It has to do with the design of the network in question. The benefits you see of an area becoming a stub (reduced LSDB size) comes at a cost, which is the loss of some routing information details. This translates into routers within a stub not having all the information necessary to make the best possible choices.
For example, suppose you have an area (which is non-zero, of course), that has multiple exit points. Now imagine that at each of those exit points there are separate external routing domains (say, EIGRP or BGP, etc). If this area is a stub, Type-5 LSAs are not allowed. This means that routers inside the stub won’t have the visibility to know the least cost path to get to external network X is through exit point 1, while getting to network Y, it is best to get through exit point 2.