Accuracy of coalescent likelihood estimates: do we need more sites, more sequences, or more loci?