I must admit that I don't know the (real) answer, but I suspect I'm not alone.
As always it depends on the image quality and density you are trying to scale and whether it's to be scaled up or down.
Here are some snippits I've gathered over time from the web that have generally served me well (but not always).
Experimentation is sometimes the best tool.
Nearest Neighbor (Jagged) for the fast but less precise method.
This method is recommended for use with illustrations containing non-anti-aliased edges, to preserve hard edges and produce a smaller file. However, this method can result in jagged effects, which become apparent when distorting or scaling an image or performing multiple manipulations on a selection.
Bilinear for a medium-quality method.
It is best to use Bilinear to shrink images
Bicubic (Smooth) slow but more precise resulting in the smoothest tonal gradations.
It is best to use Bicubic to enlarge images
Bicubic is more precise, but only when it comes to enlarging. When it comes to shrinking its exactness can actually produces pixelation, because to shrink an image pixels must be discarded anyway.
This can give acceptable results for up to 2-3 times magnifications. Upsized images become softer, they need to be sharpened with the Unsharp Mask filter.