Kernel selection for the diagnostic to reflect estimators’ extrapolation behavior

Develop a principled methodology for selecting the kernel in the proposed kernel-based sparsity diagnostic so that it accurately characterizes the extrapolation behavior of the estimators (e.g., algorithms in a Super Learner library) used for outcome and treatment modeling.

Background

The diagnostic weights observations via a kernel to quantify local data support around intervention-induced prediction points. Its conclusions depend on the kernel shape and bandwidths, which are intended to mirror how learning algorithms extrapolate.

The authors note the lack of a principled, general approach to choose a kernel that realistically captures the extrapolation behavior across potentially diverse estimators within an ensemble, highlighting a methodological gap.