One reason is that it is theoretically possible to use memory instead of computation to attack the combined hashes by pre-generating a large number of collisions under one algorithm and then simply checking those using the other one, which means you don't need to do both algorithms for every check. Can't say for sure if that works out cheaper in terms of money but if the memory is available it could definitely save a lot of time.
Another reason is that for any given hash its theoretical maximum strength against any attack will be less than or equal to its bit length, but the practical strength always trends lower over time as attacks are found, and having two algorithms to attack increases the chances of finding flaws.
