This three day course will teach you advanced topics in multilevel modelling: multilevel generalized linear models for categorical outcome data, Bayesian inference, and missing data and imputation. The three-day course builds upon the contents of the other summer school course “Introduction to multilevel analysis”. It consists of three days with lectures in the morning and computer labs in the afternoon.
The focus of the first day is on categorical outcome data, in particular binary, ordinal and event history outcomes. It will be shown why linear multilevel models are not appropriate for such data and how multilevel generalized linear models can be used to fit this type of outcome data. Attention will be paid to estimation procedures that are available and how the intraclass correlation coefficients and proportions explained variance are calculated. Special attention is paid to the interpretation of the estimated regression weights in terms of the logits and odds ratios.
The focus of the second day is on using Bayesian estimation instead of Maximum Likelihood estimation. Often Bayes is used because of a limited cluster size, or because researchers may be “forced into” the implementation because some complex models simply cannot be estimated using other approaches or result in estimation/convergence issues. Another category of reasons to use Bayesian estimation is because it is appealing to incorporate knowledge into the estimation process via priors. For example to rule out negative (residual) variance estimates, the use of approximate-zeros instead of constraining parameters to be exactly zero, or simply because researchers believe in the Bayesian paradigm of updating knowledge instead of testing the null hypothesis over and over again. During the second day, participants are gently introduced to Bayesian statistics and its specific advantages (and disadvantages) in the field of multilevel analyses. Note that this day does not assume previous knowledge of Bayesian estimation and we will only conceptually introduce Bayesian estimation. For a full course on Bayesian estimation you can subscribe here.
Missing data form a ubiquitous problem in scientific research. Ignoring the unobserved values, or insufficient treatment of the missingness problem can have a tremendous impact on our estimates. On day three participants will therefore learn what techniques can be used to solve for the missingness problem in multilevel data. We will use multiple imputation (package mice in R) to treat missing values and obtain inference on incomplete multilevel data. The mice package is a world-leading, flexible software package for multiple imputation. Participants will learn the practical workflow of solving multilevel missing data problems, as well as the common caveats of missing data theory in multilevel problems. Moreover, participants will learn how to incorporate package mice in their own workflow and how to prepare imputed datasets for analysis in their preferred statistical analysis software package. For a full course on missing data you can subscribe here.
It is expected participants have taken the course Introduction to Multilevel Analysis or a similar course with the same contents (i.e. chapters 1-5 from Hox, Moerbeek and Van de Schoot (2018). Participants are also expected to have experience with analyzing multilevel data in common software such as Mplus, SPSS, R, HLM, or MLwiN.
Hox, J., Moerbeek, M., & Van de Schoot, R. (2018). Multilevel analysis. Techniques and Applications. 3rd edition. New York: Routledge.
Book is NOT included in fee (about 45 euros)
Irma Reyersen - email@example.com