treat identifiability by requiring centered feature functions

[dswah]

pg 164 in Wood introduces this concept. elaborated in appendix A.5

[dswah]

some more posts https://stats.stackexchange.com/questions/110472/how-exactly-is-the-sum-or-mean-centering-constraint-for-splines-also-w-r-t-g?rq=1 https://stats.stackexchange.com/questions/33366/centering-constraints-in-mgcv-gam?rq=1

[ashrafabdul]

@dswah From what I understand, when using regression splines this can be enforced by setting the intercept parameter of the shape function to zero. However, I am not sure how this can be achieved with bsplines. I wonder if this can this be done post hoc. Once we have the fitted shape function, can we shift each function by some constant value such that expectation of the shape function over the covariate is close to zero. We could then add all the individual shifts to the global intercept term and finally reevaluate the model performance. What do you think about this?