If $H$ is secure in the random oracle model and of width at least $n$, then $x\mapsto P(H(x),n)$ is also secure, inasmuch as an $n$-bit hash can be (proof sketch: any attack that distinguishes $x\mapsto P(H(x),n)$ from a random oracle is trivially turned into a distinguisher for $H$). The converse is not true.
Finding a collision requires $O(2^{n/2})$ queries to the oracle (or evaluations of the inner function of $H$). More precisely it is expected to require $o(\sqrt{\pi/2}\cdot2^{n/2})$ distinct queries, and odds of success reach 50% after $o(\sqrt{\ln 4}\cdot2^{n/2})$ such queries (that's the generalized birthday problem). The standard paper on how to exhibit such collisions is Parallel Collision Search with Cryptanalytic Applications.
Finding a preimage is expected to require about $2^n$ queries to the oracle (I plan to dig out the right formulas someday). It is possible to speed-up an attack by pre-computation, e.g. with rainbow tables, in particular if the set of possible messages is known, or if $n$ is small enough.
Now we restrict to MD5, thus $n\le 128$.
The difficulty of finding collisions is about as for a good hash ($\approx 1.3\cdot2^{n/2}$ MD5 inner rounds) when the messages are tried without consideration of the inner structure of the MD5 round function (e.g. random messages or words from a dictionary). Without constraints on the message, finding collisions among messages of at least 1024-bit has been possible for any $n$ since that breakthrough, works even with some constraints, now has little cost, and was recently made possible with 512-bit messages. The lower $n$ is, the easier it gets. AFAIK there is no shortcut for heavily constrained messages, e.g. passwords of less than 10 ASCII characters.
The difficulty of finding a preimage is about as for a good hash (about $2^n$ MD5 inner rounds) when the messages are tried without consideration of the inner structure of the MD5 round function, nor knowledge on the origin of the target value. I am not aware of a practical method better than that, except when the message is in a known set smaller than $2^n$, which is when rainbow tables shine; in that case, I do not see that lowering $n$ changes the cost.
