When I first read your question, I frankly wasn't sure what "self-plagiarism" might be. I think the term is a poor descriptor but more on that shortly. This is a great question because it touches on so many issues - the peer review process, the "publish or perish" philosophy, tenure, professional integrity, and morality.
The problem with plagiarism, the way I see it, is that the originator of the work doesn't get credited for his or her efforts and someone else gets credit for something they didn't do. 'Credit' is more than simple acknowledgement, it's the currency of tenure. It's basic thievery. You can't steal from yourself, so the term seems to be a misnomer.
But whatever we call it, it's surely not good. I'm not familiar with the articles mentioned in Cohen's blog but the widely cited 1998 articles by Paternoster, Brame, Piquero, and Mazerolle [both] on the appropriate significance test for the equality of coefficients across models immediately came to my mind. One was published in Criminology and the other in the Journal of Quantitative Criminology - both top journals in the criminology field. The authors, like Hipp, are well-known and respected. I'm sure if we looked, and no one does, we could find many more examples from other authors.
I don't want to speculate about why people do this. The obvious answer is the requirements of tenure at research oriented universities, but sometimes the obvious answer is not the right one. I will say that the perpetrator - Hipp in this example, is surely harmed as a result and rightfully so. I suspect these journals, which are great journals, will not be so eager to publish his work in the future. He may have also alienated other journals as well as other academics generally, which is certainly not a boon to his career. The bottom line is that this is cheating; it's tantamount to scribbling notes on your hand before an exam. Others aren't directly harmed by this; their articles might not be published and the journal's resources are wasted reviewing and publishing something that's already been published elsewhere. Any single instance of this behavior is of minimal consequence to others - the same is true of the exam example - but in both scenarios imagine what happens when lots of people cheat. Besides, it's morally, intellectually, and professionally dishonest.
I can say that in my specific area - judges and sentencing - the convention seems to be that given a good dataset the authors publish many related articles from it that belong in a single publication. For example, in considering how different judges sentence, authors typically publish one article on black vs. white judges and another using the same data and method on male judges vs. females. They are "thinly sliced 'salami' articles" to quote Cohen. The exception here is the recent work by Brian Johnson (2006) but even his article which is comprehensive compared to many previous publications on the same topic shares some similarities (and the same data) with Ulmer's 2004 piece.
I find myself sitting on a fantastic dataset and grappling with these same issues, counting the number of publications I think I can squeeze out of the data. It’s a gray area. I think varying a couple independent variables does not warrant a separate article unless the theoretical focus of the paper is different. I guess one way to think about it is this – if the lit review and theory sections would be very similar for both articles then the methods section must be substantially different (i.e. better statistical technique, different outcome variable, etc.). Adding a couple more predictors or swapping one predictor for another is not a justification for a new paper unless the theoretical focus is changed and this should necessitate a rewrite of the frontend, discussion, and conclusion.