Art,

There are certainly some poorly done meta-analyses. But I certainly cannot share a blanket disdain for meta-analysis. There are also some meta-analyses which are well done - especially the one which I have just finished!!!

I can't give particulars on the topic because it is work which has yet to be published. However, I had in my hot little hands survival data on over 1 million individuals from 18 different cohorts. We wanted to look at the relationship between survival and an individual-level characteristic. The individual level characteristic is based on easily obtained body measurements. There would be no question about consistency of these measurements across cohorts.

The cohorts were collected in a number of different places around the world. I would note, too, that cohort enrollment began at different times with a couple of cohorts being initiated in the 1980's. Most were initiated between 1995 and 2003.

We assumed a priori that the cohort effects on survival would be random. Effects of time and place would result in some differences in the relationship between the body measurement and survival. But there are problems estimating a random effect survival model - especially with the large number of subjects available for this study. Beyond the shear volume of data, there were some inconsistencies across cohorts in how some adjustment characteristics (e.g., education) were obtained. Differences across cohorts in collection of these adjustment characteristics necessitate different models for each cohort. Note, too, that those difference contribute in part to random variation across cohorts in the survival function as it relates to the primary predictor.

In order to estimate the random effect survival model, we fitted a fixed effect survival model for each cohort using the PHREG procedure and then treated the log hazard ratio estimates from each cohort as the response in a meta-analysis.

Given parameter estimates beta_hat{i} with associated standard errors se_hat{i}, the meta-analysis is performed by assuming that

beta_hat{i} ~ N(beta, se_hat{i}^2 + tau^2)

[See: Brockwell SE, Gordon IR. A comparison of statistical methods for meta-analysis. Statist Med 2001; 20:825-840. DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials 1986; 7:177-188.] This model is fit by the code

proc mixed data=estimated_parms; class cohort; model beta_hat = / s; random intercept / subject=cohort; repeated / subject=cohort group=cohort; parms / parmsdata=VHAT hold=2 to 19; run;

Special attention is necessary to construct the data set VHAT. This data set must have a variable ESTIMATE which references all of the variances which are specified by the model

beta_hat{i} ~ N(beta, se_hat{i}^2 + tau^2)

Of course, values of the variances must be specified in the proper order. Ordered values of the variable ESTIMATE are

_n_ ESTIMATE 1 tau_init 2 se_hat{1}^2 3 se_hat{2}^2 ... ... 19 se_hat{18}^2

where se_hat{1}^2 is the square standard error for the parameter estimate beta_hat{1} which is the response value for the first record.

Note that the PARMS statement holds the values se_hat{1}^2, se_hat{2}^2, ..., se_hat{18}^2 as fixed. Typically, the value tau_init would be specified to be 0, but if there was strong a priori knowledge about the variability of the parameter estimate across cohorts (or, more generally, across studies), then one might specify a non-zero initial estimate for tau. In general, though, specifying an initial value of 0 for tau_init would not present any estimation problems.

Of course, our primary interest is to estimate the "average" effect of our predictor on survival. In the model

beta_hat{i} ~ N(beta, se_hat{i}^2 + tau^2)

this is beta. The intercept estimate from PROC MIXED is our estimate of beta.

Dale

--------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra@NO_SPAMfhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 ---------------------------------------

--- On Sat, 11/7/09, Arthur Tabachneck <art297@NETSCAPE.NET> wrote:

> From: Arthur Tabachneck <art297@NETSCAPE.NET> > Subject: Re: SAS meta-analysis commands > To: SAS-L@LISTSERV.UGA.EDU > Date: Saturday, November 7, 2009, 7:36 AM > Leonard, > > I've never been a proponent of meta-analysis and have > always clinged to the > quote from a SUGI 27 paper that went something like: "I've > never meta > analysis I liked." > > Regardless, you might want to take a look at the macro on > Michael Friendly's > page: http://euclid.psych.yorku.ca/ftp/sas/macros/meta.sas > > HTH, > Art > -------- > On Sat, 7 Nov 2009 16:37:42 +0200, Leonard Rusinamhodzi > <l.rusinamhodzi@CGIAR.ORG> > wrote: > > >Dear all, > > > > > > > >I'm looking for colleagues who have performed a > meta-analysis using SAS and > >if they can share their macros for the analysis. > > > > > > > >Thank you > > > > > > > >Leonard >