Programming the statistical procedures from SAS

Rescale the data values for convergence

Reply
Occasional Contributor RVS
Occasional Contributor
Posts: 5

Rescale the data values for convergence

Hello SAS community,

 

I was reading one of the SAS global proceeding papers related to linear mixed models. That paper suggested to rescaling the data is one of the options to trouble shoot convergence issues in parameter estimation. I'm working on large dataset that consists of multple years and locations, and response variable is yield measured in tons/hectare. Since this dataset requires to estimate a large numbe of parameters, convergence is often a problem to estimate the parameters.  I was thinking to rescale yield values to kgs or some other form.  

 

Any suggestions or comments?

 

Thank you,

RVS

Grand Advisor
Posts: 10,073

Re: Rescale the data values for convergence

It might help to post a link to the paper so we can see which approach you may be contemplating.

 

Occasional Contributor RVS
Occasional Contributor
Posts: 5

Re: Rescale the data values for convergence

Thanks for the response. Below is the link for the paper. 

http://support.sas.com/resources/papers/proceedings12/332-2012.pdf

 

I'm refering to the Section II ("TROUBLE SHOOTING CONVERGENCE FAILURES IN MIXED MODELS") from the above paper. 

 

Valued Guide
Valued Guide
Posts: 679

Re: Rescale the data values for convergence

Tons per hectare should be a good scaling. Vaules are typically between 2 and 10, so there are not extremes. Converting to kg per hectare would probably cause more problems, considering the values would be ranging up 10000 or so. The variance-covariance matrix would have lots of very large numbers (for variances), and covariances which could be very small or very large.

Respected Advisor
Posts: 2,655

Re: Rescale the data values for convergence

I suspect (unfortunately) that the model may be overspecified for the number of observations.  Consequently, some random effect values are really sensitive to changes, and thus convergence becomes a problem.

 

My usual approach is to "sneak up on" the model I want to use, starting with something less complex, and working there until it converges and is meaningful in some sense.  Then complexity can be added which, in mixed models, is almost always something in the random (or repeated, if in PROC MIXED) statement.

 

Steve Denham

Ask a Question
Discussion stats
  • 4 replies
  • 219 views
  • 1 like
  • 4 in conversation