I think that would be much more difficult (athough perhaps other readers know an easy way). The problem is that there are no real slices to make for the graph (all combinations of predictions for all continuous variables would not be in the file). ... View more

You should make the plots based on the output file you created with your code. The User's Guide gives a example that is almost the same as what you want. http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_glimmix_sect027.htm Of course, since you have binary data, the observations will consist of points across the bottom and top of the graph, with the predictions as smooth curve that goes from 0 to 1 (on the ilink scale). ... View more

I think your last posting gets at the critical point. Using the NLMIXED code (or equivalent now in GLIMMIX for this example), the within-study variance is automatically determined based on the specified distribution. One does not fix the within-study variance. If the data have a Poisson distribution, the variance is determined by the mean (by definition); i.e., the variance is completely known, and there is no reason to fix it in each study. The software takes care of this. This is why the approach you are taking is difficult -- you are specifying a discrete distribution (NB or Poisson), but you want to fix the within-study variances at something other than what is defined for those discrete distributions. You are essentially getting a overdispersed distribution scenario, where the degree of overdispersion (or possibly underdispersion) varies with the study. I attemped one workaround, where you take into account the internal variance structure (determined by the software), and then rescaled it to achieve YOUR fixed variances. However, I think you would be better off using the normal approximations, because this is easier to control the within-study variances. ... View more

It looks like you are on the right track. The results are not expected to be the same. You can't compare log-likelihoods because pseudo-likelihood is used for the Poisson distributions but true restricted likelihoods are used for the normal distribution. ... View more

You are on the right track based on my previous comments. The difference is still probably due to the estimated scale for the chosen distribution. But the more I think about it, I don't think that Poisson or NB are appropriate for your analysis. Even if the original data in each study are discrete, either Poisson or NB, you are doing the meta-analysis on an estimated parameter from each study. This parameter estimate would be continuous, and probably approximated described by a normal distribution. Your within-study variance (or its inverse, the weight) is based on large-sample (normal) theory. Thus, I think you could avoid a lot of confusion by directly analyzing the parameter estimate for the studies using normal-based approaches. You can continue with use of GLIMMIX, but base the analysis on LOGtox. The weight should be the inverse of the within-study variance for the logtox for each study. You still need a parms statement, so that the residual is held fixed at 1. The combination of the weight and the parms statements gives you fixed within-study variances and an estimated between-study variance. The output says that the residual is 1, but because of the weights, the residual is actually different for each study (=1/weight). proc glimmix data=count_data; class study; model LOGtox =  / offset=logPA2 solution;   random intercept   / subject=study ; * to test heterogeneity between study; weight inv_varln; * weight = inverse variance; parms (1) (1) / hold=2; run; You can read more at: http://www.lexjansen.com/pharmasug/2000/stats/st09.pdf (this was written before GLIMMIX was available, but it is very useful).  There are also other ways of using a long parms statement and no weight. ... View more

There are several things to keep in mind with meta-analysis. First of all, with fixed or random effects meta-analysis, you need to fix the within-study variances. With normal distributions, this is straight-forward  by first fixing the residual at 1 and then using weights. (This is easily accomplished, using either mixed or glimmix (see my other posts on a similar topic)). Or, one specifies a separate within-study (i.e., "residual")  variance for each study and holds all of these fixed. You are trying this approach here, but there are some complications, so I don't think you are getting what you want. With a discrete distribution (Poisson or negative binomial (NB)), the within-study ("residual") variance is defined based on the distribution (=mean for Poisson, =(mean+(scale)*mean^2 for NB). Here, "scale" means the within-study scale parameter for the NB (equivalent to 1/k in other parameterizations). This residual scale term is on the scale of the raw data, not the link (but the other random statements refer to the scale of the link). You have the pre-specified within-study variances, but I don't believe they are being substituted for the above-described within-study variances, using your coding. Rather, I think with the first two approaches (your latest post) you are getting the internally-defined residual variance multiplied by the listed variance parameters (or the equivalent with the weights). This gets more complicated because, by using the NB distribution, the program is also estimating a NB scale parameter based on data across the studies (that is the scale parameter being displayed). Thus, it appears that none of your choices is really fixing your within-study variances at the values you want. I have done a lot with meta-analysis, but either with normal data, or with non-normal data but without fixing the within-study variances (using inidiviudal data within the studies). I have definitely not tried this using the NB distribution (with its complication of estimating a separate scale across the studies).  I don't have an exact suggestion right now, but will think about it more. Here is another complication. If you substitute a variance for "residual", then you no longer have a true Poisson or NB, since these are defined, in part, by the variance:mean relation. Here is one approach to consider. Choose Poisson, and determine the "residual" variance based on the mean. Then re-scale your listed within-study variances, so that the product gets you back to the desired within-study variances. Then use your method 2 (last post). I need to think about this some more. By the way, your approaches 1 and 2 are closest to what you want (with the qualifiers that I list here). ... View more

HPMIXED is only for normal data. ... View more

You are getting the default choices for predicted values and residuals. These are considered conditional values (conditional on the random effects). Thus, you are getting the predicted value (predicted linear predictor) for {x1,x2,x3} for each hosp value in the data set. These are considered BLUPs (actually, EBLUPs). Even if the independent variables were the same, the predicted value would be different for two different hosp values. The residual is just the difference with the observed (calculated on the link scale). If you want the so-called marginal predictions (i.e., at the expected value of 0 for the random hosp effect), use:      output out=gmxout2 pred(noblup)=pred residual(noblup)=residual; Check out the following for more guidance, or the User's Guide. http://www2.sas.com/proceedings/sugi30/196-30.pdf ... View more

You will probably not get GLIMMIX to run with that many observations. I suggest you take a random sample of the customers (keeping all the observations for the selected customers), and do the analysis on this sample. Note: you did not define ID as a factor using a class statement. That is OK, as long as the data are sorted by ID before running GLIMMIX. If not sorted, you could get incorrect results (when you have a small enough sample to get any results). ... View more

Use: model y = height / dist=bin link=logit s; random int height / sub=center type=un; random _residual_; This will give you a random coefficient model (with random effects for intercept and slope). ... View more