On Fri, 30 Nov 2012 03:08:51 +0000, @ com (JJ) wrote:
>As my understanding, when I assign a weight n (say) to an observation in >the weighted regression (logistic or linear), it is equivalent to repeat >the same observation n times in the data set. >I have seen people assign 1 million as the weight to an observation. It >seems unreasonable to me but I don't have the knowledge to say why it is >not reasonable.
It's an ad-hoc way to force the regression line through that point, for the usual, smaller N, when you work from the variance of the point. But it surely screws up the degrees of freedom if you do it by means of "repeating the observation" by weighting outside of the regression program (as you can do, say, with SPSS).
>If you have any references (titles of books, articles, etc) discussing >the role of weight in modeling, please let me know.
Survey sampling uses weights, at times, but that does not sound like the sort of problem that you were citing.
In case no one suggests something particular, I will suggest looking at the regression books by Harrell or J. Cohen, and see if they mention the topic, and work from there.