On Thu, 18 Mar 2004 09:43:06 -0500, "RichardM" <[EMAIL PROTECTED]> wrote:
> I've tested binomial and chi-squared approximation (or z) and I found that > even with very large samples, the probability obtained are different. The > difference is often small (especially when N is large and when P is close to > .5), but knowing that many people will consider probabilities of .051 and > .049 as quantitatively close but qualitatively different (significant vs non > significant), I am reluctant to use any normal approximation when the > binomial probability can be computed. [ snip, where I said to State the Actual Proportions.] Have I missed the question? If you have a personal objection to using 'approximations' (even if they are conservative and convenient), then don't use them -- if you have the information that you do need, for the power analysis that you were mentioning. The difference in two proportions is best expressed as the statement of the two proportions; so, state it that way. Why are you looking for another way to show the effect size? - Now, if you need to look up the power according to 'effect,' then you should keep in mind that -- for the purpose of Power Analysis, with its future unknowns -- you should consider 0.049 and 0.051 to be equal. Also, your figuring might as well let either of them slop over to be 0.04 or 0.06. Your discrete tests, for small N, are lumpy, so you should not pretend to have more precision than that. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html - I need a new job, after March 31. Openings? - . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
