An intention to treat analysis should ideally have responses from everyone randomised, whether they complied with the protocol or not.
When responses are missing, the best we can hope for is to estimate the distribution of the missing respones. For each individual, this distribution will depend on whether they were lost to follow up because they stopped complying (taking the intervention as specified in the protocol) or other reasons. If they were lost because they stopped complying, then the distribution of their unseen responses should not be estimated from those who carried on complying (that would be closer to addressing the per-protocol hypothesis).
So, if by 'as observed' data you mean an analysis that includes observed responses from patients whether or not they were still complying with the intervention, then this is addressing the ITT question.
LOCF is basically saying that before the last observation was seen the patients condition stabilised, and has not changed since, so that their last observation is a valid draw from their final stable response distribution, which could have been drawn at any subsequent time (in other words, if we happened to see subesquent measurements, we could randomly change around the time they occured and the inference would be unchanged). This is implausible if dropout is a reasonable time before the end of the trial, as the patient almost certainly would have gone on to another intervention, so that their unseen response would be quite different.
These points are discussed in much more detail in the forthcoming monograph 'Missing data in clinical trials - a practical guide' by Carpenter and Kenward which is available here.