The promise that causal modelling can lead to robust AI generalization has been challenged in recent work on domain generalization (DG) benchmarks. We revisit the claims of the causality and DG literature, reconciling apparent contradictions and advocating for a more nuanced theory of the role of causality in generalization.
Suppose we are interested in predicting the diagnosis of lung cancer. It is well established that smoking causes lung cancer, but the relationship may be confounded by age, socio-economic status or occupation. Patients diagnosed with lung cancer often experience significant weight loss.
At present people aged 55 – 74 and have ever smoked are screened for lung cancer. Otherwise, patients are only tested for lung cancer if they present to the GP with symptoms. Consider the screening criteria was widened to include a wider range of ages and people who have never smoked. This could then cause some patients to be diagnosed sooner and at early stages of the disease. As such patients may not be as sick and the average rate of weight loss changes. In this case we may see a distribution shift in the weight loss variable.
Suppose we wish to predict the onset of chronic kidney disease (CKD). Type 1 diabetes is one of the leading causes of CKD. However, both diseases are commonly caused by high blood pressure. Type 1 diabetes is a chronic condition that requires constant vigilance on blood glucose levels, which some patients struggle to manage.
An insulin pump is a small electronic device that releases the regular insulin your body needs through the day and night. Insulin pumps are not widely available on the NHS to all patients with type 1 diabetes. For many this means that they would need to self-fund an insulin pump. However, if they became more widely available on the NHS then people would be able to better manage their diabetes and we may see a distribution shift in blood glucose levels.