topic Re: Variable selection in Proc Logistic - dealing with missing values in Statistical Procedures

Variable selection in Proc Logistic - dealing with missing values

Xiaoningdemao — Thu, 28 Jul 2016 22:15:33 GMT

Dear All,

Here I have a question about variale selection. Say I have the following data set:

Age Sex weight Pass

15 0 70 1

17 1 - 0

16 - 60 1 ;

And I want to use logistic regression to predice pass using age, sex and weight . also I want to use backwards variable selection to selsect significant covariate. So I coded as following:

proc logistic data=have descending;

model pass= age sex weight /selection=backward fast;

run;

However, since I have missing data here. The program only use the entry that has complete observation based on full model, therefore the first enrty.

But I want it to enclude more observation as eleminate covariates, that is when model only with age and sex, i want it to use first two entries. Is there a way to realize this in SAS?

Thank you very much!!!

Best,

HereH

Re: Variable selection in Proc Logistic - dealing with missing values

ballardw — Thu, 28 Jul 2016 22:47:27 GMT

The procedure will only use records that have values for all the model variables.

You might try imputation to replace missing values. There are a number of ways but if you have significant percentage of missing values for a given variable then your resulting model is going to be suspect.

What percentage of your records are missing one or more of the variables?

Re: Variable selection in Proc Logistic - dealing with missing values

PGStats — Fri, 29 Jul 2016 02:25:43 GMT

You could screen your predictors with PROC HPSPLIT. It offers three methods for dealing with missing predictors. The output decision tree is a crude model but will show you which variables are the most important predictors.