On p. 26 of Regression Modelling Strategies (2nd ed.) Frank Harrell argues against letting the knots in a RCS be free parameters and for setting them in advance, at various quintiles.
He cites several reasons including instability of estimates, inference problems and a lack of software.
I was wondering if this is still true? That is, have there been developments in algorithms and software that alleviate some or all of these problems?
