The cited text on JSTOR requires either an institutional affiliation or an upgraded (paid) membership, so I can only see the preview page, but I'll take your word for the average ages you quote. Whether using 20 or 25 as a likely age for brides, and 23 or 28 for grooms, I still regard it as preferable to have something in the data field for DOB. Frankly, I wouldn't create a profile without it.
The cumulative error range you suggest would apply when coming forward in time if we did not have good data for the following generations -- but here, we do. So practically speaking, the issue is what may happen if we try to go further back. But frankly, since I can't imagine anyone proposing the creation of profiles further back without _some_ believable data to anchor them, I don't see a real-world problem.
I.e., let's suppose we find some credible marriage record for believable parents for Walter or Grace, and anchor a parental profile (or two) in that information. Using the same formula I propose using for Walter and Grace, we would be unlikely to be farther off the mark with said parents than we were with their offspring. [BTW I don't think I've ever created a parent's profile absent their birth and/or marriage record, and instead based solely on the birth date of a presumably-eldest child; if I did, then I'd certainly have to agree that I'd be introducing a very wide window of uncertainty.]
If we found only a death date for a parent of either Walter or Grace, and used only that as an anchor for a profile, then, again, your general point about a widening error range would certainly be more relevant... however, I find it hard to imagine creating a profile based solely on a death date. I suppose one could do it using a contemporary average-age-at-death and calculating an estimated DOB from that. I just don't think I've seen that done, and I'd raise an eyebrow (or more) if I saw it happening with a PGM profile.