Unless you have multiple PR AUC estimates per model, or you can aggregate some models together, you probably won't be able to use the nonparametric tests you list. The same would be true of the parametric tests that come to my mind. Somehow, you need to find a measure of variability within your grouping variable/model in order to do any testing or generation of confidence bounds.   SteveDenham ... View more

Since subject is completely nested in time period, you can't fit an R side covariance structure. This explains the inability to calculate starting values that are consistent estimators. So what you have is a factorial design, and I don't see any reason to include random or random residual effects associated with time. You probably should retain the random effect of school. What happens in BGLIMM with that approach?   SteveDenham.  ... View more

Wage doesn't seem to me like a variable that would follow a Poisson distribution as it usually is not a discrete count. Since it is considered continuous, you might want to try an exponential distribution or a gamma distribution here and in GENMOD.   SteveDenham ... View more

When GLIMMIX stops (or maybe even doesn't start iterating), are there ERRORs, WARNINGs, or NOTEs in the log that indicate what might be happening? There may also be messages in the output file that can help diagnose the issues.  Let us know what these might be.   SteveDenham ... View more

You may need an offset (see Example 3 in the BGLIMM documentation). I would also recommend including an outpost= option in the PROC BGLIMM statement. The error message could be caused by having too many levels of sample as compared to the number of observations. Have you tried fitting the fixed effects model in BGLIMM rather than GENMOD? Or tried a frequentist approach using PROC GLIMMIX? The latter may have better diagnostics for what is going on.   SteveDenham ... View more

You need to watch out for type=AR(1) or type=ARH(1) for interrupted time spacing as these structures assume equal spacing of the observations. A possible substitute would be a spatial power structure ( SP(POW)(t), where t is a continuous variable that is equal to the time at which an observation is made). As t approaches equal spacing, this structure approaches the AR(1). Other possibilities for the variance/covariance structure would be to try PSPLINE which fits a penalized B spline to a G side random effect or RSMOOTH which fits a radial smoother, again to a G side random effect. Depending on the amount of data and the number of timepoints you have, you might be able to fit an R side UNSTRUCTURED or CHOLESKY structure.  Without knowing what your data layout is, I hesitate to suggest the latter two, and would recommend the SP(POW)(t) structure, unless the interruption is huge compared to the intervals where you have observations..   SteveDenham ... View more

Time for a sensitivity analysis of the NLMIXED results. Set up a fairly narrow gridsearch in the PARMS statement so that you can get an idea of how sensitive the parameter values are to the starting values. This may take care of your differences from the expected values.   SteveDenham ... View more

I'll hit the last one as it is one I have seen many times when fitting an R side effect to a binomial dataset.  The most likely cause of failure to converge is that you have run up against the default maximum iteration limit of 20 iterations. Try adding an NLOPTIONS statement with maxiter=5000 for example. If you still are not converging, your data is probably "pathological" with respect to the residuals.   Backing up one step, if you fit only an R side structure, the lack of Z columns is dependent solely on the syntax chosen, and the number of effects in your random / residual statement.   SteveDenham ... View more

Using the ilink option when the link is a log should result in a location estimate that approximates the median (geometric mean) rather than the expected value, but again it becomes a matter of distributional assumptions because it is not the same for gamma, inverse gaussian or log normal.   If you really, really want to look at differences in medians, you might consider bootstrapping as an approach.   SteveDenham ... View more

Given the various effect sizes I plugged in, 28 is going to give at least 80% power for a one-sided test, at an effect size of 0.45 (R squared, or.0.55 F squared) and alpha=0.05.   SteveDenham ... View more

Well, you did the analysis, so you should have everything you need to calculate a retrospective power (and don't get me started on why that is a bad idea.). I would avoid categorizing a continuous variable for this. This is actually a.ONECORR analysis, since regression is just fancy correlation. You mention an effect size of 0.45. If that is an f-squared estimate, what happens when you plug into this code:   proc power; onecorr dist=fisherz nullcorr = 0 corr = 0.557186 /* this is the square root of 0.45/(1+0.45), in other words, the square root of R-squared obtained by rearranging f-squared */ ntotal = 335 power = .; sides = 1 run; If your effect size is obtained some other way (omega squared, for instance), then ONEWAYMEANS is probably better suited to the task.   SteveDenham     ... View more

Hi (and particularly @StatDave 😞   Would it be possible to generate a bootstrap estimate of the residual variance?  Assume the range of X values is uniform. Assume the parameter estimates are normally distributed, with mean = the reported estimate and standard deviation = the reported standard error. Draw at random with replacement 100 values of X, and 1000 3-ples of the parameters (this is where a big assumption comes in - that the parameter estimates are uncorrelated). From these, calculate 100 x 1000 Y values. Last step, refit the equation at a range of (X,Y) values covering the interval, making sure to include X values of interest with Y set to missing. That should give predicted Y values with standard errors and confidence intervals, and I am sure quantiles could be calculated given the residual error.   Thoughts?   SteveDenham ... View more

It would be nice if you told us how it was "Done".   SteveDenham ... View more

If you add in DOSC, you have two repeated measures, most likely with correlations. See below for my first cut at code for this. One thing I noticed is that you have multiple dependent variables specified in your MODEL statement. SAS/STAT version 15.x documentation explicitly says: "The MODEL statement names a single dependent variable..."  (emphasis added).   So here is what I would try for a doubly repeated analysis:   proc mixed data=serology; class id gender group type dosc; model ilt2 = gender group type dosc/ solution; repeated dosc type /type=un@un subject=id r corr; run; What I would worry about here is the number of levels for dosc, how many unique days a single id is collected, and why you might think that dosc would have a different effect at different times of the year. If you have many levels of dosc, consider changing to PROC GLIMMIX where you can look at spline or polynomial effects for dosc, rather than soaking up a lot of degrees of freedom with a categorical approach.   SteveDenham   ... View more

See the PROC POWER documentation for code to do these.   SteveDenham ... View more