topic Re: Y/N(char) variable to 0/1(num) with an option in Regression in SAS Programming

Y/N(char) variable to 0/1(num) with an option in Regression

Cecillia_Mao — Tue, 24 Aug 2021 14:27:32 GMT

Hi everyone,

I'm doing a logistic regression. However, I need to convert about 20 independent characteristic variables to numeric ones in order to check assumptions. I have seen something like an option(probably in a proc step) to change yes/no to 0/1 in just one step (Not if then, not proc format, not array), but I totally forgot where it came from. I tried my best to search similar stuff( "categorical variable to 1 2 3 in one step",etc), but I haven't found anything like an option that can do the job. I'm sorry that I can't provide more information, any hints would be APPRECIATED!

Re: Y/N(char) variable to 0/1(num) with an option in Regression

ChrisNZ — Tue, 24 Aug 2021 02:39:12 GMT

Like this?

NUM=(CHAR='yes');

Re: Y/N(char) variable to 0/1(num) with an option in Regression

PaigeMiller — Tue, 24 Aug 2021 11:43:04 GMT

@Cecillia_Mao wrote:

I'm doing a logistic regression. However, I need to convert characteristic variables to numeric ones in order to check assumptions.

What assumptions? There are not usually any assumptions on character variables, and so no checking is required. So, no conversion to numeric is required either.

Logistic regression will run just fine with a character variable as the response.

Logistic regression will run just fine with character x-variables if you indicate the variables are character in the CLASS statement.

Again, no conversion from character to numeric is needed for logistic regression. Do the easy thing in this case, use your data as is.

Re: Y/N(char) variable to 0/1(num) with an option in Regression

Tom — Tue, 24 Aug 2021 12:06:59 GMT

Do you really need to create dummy variables? Can't you just use a CLASS statement in your analysis step?

Perhaps you are thinking of the GLMMOD procedure? That should be able to generate dummy variables from your CLASS variables.

Re: Y/N(char) variable to 0/1(num) with an option in Regression

Cecillia_Mao — Tue, 24 Aug 2021 14:13:43 GMT

Hi Paige,

Thanks for the reply! I'm checking the MULTICOLLINEARITY assumption, and I'm using the "Correlation matrix" criteria, which means I need to use “proc corr ” that needs numerical variables. BTW, this is the material I'm referring to
https://www.lexjansen.com/wuss/2018/130_Final_Paper_PDF.pdf .

Re: Y/N(char) variable to 0/1(num) with an option in Regression

PaigeMiller — Tue, 24 Aug 2021 14:18:39 GMT

Perhaps something needs to be clarified in your original post.

Are these character variables that you need to convert to numeric x-variables or y-variables? How many are there?

Re: Y/N(char) variable to 0/1(num) with an option in Regression

Cecillia_Mao — Tue, 24 Aug 2021 14:21:52 GMT

Thanks so much, Chris! It's not exactly what I'm looking for, but it's much better than if-then and proc format in some cases! Thanks for sharing!

Re: Y/N(char) variable to 0/1(num) with an option in Regression

Cecillia_Mao — Tue, 24 Aug 2021 14:26:31 GMT

Thanks for pointing that out, it had been a little bit confusing. I revised the original post a little bit. They are about 20 independent character variables that need to be converted to numeric ones.

Re: Y/N(char) variable to 0/1(num) with an option in Regression

Cecillia_Mao — Tue, 24 Aug 2021 14:30:40 GMT

Hi Tom,

Thanks for the reply! I made the case a little bit confused in the original post. If only considering running the model itself, the class statement would be enough. However, I just want to change character variables to 0,1,2,3... for assumption check, which sometimes requires the data in numeric format. Thanks again!

Re: Y/N(char) variable to 0/1(num) with an option in Regression

PaigeMiller — Tue, 24 Aug 2021 14:37:08 GMT

@Cecillia_Mao wrote:
Thanks for pointing that out, it had been a little bit confusing. I revised the original post a little bit. They are about 20 independent character variables that need to be converted to numeric ones.

So you really ought to create dummy variables, and if there are more than two levels of a specific variable, this is more complicated than the suggestion from @ChrisNZ . Here is a long thread on many different ways to create dummy variables.

Also, checking multicollinearity is not checking "assumptions", so that threw me off. An assumption in modeling would be something like does a variable have the proper distribution. Multicollinearity is certain a concern, but it isn't what I think would be called an "assumption".