Why do i get "Convergence Questionable: check iterations" when running standard least squares mixed model?

Question

Hi All,

I am a biomed PhD student, and hypothesise that a particular cell type increases at with different length of drug treatments (e.g. comparing PBS control to d12 of treatment to d19 of treatment). I have used JMP pro12 and run a standard least squares model. I have performed two experiments to test this, and have set the experiment code as a random effect. The other effect is treatment.

I am looking at the Fixed effect test -> effect details -> LS means differences Tukeys HSD -> ordered differences report to see if the treatments are significantly different. The fixed effect test has a dot instead of a p value in the prob>f column, and there is also dots in the Lower CL/Upper CL columns. I am therefore not sure whether to believe the p-value (<0.0001).

I think this might be to do with the "Convergence Questionable: check iterations" error?

Do you know how to fix this? Can I trust the p values and say that there are significant differences between all my groups or is my data not suitable for this test?

Thanks a lot!

Emma

Phil_Kay · Answer

That all makes sense, Emma.

emmahoulder · Answer

Really helpful, thanks a lot. I got the following reply from JMP technical support that has answered my question I think:

Hi Emma,

Your question about the failure to converge was passed over to JMP Technical Support.

In examining your model, I noticed a few things:

At the last iteration, the variance component estimate for Expt was negative, suggesting there may not be enough evidence of variability between Expt's to model that random effect.

Indeed, when plotting the data, there does not seem to be obvious evidence of any difference between the Expt's.  The graph also emphasizes the uncertainty in the estimates for the treatment combinations in which there are only a few rows of data.

I think this is the cause of the instability and failure to converge to a solution.

Notice if you de-select "Unbounded Variance Components" on the model dialog (bound the vc's), then the model will converge to a solution where the Expt random effect variance is estimated to be zero.

That model would be equivalent to the model where you remove the Expt random effect and only fit the fixed treatment effect. I think that would be a reasonable approach here.

I hope this information is helpful. Please let me know if you have follow-up questions, or need clarification.

Thanks,

Adam Morris

JMP Technical Support

Phil_Kay · Answer

A couple of other ideas:
1. At JMP we always say you should plot the data before going to a statistical analysis. We designed JMP for this. When you plot your response vs treatment does it look like there is a big difference? Do the data look sensible? Any strange outliers?
2. I would like to plot the distribution of the response. From a quick look at your screenshot of the table it seems that it is varying over orders of magnitude. So a transformation of the response may be useful for modelling.