In the current age of open science, sharing research code and data is often required when publishing a scientific paper. Moreover, the open dissemination of research data is a potential gold mine for answering many research questions. However, privacy and confidentiality constraints often impede the open dissemination of research data. Synthetic data can be an excellent solution to this problem: the real data is kept secret, but a "fake" version of the data is made available. This synthetic dataset can serve many purposes. For example, it allows those in the process of obtaining access to the real data set to get familiar with the structure of the data, and it allows reviewers (or other researchers) to rerun scripts and assess whether the original analysis code is reproducible and runs as intended. Additionally, the synthetic data itself can be used to run completely different analyses, unrelated to the original research problem. In this course, you will learn what synthetic data is, how to generate synthetic data, how to evaluate its quality in terms of utility and remaining privacy risks, and how to obtain statistically valid results from analyses on this data.

 

In three half days, we will cover the origins of synthetic data (including its relation to multiple imputation of missing data), practice generating our own synthetic version of a realistic scientific dataset, and evaluate its quality and disclosure risks. We will discuss how to make inferences from synthetic data and work on increasing the synthetic data quality through advanced modeling or solving practical problems that arise when working with complex data structures (for example, how to deal with deterministic systems/composite variables or logical constraints). On the final day, there will be room for individual consultation.

 

The course will have a hands-on format, with more time scheduled for practicals (+ discussion) than lectures (approximately a 60/40 division). In principle, (social) scientific datasets are provided for all practicals, but participants can also bring their own data (this might not be ideal if (1) this data is so privacy-sensitive that instructors cannot look at it; (2) the dataset is so large that running code takes too long). All practical exercises are in R, but only little programming experience is required (a recent 'introduction to R'-course or some working experience with R or another scripting language suffices). A good understanding of basic statistics will definitely be beneficial (i.e., working experience with regression analysis).

 

Organizational structure of the course

Each day will consist of two blocks of two hours, containing a live lecture of approximately 45 minutes and a hands-on practical (that can be completed individually or in small groups), and a discussion of approximately 60 minutes. The lecturer will be available for questions during the practicals. On the last day, there will be some time for individual consultation (but project-related questions can also be asked during breaks or before/after class).

Software requirements