Machine Learning + Causal Inference:  A new model building strategy for big data? Experiences from air-pollution research

Associate Professor Theis Lange
University of Copenhagen

Traditional epidemiological model building is based on specifying the functional form linking exposure variables to outcomes; typical a linear relation or linear with a few interactions. This provides models that are easy to interpret, but imposes great many assumptions and is unstable for highly correlated exposure variables. In contrast, machine learning methods avoid the need for such limiting assumptions. As the catchphrase goes: we let the data speak. However, despite this catchphrase machine learning methods can be extremely difficult to interpret – we cannot understand what the data says.
In this talk I will propose a method to get the best of both worlds. The proposed method is illustrated by applications to household specific air-pollution at high temporal resolution from Denmark.