topic Re: How to score a data set with linear mixed effect model in SAS Forecasting and Econometrics

How to score a data set with linear mixed effect model

NaNaN — Fri, 20 Apr 2018 19:28:17 GMT

Hi there,

I read a paper about applying mixed effect model to predict a fractional response attribution with the range of [0,1]. Please see the code in below:

%let level=obligor_id;
proc sort data=MyData;
 by &level;
run;

proc nlmixed data=MyData noad qpoints=80 tech=quanew maxiter=3000 maxfunc=3000
 qtol=0.0001;
 parms b0-b14=0.0001
 gamma1-gamma2=0.4;
 cov_mu=b0+b1*Var1+b2*Var2+…+b14*Var14;
 con_mu=cov_mu+gamma1*z;
 con_sigma=gamma2**2;
 model RR~normal(con_mu,con_sigma);
 random z~normal(0,1) subject=&level;
 predict con_mu out= mix_output_&level (keep=instrument_id RR pred);
run;

For my problem, I replaced the obligor_id with account_num, which is unique in my dataset. The result returned from this model was really good. The problem now is how to apply this model to score a out-of-time data set with the random effect?

Looking forward to any suggestions.

Thanks!

-Shelly

Re: How to score a data set with linear mixed effect model

Rick_SAS — Fri, 20 Apr 2018 20:13:08 GMT

You can use the "missing value trick" to score additional data in conjunction with the PREDICT statement that you are already using. Some caveats and examples are discussed in another thread and on the excellent UCLA statistics site.

Re: How to score a data set with linear mixed effect model

NaNaN — Fri, 20 Apr 2018 20:32:44 GMT

Thanks for your suggestion! I’m not sure how this missing value trick can be applied if I want to implement this model in a system that can score in live.

Re: How to score a data set with linear mixed effect model

Rick_SAS — Fri, 20 Apr 2018 21:04:16 GMT

I don't think that trick works for new data. You'll have to write the DATA step code to score the model from the parameter estimates.

Now that I look at your model, isn't it linear? If so, you can switch to PROC MIXED and use the STORE statement to save the model and PROC PLM to evaluate the model on new data.