By way of background: I'm inclined to think that, while we can of course have impersonal reasons to take non-existent people into consideration (e.g., by preventing miserable lives from coming into existence), we cannot have personal reasons that ultimately stem from non-existent entities. Roberts argues against this view by means of a temporal analogy (p.769):
[T]he judgment that continuing to exist would be worse than dying for me, on the face of things, would seem perfectly cogent. I am not sure why the judgment that coming into existence makes things worse for the miserable Meg is not just as cogent. The subject doesn’t exist at the world where the choice is made not to bring Meg into existence, just as the subject doesn’t continue to exist at the world where the choice is made to die. But in both cases there is still a subject; the future could have unfolded in a way that includes Meg, just as the future could have unfolded in a way that (still) includes me.
But this strikes me as a bad (or at least limited) analogy. Given Eternalism, I exist (here and now) regardless of whether I die tomorrow or not. My current existence suffices to provide a subject that can continue to be referred to even after I die. But in non-identity cases, whether there is a subject at all depends on what choice we make. If we choose to bring Meg into existence, then there is a subject and we can regret her suffering for her sake. But if we don't, then there isn't, and our reason for relief at avoiding such suffering can only be impersonal.
(Roberts adds, on p.770, that "even if some form or another of modal actualism is true, any plausible articulation of moral law is going to require some reference to – not a de re reference; not a singular reference; but some way of talking, at least in some descriptive way, about – people who will in fact never exist at all." Which is true enough, but compatible with my view that this descriptive way of talking can only ground impersonal reasons.)
So, let's accept that we have only impersonal reasons to prevent the suffering of the non-existent (by ensuring that they don't exist), but both personal and impersonal reasons to relieve the suffering of actual people. Does that mean that the latter "count for more"? One difficulty with this idea is that, were we to discount a supposedly non-existent person's interest in non-existence to the point where we allow them to come into existence (say to prevent a lesser harm to others), then suddenly they turn out to be an actual person whose interests count with full force, and hence who we have wronged by discounting. So we can't discount the value of preventing miserable existence in this way, relative to relieving the misery of actual people. The personal value of preventing suffering must substitute for, rather than add to, its impersonal value.
On the other hand, it seems we can coherently discount the positive value of bringing more happy people into existence, compared with the value of making actual people happier. (Were we to bring them into existence, we would turn out to have personal reasons to be glad we did so. But no personal wrong is done by failing to bring them into existence. This seems like a curious case where the moral quality of one's options depends upon which option one actually takes.) So it's open to us to hold that personal value adds to the impersonal value of increasing happiness for actual people, in contrast to the case of preventing misery. This would seem an odd asymmetry though. What could explain such an asymmetry between promoting the good and preventing the bad? Perhaps just the observed point: If we are to count personal value over and above impersonal value, this is the only coherent way to do it. (Still, I want to stress that it strikes me as very much an open question whether we should want to count personal reasons in addition to impersonal reasons, rather than as merely substituting the one kind of normative force for an equally weighty other kind.)
For a point of comparison: Roberts proposes a view she calls Variabilism to systematize the asymmetry (p.773):
The loss of wellbeing incurred at a world where the person who incurs that loss does or will exist has full moral signiﬁcance for purposes of evaluating an act performed at a given world that imposes that loss and any alternate act performed at any alternate world that avoids that loss, while a loss incurred by that very same person at a world where that person never exists at all has no moral signiﬁcance whatsoever.
This accommodates some central intuitions, but seems to lack any theoretical rationale. So I wonder whether the central intuitions might be better captured by the view I describe above. If we take seriously the idea that personal value can only arise for actual people, together with the idea that discounting the misery of non-actual people is incoherent, then this also gets us the desired results that (1) We have strong reasons to prevent miserable lives from coming into existence, and (2) We have weaker reasons to bring happy lives into existence (at least if we actually refrain from doing so).
There are a couple of subtle differences between the two proposed views that may be brought out as follows. (i) Suppose Adam actually exists, and consider a world w where he never exists. Obviously w is in no way accessible from our world, but we might still adopt various evaluative attitudes towards it. On one natural understanding of my account, we should regard w as lamentable for its lack of Adam (even if some other, similarly happy person 'Bob' exists in his place). Not so on Roberts' account, as I understand it: Adam doesn't exist in w, so when evaluating w his loss counts for nothing.
(ii) Imagine a world w+ where the (actually non-existent) 'Bob' is even happier than in w, and another world @+ where Adam is similarly happier than he is in the actual world @. On my proposed view, we have merely impersonal reasons to prefer w+ to w, and whereas we have both impersonal and (non-substituting) personal reasons to prefer @+ over @. So the latter preference should be stronger than the former. On Roberts' view, since the beneficiaries each exist in their relevant world (w+ and @+, respectively), they both have "full moral significance" and hence there would seem no grounds for any difference in preference.
While these are subtle differences, I think they count in favour of the view I've described. If we are to countenance any fundamental asymmetry at all (and I'm not yet convinced that we should: the merely 'derivative' asymmetries offered by value holism seem quite adequate to me), then it seems to make most sense -- on both theoretical and case-based grounds -- to have it stem from a kind of partiality towards actually existing people (within the above-noted constraints).