Struggling with -mkspline-

[Guest]

I'm struggling on interpreting the output after using the mkspline command.

I have a continuous exposure (N=721) called "norepi_equi_x" and want to model this independent variable using a competing risk model for three possible outcome: 1) event of interest, and (2) competing event, where participants are censored by 96 hours.

mkspline sp_vp = norepi_equi_x, displayknots cubic


knots are then located as below:

| knot1 knot2 knot3 knot4 knot5
-------------+-------------------------------------------------------
norepi_equi_x | .02 .05 .1 .2 1.0564


stset rtime, failure(rstatus==1) exit(time 96)
stcrreg sp_vp* age i.sex elixsum ap2_admit sofa_score_day_before_ext i.icu_class i.admission_type, vce(cluster icu_site) compete(rstatus==2)

I don't quite understand how to trouble shoot what is going with these non-sensical results.

The distribution of my exposure variable (and this captures the dose of a medication):

The first spline models the exposure as a linear relationship, which isn't appropriate for this setting. I don't now how to troubleshoot why each of the subsequent splines don't produce a result.

Thank you for your help.

[Nick Cox]

Your norepi_equi_x is strongly skewed -- and by default some knots will be very close together, as Stata is telling you -- so spline construction is problematic here.

You might try fewer knots or choosing knot locations deliberately.

Some experiments indicate that transforming it will help a bit, say by using asinh() as you have an exact zero.

Do you expect a complicated dependence on this predictor? Why not see how far adding a square term helps?

Most radically, if a predictor doesn't help, leave it out.