Science Jam #63: When Models Fail: Using Prediction Gaps to Reveal Hidden Social Structures

Thanks to the Young Complexity Researchers Utrecht (YCRU) group, they brought these popular sessions back to the CCSS for our researchers to discuss challenges in Complex Systems Studies. Everyone is welcome!

Therefore, we cordially invite you to the Science Jam #63 on Monday 23 June (12:00-13:00) at the Centre for Complex Systems Studies (CCSS) where you can:

• Get a free gourmet lunch with the best sandwiches you can get at the Utrecht Science Park plus nice drinks;

• Know one senior complexity researchers' work over one-hour lunch time;

• Contribute your professional knowledge and experiences in a relaxing and informal setting;

• Develop potential collaboration.

Leading complexity researcher: dr. Javier Garcia Bernardo, Methodology and Statistics, Utrecht University 

Abstract:

Decades of sociological research have extensively examined factors impacting socioeconomic outcomes, often emphasizing social contexts such as families, schools, or neighborhoods. However, understanding their relative importance is challenging because individuals are embedded in complex networks of these social contexts, which interact with each other and with individual characteristics to constrain or facilitate opportunities. In this talk, I introduce the concept of prediction gaps—differences in performance between statistical models of increasing complexity—as a diagnostic tool to identify where conventional theories and models fail to capture the hidden structure of social outcomes. Using population-scale register data from the Netherlands, we show that while parental characteristics dominate predictions of university completion, flexible models like gradient boosting and graph neural networks reveal meaningful discrepancies—especially for disadvantaged subgroups, such as girls without fathers or students from Turkish and Moroccan backgrounds. These gaps highlight where social processes are nonlinear, context-contingent, and likely under-theorized. By framing prediction errors as signals of structural complexity, this work bridges machine learning, sociology, and complex systems science.

Everyone is welcome and please feel free to invite your colleagues/friends/classmates/students to join us.

If you would like to have the lunch arrangement, please sign up before 15:00 Friday 20 June.