09-23-2014 04:54 AM
I am a basic to intermediate level SAS user. I am currently doing meta analysis on Low Back Pain with original data we obtained from 27 clinical trials across the world. I created a master data set to combine the commonly used variables (like age, sex, etc.) and important outcome variable. I was just asked to run some overall meta-analysis on the master dataset, to see the overall effectiveness of exercise vs other treatments. I am thinking of random effect model. It seems PROC MIXED is widely used. But I suppose PROC MIXED for individual study must be somewhat different from PROC MIXED for pooled analysis? For example, which study the data come from should be one of the random effects. I am wondering if someone here could give me suggestions or a sample code to run this type of analysis, given that each trial is run under different conditions and their treatments are not exactly the same? I guess what I am saying is how to deal with the heterogeneity across the studies.
I need to get this done ASAP. Any suggestions will be greatly appreciated! Thanks so much!
09-23-2014 08:50 AM
It sounds like you will first need to read more about meta-analysis. The excellent book by Anne Whitehead (Meta-Analysis of COntrolled Clinical Trials) is a good place to start. It is not clear if you are wanting to do a series of univariate meta-analyses (control - treatment), or a multi-treatment meta-analysis (control, treatmentA, treatmentB, ...). You have to decide on an effect size of interest. It seems you have the estimated effect sizes from the different studies. It is critical that you also have the so-called sampling variance for each study for each effect size (these are just the within-study variances of the effect sizes). For instance, if your effect size was the mean, then you need the estimated variance of the mean for each study, not just the means. You need to fix these 'sampling variances' in the analysis, which is done through some clever use of weights in MIXED. You also estimate the among-study variance (at a minimum) in the analysis. I can't give you code until you give more details.