Fig. 3: A more complete model of disease, including a predicted target Y.
From: Why do probabilistic clinical models fail to transport between sites
Conditioning on observations \(O\) (shaded green) to produce \(P(Y{\rm{|}}O)\) allows the prediction to be affected by site T and selection variables S. In contrast, conditioning on the latent but true causes and effects (shaded red) to produce P(Y|C,E), does not. As in other figures, arrows represent the direction of causality, not the direction of inference. Variable names as in Table 1.
