That's not really true. In fact, practical collision attacks vs. MD5 preserved the size - which if you think about it, makes sense. Unless a has function is utterly terrible, it's not unreasonable to assume it's easier to find a few correlated bits than it is to append bits. After all, to append bits you're going to need to somehow ensure that the internal fixed-size state of the hash algorithm cascades into the same state or nevertheless gives the same output, and since it's trivial to have a very, very high period state machine, that may mean appending huge numbers of bits.

If safety were an argument you'd add an extra, unrelated hash function. E.g. even md5 is likely much harder to break if you also have the CRC32, even though CRC is a thoroughly insecure hash (and of course, you wouldn't use an insecure hash, now would you?)