Have you made normal probability plots and scatterplots of the b0's and b1's? Especially if the numbers seem reasonably normally distributed, possibly with some outliers, I'd throw out the outliers, compute mean, standard deviation and correlation from what I had left, and hope that this all was adequate for the purpose at hand.

If this were really important and I didn't have anything else to work with, I might compute some kind of "Mahalinobis distance" of each point from the (trimmed?) mean relative to the (trimmed?) covariance, and then I might try to plot this Mahalinobis distance" vs. the x range and R-squared numbers. However, I'm not certain what I would do with the result.

hope this helps. spencer graves

Remko Duursma wrote:
This is really helpful, however: i do not have the covariance
matrix for each study. In fact, i only have b0, b1, R-squared
and the range of x's used for the fit.


Rekmo
--------- Original Message ---------

DATE: Mon, 01 Sep 2003 13:50:04
From: Spencer Graves <[EMAIL PROTECTED]>
To: [EMAIL PROTECTED]
Cc: rhelp <[EMAIL PROTECTED]>

Can you get the covariance matrices of the vectors b = c(b0, b1)? There is a reasonable literature on meta-analysis with which I'm not very familiar. However, a standard thing to do is to compute a weighted average with weights proportional to the inverse of the covariance matrices, while testing to evaluate whether the b's plausibly all estimate the same thing.

The theory is as follows: Suppose b.i ~ N.k(mu, Sig.i), i = 1, 2, ..., n. If you have a covariance matrix for each vector b.i, then you have this set-up. Assuming you do have (or can approximate) Sig.i, then

l.i = log(likelihood(b.i)) = (-0.5)*(k*log(2*pi)+log(det(Sig.i))+t(b.i-mu)%*%solve(Sig.i, (b.i-mu))).

The first derivative of l.i with respect to mu is as follows:

D.l.i = solve(Sig.i, (x.i-mu)).

The solution for mu of sum(D.l.i)=0 is as follows:

mu.hat = solve(sum(Sig.i), sum(solve(Sig.i, (x.i-mu)))).

One could also derive various statistics for evaluating whether it is plausible to believe that these b.i's all come from the same population. I would assume that the literature on meta-analysis would deal with this, but I have not looked much at that literature, and I'll leave that question to others.

hope this helps.
spencer graves

Remko Duursma wrote:

Dear R-helpers,

i have the following situation: i have a bunch of

y=b0 + b1*x from different studies, and want to estimate a "general" y=f(x). I only have the b0,b1's and R-squareds. Should i weigh the separate equations by their R-squared?

thanks

Remko


^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~' Remko Duursma, Ph.D. student Forest Biometrics Lab / Idaho Stable Isotope Lab University of Idaho, Moscow, ID, U.S.A.

______________________________________________
[EMAIL PROTECTED] mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help

______________________________________________ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help



______________________________________________
[EMAIL PROTECTED] mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help

______________________________________________ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help

Reply via email to