On Thu, 02 Nov 2000 17:44:08 GMT, Ross J. Micheals <email@example.com> wrote:
> Is there a statistical definition of "representative sampling?"
No. It is more broadly *logical* than statistical. My notion right this minute is, it is most clearly defined by its negation. That is, if you avoid the traps of all potential non-representative-ness, then you may have representative samples.
Someone just has to find a single way that your system really messes up, and that could show that your samples are not representative, no matter how many other tests your selection system may have passed.
> That is, > is there a way to show that a sampling strategy will yield (within some > tolerance) a sample that is "representative" of the population?
Sure, you should be able to draw prospective confidence limits for different schemes. Is that all you are asking?
Let me point to an obscure problem. When you are drawing a set of random samples, should you *insist* that extreme relationships of one sort and another be present, at least occasionally? And, is there a difference in the requirement if you are drawing ONE sample, as opposed to drawing a large sequence of samples?
- I have a computer program for doing Random Assignment to trials, which produces your 100 (or 500 or however many) IDs and Groups, to be used in consecutive order. A runs-test is included in the output.
I expect that if you produce a random assignment with your chosen 'seed number', you might be unhappy if one group got the assignment 8 times in a row; and the like. So, if the program is extreme according to the runs-test, you can try another seed and get another list. Then you use that one if you like it better.
- BUT you should see: A single selection will look 'representative.' A million selections done like this will be lacking, grossly, in Runs. Does that matter? - well, it might.