Many thanks again for your suggestions. Sorry. Effectvely there was a mistake in the HR I previously submitted. As suggested, I started from scratch and the correct HR are: HR 3,12 (2,24-4.35) HR 1,25 (1,03-1,52) Mario On 23/11/2014 09:23, Viechtbauer Wolfgang (STAT) wrote: I actually don't see Michael suggesting that you should work with the individual studies. My interpretation of his reply is that he is suggesting the same thing that I have done. But in the end, you should get similar results whether you test those two summary (log) HRs against each other or if you work with the individual studies and test whether the summary (log) HRs of the first set is different from the second. I have cc-ed Michael in case I am misinterpreting his suggestion. I had originally thought of using the individual studies as Mario interpreted but I do not have strong feelings either way. I suppose I was thinking of cases I have seen where either injudicious rounding or typos has meant that it was better to start from scratch. [Show Quoted Text - 122 lines][Nascondi Testo quotato] Best, Wolfgang -----Original Message----- From: Mario Petretta [mailto: <https://inbox.unina.it/horde/imp/message.php?mailbox=INBOX&index=101432> petretta at unina.it] Sent: Friday, November 21, 2014 18:19 To: <https://inbox.unina.it/horde/imp/message.php?mailbox=INBOX&index=101432> r-help at r-project.org Cc: Viechtbauer Wolfgang (STAT) Subject: R: [R] Comparing summary hazard ratios in meta-analysis Many many thanks to Michael Dewey and to Viechtbauer Wolfgang for the kindly and useful replay !! I only ask to Wolfgang if I should log-transform hazard ratios and compute standard error only for the summary hazard ratio estimates or, as suggested by Michael, for each single study, combining them in a single data frame and thereafter performing the meta-regression incorporating the moderator. Mario -----Messaggio originale----- Da: Viechtbauer Wolfgang (STAT) [mailto: <https://inbox.unina.it/horde/imp/message.php?mailbox=INBOX&index=101432> wolfgang.viechtbauer at maastrichtuniversity.nl] Inviato: venerd? 21 novembre 2014 15:37 A: Michael Dewey; Mario Petretta; <https://inbox.unina.it/horde/imp/message.php?mailbox=INBOX&index=101432> r-help at r-project.org Oggetto: RE: [R] Comparing summary hazard ratios in meta-analysis Those hazard ratios and CIs seem a bit strange. On the log-scale, they should be symmetric, but they are not. Could be due to heavy rounding though. At any rate, it comes down to this: hr <- c(3.12, 1.15) ci.lb <- c(2.2, 1.03) ci.ub <- c(4.1, 2.6) meta <- c(1,2) ### log-transform hazard ratios and compute standard error based on the CI bounds yi <- log(hr) sei <- (log(ci.ub) - log(ci.lb)) / (2*1.96) library(metafor) res <- rma(yi ~ factor(meta), sei=sei, method="FE") res So, yes, the two hazard ratios are significantly different from each other. Best, Wolfgang -- Wolfgang Viechtbauer, Ph.D., Statistician Department of Psychiatry and Psychology School for Mental Health and Neuroscience Faculty of Health, Medicine, and Life Sciences Maastricht University, P.O. Box 616 (VIJV1) 6200 MD Maastricht, The Netherlands +31 (43) 388-4170 | <http://www.wvbauer.com/> http://www.wvbauer.com -----Original Message----- From: <https://inbox.unina.it/horde/imp/message.php?mailbox=INBOX&index=101432> r-help-bounces at r-project.org [mailto: <https://inbox.unina.it/horde/imp/message.php?mailbox=INBOX&index=101432> r-help-bounces at r-project.org] On Behalf Of Michael Dewey Sent: Friday, November 21, 2014 13:25 To: Mario Petretta; <https://inbox.unina.it/horde/imp/message.php?mailbox=INBOX&index=101432> r-help at r-project.org Subject: Re: [R] Comparing summary hazard ratios in meta-analysis On 21/11/2014 08:51, Mario Petretta wrote: Dear all, I use R 3.1.1 for Windows. I performed two different meta-analysis assessing the prognostic value of two different tests in patients with coronary artery disease. The study included in the two analysis are different. That makes life simpler. The variable of interest in dichotomous (normal/abnormal result) for both tests. The effects size is hazard ratio and its standard error (ln units) for both meta-analysis. It sounds as though you might want to use meta-regression. You will need a single data frame containing at least log hr, se of log hr, an identifier for the test. I would use the metafor package for this, look in the documentation for how to incorporate a moderator (your test variable). The advantage of meta-regression is that you not only get a test but also a measure of how different the hr are with a confidence interval. I would like to statistically compare the two summary hazard ratios and 95% CI (eform) obtained from the two meta-analysis. For one meta-analysis: HR 3.12 (95% CI 2.2 - 4.1) For the other: HR 1.25 (95% CI 1.03 - 2.6) It is possible or I'm comparing apples with oranges? Any suggestion is welcome. ------------------------------------------------------- Mario Petretta Associate Professor of Internal Medicine Department of Translational Medical Sciences Naples University Federico II Italy -- Michael <http://www.dewey.myzen.co.uk/> http://www.dewey.myzen.co.uk [[alternative HTML version deleted]]