Diana Karimova

Tilburg School of Social and Behavioral Sciences
Methodology and Statistics
Tilburg University

Project

Multinomial Choice Model for Relational Events Networks
My PhD project is devoted to developing exploratory methods for better understanding temporal features of dynamic interaction processes, such as speed, lag, rhythm, pacing, and memory. In this working project, we will analyze the relational event history of 70,000 information-sharing events between colleagues through email in large organizations. The interest in the study is in how (fast) new information flows through the network, whether there is a lag between members in different teams, whether there is a specific rhythm of information sharing events, and how this changes in continuous time during the period of working projects. External information of the employees is available such as team membership (which changes over time), common working projects including important deadlines, the time of employment, the department and location, among others. All this information can be interpreted as a covariate that influence the process of interaction. For example external covariates such as gender, department, location, hierarchical level, religion, time of employment, ‘activity’ of the employee to send messages, ‘popularity’ of the employee as receiver. Network effects are dynamic in nature: If teams have to work together to meet a certain deadline, it is likely that the effect of team membership dies out near a deadline. To explore such temporal properties of the network effects, we will use Bayesian MCMC methods to fit the model sequentially over the event history sequence. Through visualization we will learn about the dynamic nature of network drivers and temporal aspects of the dynamic interaction process. There are endless possible internal and external covariates, and interactions that can potentially drive the temporal interaction process. Therefore overfitting and false positives can be serious issues. To account for this, I will develop shrinkage priors, such as the Bayesian Lasso and horseshoe priors to shrink nonexisting effects towards zero and leave true effects unaffected.

Supervisors
Prof. dr. R.T.A.J. Leenders, dr. ir. J. Mulder

Financed by
ERC Grant

Period

1 September 2018 – 31 August 2022