From chemistry-request@ccl.net Thu Apr 14 12:26:40 2005 Received: from smtp.goldrush.com (smtp.goldrush.com [206.171.171.11]) by server.ccl.net (8.13.1/8.13.1) with ESMTP id j3EGQZZF016535 for ; Thu, 14 Apr 2005 12:26:35 -0400 Received: from Compaq (x2-04-151.goldrush.com [64.162.10.151]) by smtp.goldrush.com (8.12.8/8.12.8) with SMTP id j3EGQJC1023826; Thu, 14 Apr 2005 09:26:27 -0700 From: "Steve Bowlus" To: , Subject: RE: Correlation Coefficient Date: Thu, 14 Apr 2005 09:26:20 -0700 Message-ID: MIME-Version: 1.0 Content-Type: text/plain; charset="Windows-1252" Content-Transfer-Encoding: 7bit X-Priority: 3 (Normal) X-MSMail-Priority: Normal X-Mailer: Microsoft Outlook IMO, Build 9.0.2416 (9.0.2910.0) Importance: Normal In-Reply-To: X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2900.2180 X-MailScanner: Found to be clean X-MailScanner-SpamCheck: X-MailScanner-From: chezbowlus)at(goldrush.com X-Spam-Status: No, score=0.1 required=5.0 tests=DNS_FROM_AHBL_RHSBL autolearn=failed version=3.0.1 X-Spam-Checker-Version: SpamAssassin 3.0.1 (2004-10-22) on server.ccl.net Cross-validated r-sq (or q-sq) provides a measure of the homogeneity of the data set. Depending on how the model is developed, it reveals whether there are influential points in the data set which are unduly weighted in the model. In this regard, q-sq replaces the influence statistics common in univariate statistics. By "reliable" are you actually meaning to ask whether predictions made by the model will be accurate (i.e. useful)? One must remember that 1) predictions are associated with a confidence interval, and 2) the new compound must lie in the prediction space of the model (not be an outlier wrt any of the descriptors). So prediction made with a large CI may be statistically correct, but practically useless ("unreliable"?). Bottom line, the model should have high r-sq, high cross-validated r-sq, and a reasonably tight standard error of prediction. What is useful is of course dependent on the situation. Small, noisy data sets at the beginning of a project, where the intent of the model may be to determine areas or descriptors for exploration, might use r-sq on the order of 0.7 and q-sq as low as 0.3. As the dataset grows and becomes more homogeneous in the descriptor space, r-sq and q-sq should both increase and converge on the same value, while the SE shrinks to a (practical) limit of the accuracy of the (bio)assay. One or more journal editors may weigh in with their favorite cutoffs, but I am not aware there is any magic number to assure the "reliability" of a model. Steve Bowlus -----Original Message----- From: Computational Chemistry List [mailto:chemistry-request)at(ccl.net]On Behalf Of lhhu)at(yangtze.hku.hk Sent: Wednesday, April 13, 2005 9:49 PM To: chemistry)at(ccl.net Subject: CCL:Correlation Coefficient Dear All, I have a question about Correlation Coefficient of cross-validation. Can anyone tell me how much the value of Correlation Coefficient indicate the result is reliable? Certainly I know the larger the value, the more reliable the result. I want to know the lowest limit of that value. I read some papers that used the Correlation Coefficient value, some values are lower than 0.8, is it OK for present them in journals? Or it depends on different situation or topic? Thanks a lot to you all, Holly -= This is automatically added to each message by the mailing script =- To send e-mail to subscribers of CCL put the string CCL: on your Subject: line and send your message to: CHEMISTRY)at(ccl.net Send your subscription/unsubscription requests to: CHEMISTRY-REQUEST)at(ccl.net HOME Page: http://www.ccl.net | Jobs Page: http://www.ccl.net/jobs If your is mail bouncing from ccl.net domain due to spam filters, please use the Web based form from CCL Home Page -+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+