Excellent answers. A few other thoughts:
- Include a graph showing the fits with confidence bands, and spike histograms along the x-axis to show the data density for O_2.
- Aside from the fact that the number of knots is the most important aspect of restricted cubic spline modeling, we tend to put knots where we can afford to put knots. Hence the use of quantiles for default knot locations. The spike histogram mentioned above will help on this count.
- Most analysts would just force linearity on the predictor, so you’re way ahead of the game.