The failure rate of a product generally follows a thing call the 'bathtub curve', so called because a plot of failures vs time looks somewhat like a longitudinal cross section through a bath.
The steep side is located at t = 0, and represents 'infant mortality', ie. those units which fail prematurely because of some manufacturing defect such as a cracked solder joint or ESD damage. It's because of this part of the curve that we do burn-in and temperature cycling at the factory, if we're looking to sell a product that's required to have very high reliability straight from day one in the customer's hands. Better that a product with a latent defect fails inside your own QC lab than finding it's dead on arrival at the customer's site.
The failure rate then falls close to zero and stays that way for a long period of time. This is the product's normal service life. The factory defects have been weeded out, and the remaining units are working exactly as they should.
Finally, the failure rate starts to slowly rise again, as units start to reach the end of their service lives and simply wear out. Exactly when a given unit fails depends on how it's been used - how hot it's been running, how many temperature cycles and how extreme they were, how many hours/day it was switched on, any stresses to which it's been subjected by the user, its power supply, ancillary equipment, and anything else you can think of that's relevant to your product.
Your warranty should, of course, cover the early life failures, and ensure that customers are into the flat portion of the bathtub curve before the warranty expires. However, once they're into that region, it doesn't really matter how long a warranty you offer, because the expected failure rate is very low anyway. Units that fail may well do so because they're being used outside their design parameters (intentionally or otherwise), so you may or may not decide you want to cover these, and can adjust the length of the warranty you offer accordingly.
Calculating the MTBF of any particular product as a whole is complex, and depends on various assumptions about what happens in the event of a failure. If a component fails, is just that component replaced, or a whole assembly, or the whole product? It makes a huge difference.
Normally when I'm asked for an MTBF figure, I start by picking out any components which have an especially short MTBF compared to the other parts in the design - fans and other moving parts, followed by electrolytic capacitors, then maybe power components that are subject to temperature cycling, or very fine geometry digital semiconductors (these increasingly come with a limited lifetime spec which is strongly temperature dependent). These parts will dominate your MTBF to the extent that you can probably ignore everything else.