gam in mgcv performs centering constraint at a common mean if smoothing a covariate for a categorical variable.
If I center the covariate at mean in advanced for each group, doing the analysis via s(x, by = group), and add them back in visualisation, does it mean the reference point is at mean of each group?