Lewis famously identified 'possible worlds' with spatiotemporally isolated regions (call them 'Lewis-worlds'). But, as Cian Dorr points out in 'How to be a Modal Realist', the view is much improved by jettisoning this claim. There doesn't seem any principled reason for taking spatiotemporal boundaries to be of any fundamental modal significance (and it makes it harder for Lewis to accommodate extra-worldly modal facts, like "possibly, there are at least two Lewis-worlds"). Instead, the best version of the view goes something like this:
Accept Lewis' ontology. So there's a plenitude of concrete objects, scattered across countless Lewis-worlds. This all actually exists (we're not going to arbitrarily limit the scope of the actual to the stuff around here -- restricted quantification suffices to make true ordinary assertions of "there are no talking donkeys"). For qualitative claims using unrestricted quantifiers, modal operators are redundant: possibly there is a talking donkey, just because there really is a talking donkey. (There's no need to add "in some Lewis world", any more than we would add "in some grassy field".) Then, to accommodate de re modal claims, we simply overlay the qualitative facts with an identity interpretation (or what Dorr calls a 'counterpairing': a function from objects to counterparts). So, for example, 'I could have been standing right now' is true because there's a guy X standing up at a time t, and a counterpairing relation R which pairs me with X and now with t.
(Note that it's then the pluriverse-spanning counterpairing relations, rather than the individual Lewis worlds, that play the role of "possible worlds". To be "true at" a counterpairing relation is to be true according to the implied assignment of identities to the qualitative pluriverse.)
It's a neat picture. I should clarify that Dorr himself doesn't think that it's true, but merely an improvement over Lewis' formulation. I'm inclined to agree with him on both counts!