Apply quadratic model from regstats to predict unknown values

Here's the setting - you get the quadratic model 'beta' from regstats using:

 beta = regstats(Y,[X Z],'purequadtratic','beta');

where Y, X and Z are n by 1 vectors and X and Z are explanatory variables used to predict Y.

I understand how to extract residuals and yhat from regstats. What I want to know is whether there is a function that will allow me to predict values of Y for additional values of X and Z.

I've figured out how to do this manually, but not being particular tuned-in to maths, it took me a while to discover the formula for a quadratic model. Is there a Matlab function for this procedure? What have I missed in the documentation?

Best regards,

Sam

----------------------------- 7.11.0.584 (R2010b) Win 2007 - 64 bit HP xw4600

Answer 1

[Sam Roberson]

Hey Grega,

I'm glad you asked this question, because I now know 2 different solutions. The first solution is the manual solution that I worked out myself, which I won't give here. The second is from a Mathworks workshop I went to, and looks like this:

X = x2fx(predictors,'quadratic');
y = responses;
[Q,R] = qr(X,0);
beta = R\(Q'*y);
yhat = X*beta;
residuals = y - yhat;

where predictors is a p by m vector or row matrix of data and responses is a p by 1 vector. This is all given in the code for regstats. I should have looked there first it seems.

Best,

Sam

Answer 2

[grega]

Hey Sam!

I have similar problem with 'interaction' and 'quadratic' models.

Have you figure it out please?

Thank you!