Mmh not really 
Maybe because you just give a probability vector of shape 1 to the likelihood instead of a vector of length len(outcome) as before? This maybe takes advantage of vectorized operations under the hood?
Really not sure but glad you found a solution 