Speaker: dr.ir. Jean Paul Fox
Faculty BMS - Department of Research Methodology, Measurement & Data Analysis (OMD)
Keywords
- Covariance Structure Modelling (BCSM)
- Bayesian Statistics
- Multilevel Models/Mixed Effect Models
- Clustered (Response) Data
- Personalized Treatment (Application)
- Individual Treatment Effects (Application)
Title: A New Approach to Model Clustered Data: Bayesian Covariance Structure Modelling
Abstract:
Datasets in the behavioral, social and medical sciences are often hierarchically structured: variables describe individuals, and individuals (often) adhere to groups. Typically, data within a cluster (for instance, individual with repeated measurements, group of individuals) are correlated, and observations within a cluster are more alike than observations from different clusters. Traditionally, dependences within a cluster are statistically modeled through random effects. Conditional on the random effects observations within a cluster are assumed to be independently distributed. I will give a few examples showing that this modeling approach has several disadvantages (e.g., sample size restrictions, type of correlation, estimation).
A new modeling approach is discussed based on models for the covariance structure of the data. This so-called Bayesian covariance structure model (BSCM) has the advantage that dependences implied by random effects are directly modeled without needing to estimate the random effects. As the estimation of random effects is difficult with small data, this greatly improves the generalizability of conclusions. Furthermore, the number of parameters is drastically lower than the standard multilevel modeling approaches, while the interpretation does not change (which is beneficial for big data). Thus, BCSM allows for modelling of complex theories with either a limited or a huge data set.
In the presentation, different applications will be discussed to show the potential of the BCSM. The applications include modeling effects of personalized treatments when patients are treated by the same doctor, and identifying individualized treatment effects when patients receive the same treatment.