topic Re: GLM in Statistical Procedures

GLM

FREDY_07 — Tue, 08 Mar 2016 18:10:08 GMT

Hi, I would like to create a model by using input x which is a matrix and output y which is a vector. I saw several examples for GLM but using input variables x1 ,x2,.. so on. But I have one variable x which is a matrix. Many thanks.

Re: GLM

Reeza — Tue, 08 Mar 2016 19:08:26 GMT

If you have matrix, you may want to look at IML.

Otherwise, post a sample of what your data looks like.

Unless your matrix is Nx1, it's likely your data structure will support GLM.

Re: GLM

Rick_SAS — Tue, 08 Mar 2016 19:28:21 GMT

Could you describe your data further? How does the matrix enter into the model?

Re: GLM

FREDY_07 — Tue, 08 Mar 2016 20:39:31 GMT

I have 55 input variables in the data and 20 samples. so its 20*55. and y is a 1*55 matrix. so it is a vector. Many thanks for helping.

Re: GLM

FREDY_07 — Tue, 08 Mar 2016 20:40:54 GMT

I am so sorry y is a 1*20 matrix.

Re: GLM

Ksharp — Wed, 09 Mar 2016 01:19:07 GMT

Here is an example .

http://blogs.sas.com/content/iml/2016/03/02/dummy-variables-sasiml.html

Especial this part:

Y = Cholesterol;                 /* response variable */
Intercept = j(nrow(Y), 1, 1);
X1 = designf( Sex );
X2 = designf( BP_Status );
X = Intercept || X1 || X2;      /* design matrix with EFFECT parameterization */
 
/* Matrix formulation of one-way ANOVA (cell means model): Y = X*beta + epsilon
   See https://en.wikipedia.org/wiki/Design_matrix       */
b = solve( X`*X, X`*Y );       /* solve normal equations */
print b[rowname={"Intercept" "Sex:Female" "BP_Status:High" "BP_Status:Normal"}];

Re: GLM

Reeza — Wed, 09 Mar 2016 02:28:55 GMT

Are your matrices SAS data sets - .sas7bdat files?

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 04:11:51 GMT

No. It is a csv file.

Re: GLM

Reeza — Thu, 10 Mar 2016 04:27:07 GMT

Please post a sample of your data. I think you can use proc glm.

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 04:43:08 GMT

Many thanks for your response. I looked at the example that you provided. Hopefuly it would work.

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 12:46:09 GMT

I attached the data. There are 20 samples and 55 variables in the data. here is the output: Many thanks.

8.14

5.77

5.89

4.78

4.48

3.96

5.1

5.57

5.27

4.2

5.23

2.94

4.77

4.67

4.8

5.21

5.28

3.91

4.21

5.17

Re: GLM

Reeza — Thu, 10 Mar 2016 15:43:17 GMT

This is how the code could look, if you wanted a simple GLM model:

*import dataset into SAS;
data have;
infile 'C:\_localdata\data.csv' dlm="," dsd truncover lrecl=1000;
informat var1-var40 best32.;
format var1-var40 best32.;
input var1-var40;
run;

*Run model;
proc glm data=have;
	model var1 = var2--var40;
run;quit;

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 17:47:47 GMT

Many thanks for your response. I have some questions for you. Where is the output variable 'y'. Also why did you take 40 variables is there a specific reason for that?

Re: GLM

Reeza — Thu, 10 Mar 2016 18:09:52 GMT

Your file had 40 variables, it was a demo of how it could be accomplished.

It had no headers so I don't know which is y, which is x etc.

I simply took the first variable as the Y and the remaining as X to demonstrate the GLM procedure.

To develop a final model you might want to use a selection method to subset the variables. PROC GLMSELECT is helpful here.

Please consider my code a sample, not what you should do. You should know what your variables are and how they should be handled.

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 18:17:19 GMT

Many thanks for your response I understood you mean.

My bad. Mistakengly I put 55*40 input. It should be 55*20 but I see you mean. The 40 is the sample number. I do have 55 variables so is the following right?

proc glm data=have;
	model y = var1--var55;
run;quit;

I don't now what GLMSLECT is doing at the final stage. I will research for it. Many thanks for your responses.

Re: GLM

Reeza — Thu, 10 Mar 2016 18:28:28 GMT

You had 55 observations, not variables.

The columns are variables not rows

In general, your code looks correct, assuming that matches your data structure.

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 20:19:16 GMT

Thank you so much so I need to transpoze my data in order to make right model. Many thanks.

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 20:26:34 GMT

I have got the model but it produces very impossible results. R Square equals to 1. and F equals to infinitive:((

Re: GLM

Rick_SAS — Thu, 10 Mar 2016 20:45:35 GMT

Since no one else has mentioned it....

With 55 variables and only 40 observations, you have an overspecified system. There are enough parameters to fit the data exactly, which is why you are seeing R-square=1. However, the model is meaningless.

An analagous example is asking for the "line of best fit" through one data point.

Re: GLM

FREDY_07 — Thu, 10 Mar 2016 21:24:51 GMT

Many thanks. so you mean like finding a hyperline, right?