistent Data Collection},
  author = {Bahrami, Sara},
  year = {2021},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-49487},
  urn = {urn:nbn:de:bvb:473-irb-494874},
  url = {https://fis.uni-bamberg.de/handle/uniba/49487},
  language = {eng},
  abstract = {Respondent burden due to long questionnaires in surveys can negatively affect the response rate as well as the quality of responses. A solution to this problem is to use split questionnaire design (SQD). In an SQD, the items of the long questionnaire are divided into subsets and only a fraction of item-subsets are assigned to random subsamples of individuals. This will lead to several shorter questionnaires which are administered to random subsample of individuals. The completed sub-questionnaires are then combined and the missing values due to design are imputed by means of multiple imputation method. Identification problems can be avoided in advance by ensuring that the combination of variables in the analysis model of interest are jointly observed on at least a subsample of individuals. Furthermore, including an appropriate combination of items in each sub-questionnaire is the most important concern in designing the SQD to reduce the information loss, i.e. highly correlated items that explain each other well should not be jointly missing. For this reason, training data must be available from previous surveys or a pilot study to exploit the association between the variables.

In this thesis two SQDs are proposed. In the first study a potential design for NEPS data is introduced. The data consist of items which can be divided and allocated into blocks according to their context, with the objective that the within block correlations are higher relative to the between block correlations. According to the design, the target sample is divided to subsamples. In addition to the items of a whole block which is assigned to each subsample, a fraction of items of the remaining blocks are randomly drawn and assigned to each subsample. Where items that belong to blocks with relatively higher correlations are drawn with lower probability. The design is evaluated by means of several ex-post investigations. The design is imposed on complete data and several models are estimated for both complete data and data deleted by design. The design is also compared with a random multiple matrix sampling design which assigns random subset of items to each sample individual.   

In the second study, a genetic algorithm is used to search among a vast number of SQDs to find the optimal design. The algorithm evaluates the designs by the fraction of missing information (FMI) induced by the design. The optimal design is the one with the smallest FMI. The optimal design is evaluated by means of several simulation studies and is compared with a random MMS design.},
  keywords = {Planned missing design, Split questionnaire survey design, Multiform design, Multiple imputation, Multiple matrix sampling design, Missing data patterns, Optimizing survey response, Genetic algorithm},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{schnapp2019_46497,
  title = {Investigation of Automated and Augmented Covariate Imputation during the Estimation of two Confirmatory Multidimensional Bayesian Latent Variable Models},
  author = {Schnapp, Thorsten},
  year = {2019},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-46497},
  urn = {urn:nbn:de:bvb:473-irb-464970},
  url = {https://fis.uni-bamberg.de/handle/uniba/46497},
  language = {eng},
  abstract = {The purpose of this thesis is to investigate parameter estimation and simultaneous imputation of covariates for two multidimensional Bayesian latent variable models. These models encompass a multivariate confirmatory factor analytic model and a multivariate normal-ogive item response theory (IRT) model. The basic idea for the research conducted stems from the positive findings for IRT models with one latent variable of Aßmann et al. (2015) and Gaasch (2017) respectively, and the research of Preising (2018) concerning dynamic linear panel models, with presenting, aside from estimation, a relief for the burden of determining an imputation model for missing values in covariates by using an automated procedure, i.e. imputation via a classification and regression tree approach (CART).

The models considered within this thesis and later to be extended, originate from the class of confirmatory factor analytic and normal-ogive IRT settings, who find application in psychometric modelling, social and political sciences, medicine and many areas more. Their features get described by means of discussing concepts, first considerations with respect to estimation within a Bayesian framework, mathematical model formulations, different terminologies used in varying contexts, and an overview of considerations with respect to model and parameter identification.

By comparing different latent variable models (i.e. Aßmann et al., 2015, Gaasch, 2017, and Conti et al., 2014) and their application, ways of including covariates and handling of missing values get discussed. Using Conti et al.'s (2014) model as a basis, our derivations transform their explorative factor model into a confirmatory matrix-variate factor and a normal-ogive IRT model with several latent variables and incorporating covariates in two different ways. In order to handle missing values within these covariates, CART imputation is employed. Next to binary and metric scaled measurements, estimation of ordered scaled items is facilitated with the help of an algorithm of Albert et al. (1997), which was already implemented and tested by Gaasch (2017) in his research on IRT models with one latent variable. In order to allow the possibility of defining restrictions, the indicator matrix employed by Conti et al. (2014) gets extended to specify equality and multiplicative structures for loadings and item discrimination parameters respectively. In addition to the dedicated loading structure of Conti et al. (2014) this matrix can be used to specify cross-loading, such that a measurement can be related to several latent variables simultaneously.

In order to test the proposals made, nine factor analytic and twelve IRT model related simulation studies, and an empirical application to competency data stemming from the National Educational Panel study (see Blossfeld et al., 2011) are performed. The findings indicate the general ability to conduct the estimation and benefits for using the CART imputation within the IRT approach. However the factor analytic specification, although in general estimable without missing values, does not provide reasonable results compared to complete-case analysis when using the CART imputation approach. Thus, future directions of further research can be investigating reasons and alternatives for the shortcomings with respect to the factor analytic approach and general extensions to the IRT model.},
  keywords = {Multidimensional Latent Variables, Classification and Regression Trees, Bayesian, Gibbs Sampling, MCMC, Item Response Theory, Factor Analysis},
  school = {Otto-Friedrich-Universität Bamberg}
}

