On Fri, 3 Jan 2003 00:18:42 +0800, "K?" <[EMAIL PROTECTED]> wrote:
[snip, some] > > For Question 2, it was an non-equivalent group design for a community > program. The experimental and comparison groups had rather different samples > sizes and baselines, and I am trying to consider if t-test is appropriate > for testing the difference between the mean differences for the two groups. - "baselines" implies that these differences are pre-post. - "non-equivalent group" suggests that there could be initial differences that make any firm conclusions impossible. The three main ways to look at outcomes are these: a) ignore baseline, and compare followup; b) compare simple change scores (by t-test, say); and c) compare regressed change scores. Any or all of these can be hazardous, or irresponsible, if the baselines are sufficiently different. But if you do all three analyses and they come out the same, then you are probably in good shape with your conclusions, and you can report the results with pretty good confidence. Assuming that there is some pre-post correlation: The one-way analysis of covariance (ANCOVA) is typically the most powerful, and the most popular in controlled designs where the baselines won't differ. In other designs, you *must* look at the initial means, and if they do differ notably, then you must consider the pattern of means. Potential problems include scaling problems, which show up as differences in variances, perhaps suggesting 'ceiling' effects, etc. Oh, with non-equivalent groups, if there is any effect that you want to claim, you are also taxed with explaining away other potential confounding variables. For that, you need to describe the groups on other dimensions that *might* matter, and try to show why they didn't matter. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
