SparSNP — Supplementary Information

where yi ∈ {−1,+1}. SparSNP uses the squared hinge loss as the classification model. We use the squared-hinge loss rather than the standard hinge loss since the former is twicedifferentiable whereas the latter is not. Twice-differentiability means that the Newton stepsize used in coordinate descent is the second derivative of the loss wrt βj , whereas to use coordinate descent with the hinge loss requires some pre-chosen step size, which must be found through manual tuning or a line search procedure. The disadvantage of the squaredhinge loss is that it is potentially more sensitive to outliers in the data, with the loss increasing quadratically for mis-classified samples, whereas the hinge loss increases the loss only linearly.