When considering the "model" aspect of a game, remember that models are always abstractions. Abstractions have levels, such as the Bjourne Stroustrup discussion in his original C++ book of you don't talk about building a house of wood atoms, bricks, and rooms all at the same time. And when you are considering things at the "brick" level, it's rarely meaningful to talk about "rooms" and vice versa.
A game should be built on the proper level of abstraction. The exact level can vary from game to game, and sometimes from sub system to sub system. However, within a given game, and certainly a sub system, there is some level of abstraction that fits, and as much as possible the game should use that in its modeling. There are, of course, other concerns besides the model, and those can sometimes lead to elements in the system that stray from an exact fit on the abstraction. Even so, something that starts to diverge too much will still be a problem.
Given that then, I would say that ability score can be somewhat analogous to real world metrics, in a sort of rough and ready way, subject to the abstraction level of the model. I would not call this comparable. For example, in D&D, given what "strength" actually does, it's not merely lifting. It's muscle tone, some general fitness (though not all of it), muscle/eye coordination (though not all of it), etc. And in fact, I find game models typically work better when the label is not read too literally, with some overlap between the elements, as with D&D strength, dexterity, and constitution.
If you want, for example, that model to be more closely analogous, then the underlying system needs to acknowledge that hitting something with a sword is really influenced by Str and Dex, while being able to fight for more than a few seconds is influenced by Str and Con. This is necessarily a complication of the model and the system. Likewise, the model could pull back a little, say, collapsing Str and Con into one score or even collapsing all three into some kind of "physical" or "body" or whatever score. The proper level of abstraction depends on what is happening in the rest of the game that uses those ability scores.
Finally, you also have to consider the rest of the the model and what it is doing, how much overlap you want, etc. For example, in the D&D example, one of the things hit points model is rough and ready stamina in a fight. The D&D model is concerned with two states and a counter. I'm either still in the fight or I'm not. The hit point counter is how close I am to not being in it any more, subject to the uncertainty of getting hit or not. So it is probably a not very useful complication to the system to make some kind of directly comparable stamina mechanic based on Con or even all the physical attributes. That's already covered by a different mechanic--at a different level of abstraction. Or if you do want to have such a mechanic (similar to Hero System's endurance), then that means how hit points work also needs to change to remove that part of it.