To reiterate a point implied by other commenters: macro-expansion is a compile-time operation. If the language utilizes a typing system to help promote sound programming then the results of the expansion will be compiled in the context of type constraints - the expansion is either sound or unsound w.r.t. such constraints.
To some extent there is a continuing popular schism with the idea that "strongly typed" must be synonymous with the lack of "eval", which, at least in principle, cannot be true. In fact, StandardML does have the ability to extend the environment of a program at run-time (as well as Typed Racket). All the necessary typing and meta-data is retained at run-time so that newly evaluated code is checked for type-consistency just like when the "main code" was compiled. So, not only can one expand macros at compile time and verify that the expansions are sound, one can also do the same at run-time.