Programming the statistical procedures from SAS

NLoptions choice of technique

Accepted Solution Solved
Reply
Contributor
Posts: 24
Accepted Solution

NLoptions choice of technique

Hello,

 

can anyone tell me what small, medium and large problems in terms of sample size is?

When I went through the different optimization techniques for the nloption statement in

http://support.sas.com/documentation/cdl/en/stathpug/66410/HTML/default/viewer.htm#stathpug_hpnlin_d...

I found QUANEW is recommended for medium problems whil newrap and nrridg are suitable for small problems. Any experience how this translate into number of observations?

 

Thanks in advance and regards,

M


Accepted Solutions
Solution
‎09-20-2016 02:35 AM
SAS Super FREQ
Posts: 3,307

Re: NLoptions choice of technique

That question is  too vague.The complexity of the model (linear model vs nonlinear; fixed effects vs  random) plays a role, so what SAS procedure, method, and options are you trying to use?

 

Nevertheless, here are a few thoughts.

 

The size of an opimization problem usually refers to the number of parameters that you are opimizing.  For most regression-type problems that opimize the likelihood function, this means that number of effects in the model. Each classfication effects requires k-1 parameters, where k is the number of levels in the categorical variable.

 

The number of observations are important because that is how you form the X`X matrix that is used in regression. Many SAS procedure can use multithreaded code to fit this matrix, so the number of threads that you use can be important. MLE also has to run through the observations for each iteration. For generalized linear models, SAS will zip through hundreds of thousands of observations easily.

 

If I am optimizing a nonlinear function by using a SAS/IML NLP routine, I consider the poblem to be small if it has less than a dozen parameters. Medium problems might have a few dozen parameters, and large is more than that. I don't usually worry about the number of observations.

 

I think this is a good time for me to post a disclaimer: I work for SAS but I do not speak for SAS. If you provide more information about the procedure and syntax, we can provide better answers that are tuned to your problem.

 

 

 

 

, and the Small problems might generally be

View solution in original post


All Replies
Solution
‎09-20-2016 02:35 AM
SAS Super FREQ
Posts: 3,307

Re: NLoptions choice of technique

That question is  too vague.The complexity of the model (linear model vs nonlinear; fixed effects vs  random) plays a role, so what SAS procedure, method, and options are you trying to use?

 

Nevertheless, here are a few thoughts.

 

The size of an opimization problem usually refers to the number of parameters that you are opimizing.  For most regression-type problems that opimize the likelihood function, this means that number of effects in the model. Each classfication effects requires k-1 parameters, where k is the number of levels in the categorical variable.

 

The number of observations are important because that is how you form the X`X matrix that is used in regression. Many SAS procedure can use multithreaded code to fit this matrix, so the number of threads that you use can be important. MLE also has to run through the observations for each iteration. For generalized linear models, SAS will zip through hundreds of thousands of observations easily.

 

If I am optimizing a nonlinear function by using a SAS/IML NLP routine, I consider the poblem to be small if it has less than a dozen parameters. Medium problems might have a few dozen parameters, and large is more than that. I don't usually worry about the number of observations.

 

I think this is a good time for me to post a disclaimer: I work for SAS but I do not speak for SAS. If you provide more information about the procedure and syntax, we can provide better answers that are tuned to your problem.

 

 

 

 

, and the Small problems might generally be

☑ This topic is SOLVED.

Need further help from the community? Please ask a new question.

Discussion stats
  • 1 reply
  • 134 views
  • 3 likes
  • 2 in conversation