Sure sampling bias is probably a factor. But I'm sure it's also a factor in nearly every one of these discussions that isn't a serious and meticulous study. Fo

Sure sampling bias is probably a factor. But I'm sure it's also a factor in nearly every one of these discussions that isn't a serious and meticulous study. For wondering, I believe this might help. If we parametrize sigmoid to be 1/(1+e^{-ax}) then our a factor is at play. We're interested in x<=0. When a < 1 then the function becomes flat and much more linear like. The larger a becomes the more quadratic it looks at near the origin (but still x<=0) I still think the model is naive. Even just a sigmoid is. You need the composition of them and further parameterization. They are good for approximating as you originally suggested but this is similar to using fourier series. It's the fact that you have the three classes of second derivatives allowing you to strongly control the composite function through minimal parameterization and complexity. But that's also why I immediately understood this to be being using in an illustrative way rather than a serious one. I mean realistically so are you.