Even without the purity question there are other things to consider. ML is evaluated strictly, for example. I don't know much Haskell but I don't believe it has any direct counterpart to ML's functors, either. There are probably many other differences. Haskell is definitely more popular and is still a niche language, so the reason you don't hear much about ML's benefits is probably just because it has so few users.
"...we are placing a strong emphasis on verification and proof as tools for the practicing programmer, chief among these being the infamous hellhounds of Computer Science, induction and recursion."
andrejbauer wants to know why Prof Harper is using SML instead of Haskell. There is a slight technical difficulty about proofs in a lazy language but "Well, in Haskell you’d have to use coinduction rather than induction, since the recursive definitions of datatypes are understood as final coalgebras."
That seems to make the point that ML is easier due to synergies with learning to create inductive proofs in maths lessons.
Two are two particular issues: laziness and the type system. Laziness is both complicated in practice and makes reasoning about the evaluation semantics more challenging. It can also lead to unexpected behavior for those unaccustomed to it (which will be nearly everybody). Furthermore, the type system is much more complex which makes it a bad model for learning about implicit typing and type inference.
However, for some programs, having a type-system that is fussy about side-effects imposes a significant effort burden for a meager correctness payoff. (For other programs, this bean counting is a huge win.) Also, I find it tricky to reason about the performance of lazy programs.
OCaml (and presumably SML) hit a practical sweet-spot in the statically-typed functional programming world.