In Matlab when you do polynomial regression using polyfit and then evaluate the predicted trend using polyval, you can use error estimation structure S to obtain error estimates, which allows you to draw confidence intervals at any given confidence level. However, it differs from typical method of adding +/- 1.96 standard deviations for 95% confidence bounds in that typically the distance between a bound and prediction is constant, and here it differs for every x (so as if standard deviation was computed separately for every x value).

I want to compute confidence intervals for multiple regression of certain dataset (so two or more input variables, one output variable). How to do it like polyfit does? Right now I can just offset the "prediction surface" by +/- 1.96 SD to estimate 95%CI, but I wanna do better.