On 30 Mar 2002 09:46:45 -0800, [EMAIL PROTECTED] (wuzzy) wrote: > Maybe someone will point me to other newsgroups or mail groups on > biological or clinical statistics as I know that sci.stat.edu is about > the education of statistics not really about stats itself.. > > My question (frustration, rather) is: how do you deal with the fact > that signs on coefficients of multivariable models change direction > and size when you remove a predictor of the dependant variable(s).
How do you deal ... ? You decide on the order -- what comes first rationally. You argue for your choice. You try to show that other choices are irrational for external reasons, or are contrary to other data. Some of this has been discussed as "path analysis" and "structural equations" but the basic information is often available as the zero-order correlations and the various partial correlations (or regression coefficients). > > is there a test for this? There is a test for whether one coefficient changes notably when another variable is added/ dropped. You can find a discussion with detail, by using groups.google.com -- specify the subject line or the message-id. This post has a lot of the detail -- ========== header From: Gary McClelland <[EMAIL PROTECTED]> Newsgroups: sci.stat.consult Subject: Re: sig. of diff. between r and partial-r Date: Thu, 04 Jan 2001 16:04:43 -0700 Message-ID: <[EMAIL PROTECTED]> ========== end of header. > > It seems to me that if genetics place an important part in determining > cholesterol levels (say), and you study diet as it relates to > cholesterol, but forget to insert genetics (say in a twin study), then > you might find that eating eggs raises your cholesterol whereas if you > don't include eggs might lower cholesterol. (hyopthetial) > - that doesn't parse, but we get the idea. > > How do observational and especially exploratory studies overcome this > problem? The good ones will admit the problem and explain it. Secondary variables are helpful for trying to nail down the primary ones. - This is one reason why observational studies and surveys that are turned to questions not originally planned become particularly untrustworthy. For any survey, you worry about potential biases that can't be thoroughly nailed down, even though you thought of them beforehand. For new questions, there might be no relevant attempt, not even a simple one, to control for some obvious bias. > > I think you can only overcome it (given lack of theoretical grounds to > overcome it) by looking at the size of the relationship and "guess" at > how true your result is, even though the opposite of your result might > be what happens in real life. Size of relationship helps. A BIG effect (like a 5-fold mortality risk) is more convincing than a 50% increase. A tiny relationship can't account for a huge outcome. But 'guess' is not the right word, unless you mean the educated guess of the professional. In epidemiological research, the biological underpinning has to be believable. -- 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/ . =================================================================
