topic Warning: "Did not converge" in Proc mixed in Statistical Procedures

Warning: "Did not converge" in Proc mixed

NBriggs — Sat, 22 Dec 2018 19:40:56 GMT

I am trying to analyze a blood serum in a 6x6 latin square design and 7 different time points. I am missing 5 data points, however I had this syntax work for 4 out of 6 minerals (Everything but Mn and Co) set up the same with the same data points missing. This is the syntax I have been using, which worked for 4 other minerals set out in the same manner. I am unfamiliar with SAS and have been trying to troubleshoot on my own and it has been difficult. I have also used type= un@ar (1) with some luck on the others.

title1 'Serum TM Data'; 
proc mixed data=one covtest cl ; 
class Period Cow Time Trt; 
model Mn = Period Time Trt Time*Trt/ddfm=kenwardroger; 
repeated Period Time /subject=cow type=un@cs r rcorr; 
lsmeans Period Time Trt Time*Trt /adjust=tukey; 
ods output diffs=ppp lsmeans=mmm; 
ods listing exclude diffs lsmeans; 
run;

Is there anyone who can advise what I could change?

File is attached.

Re: Warning: "Did not converge" in Proc mixed

StatsMan — Mon, 24 Dec 2018 13:30:56 GMT

Check the iteration history. Is the model marching towards convergence? If so, then you may need to increase the number of iterations to allow MIXED to reach convergence.

If the convergence criteria is bouncing around, then the model just may be too complicated for this response. Try simplifying the mean structure of the model and see if that helps. You may also need to simplify the covariance structure, although I understand the structure you are using makes sense for the design.

Re: Warning: "Did not converge" in Proc mixed

sld — Fri, 28 Dec 2018 03:31:11 GMT

The model converges if the response variable is log(Mn).

However, I will take @StatsMan's comment a bit further and suggest that your design (only six subjects for a six-period, six-treatment Latin square design) may be (and probably is) much too small to be useful or reliable. I would be concerned about balance and carryover, especially because of a period effect (Mn decreases over periods). A model that converges is necessary but may not be sufficient. It is convenient but treacherous to underestimate the complexity and needs of Latin square/crossover designs.

I hope this helps.