From the standpoint of collision-resistance (finding two colliding messages) and second-preimage-resistance (finding a different message colliding with a given one), the concatenation of multiple hashes is at least as secure as the strongest of the hashes (Proof: for any of the two properties, any attack that breaks the concatenation can be turned into an attack that breaks each hash. For example, hypothetical colliding messages for the 640-bit MD5||SHA-512 also collide for SHA-512, thus if SHA-512 is collision-resistant, then MD5||SHA-512 is collision-resistant).
From the standpoint of finding a message hashing to some given random value, or finding an unknown random message known to hash to a given value (loosely speaking, first-preimage resistance), it is prudent to consider that the concatenation of multiple hashes is insecure. We can construct hashes so that each is individually resistant to finding an unknown message (chosen randomly in some set too large to explore) from its hash, but the concatenation is totally weak against that.
For many iterated hashes including all Merkle-Damgård hashes (thus MD5, SHA-1, the various SHA-2), using the concept of multicollisions, it can be shown that the concatenation of multiple hashes is not much more secure against collisions than the strongest of the hashes. See Antoine Joux: Multicollisions in Iterated Hash Functions. Application to Cascaded Constructions, in proceeding of Crypto 2004.
The above article strongly suggests that among using 128 bits from SHA-512 plus 128 bits from MD5 (or) using 256 bits from SHA-512, the later is much safer. The only assurance we have about the former is that it is not worse than SHA-512 truncated to 128 bits, which can be attacked in about 264 hashes, when we expect the later to require about 2128 hashes to attack.