In the interview with statistician Susan Murphy and earth scientist Brendan Meade, I was interested by the difference between detection of current events and prediction of future events, along the spectrum of frequent to rare events. For the prediction of rare events such as major earthquakes, Meade uses a larger quantity of smaller earthquakes to better understand the physics of earthquakes and provide more data for larger earthquakes. On the other side of the spectrum, Murphy detects more frequent events, such as when you are stressed. However, she notes that detection is much easier than prediction of future stress. As a result, work in mobile health is experimenting with various intervention plans to help predict and influence future health conditions such as stress.
A question I had for Meade is based on modern data collection of the millimeter-scale movements of tectonic plates and data on smaller earthquakes, how accurate and reliable are simulations of large earthquakes in predicting the location and time of future major earthquakes? While he mentions that a clear geodetic precursor has not yet been identified for earthquakes, can simulation help identify one?
Another question I had for both Murphy and Meade is how important is interpretability of model predictions in both mobile health and predicting earthquakes and natural disasters? If accuracy and fit is high but the features are not interpretable, is that sufficient given appropriate responses and interventions? Or should people have some understanding of why the model is predicting that they are at risk of harm? It seems that society can help shape this attitude, whether people “blindly” trust the models as long as the results are good, or whether they want some understanding and “control” over intervention.