Multiply your original estimate by 3, works most of the time
replies(6):
So: 2 hours -> 4 days, 1 week -> 2 months, etc.
(I'm not sure where this turned up, but it's a long time ago, going on three decades.)
The other option is to carefully track tasks, relevant dimensions, estimates, and actual performance, and see if there's any prediction modelling which can be derived from that. Problem is that this a classic instance of modelling in which the model itself affects the domain (as with economics and sociology, contrasted with astronomy and meterology where this isn't the case), such that estimates incorporating past performance data will incorporate the fact that the prediction model is being used.
Note that this also comes out about the same if you're using months (12 months -> 24 years (2.4 decades)) and (at least within a factor of two-ish) weeks (52 weeks -> 104 months (0.867 decades).
I'm not claiming this is accurate, I'm stating that it's a heuristic I'm familiar with.
This may have been in Arthur Bloch's Murphy's Law and Other Reasons Things Go Wrong, though I'm not finding it in comparable collections. Possibly from project planning literature I was reading in the 1990s (DeMarco & Lister, Brooks, Boehm, McConnell, etc.).