7 We have also run non-linear models treating the position on the scale as an ordered outcome.
We evaluated the predictors of increasing grade (severity) of CIN (on the ordered outcome variable: no CIN, CIN1, CIN2, and ≥CIN3) using bivariate (unadjusted) and multivariable ordinal logistic regression analyses.
To avoid over-fitting, the multivariable model was limited to 10 covariates based on frequencies within the ordered outcome variable.[23] For all models, continuous covariates were first included in the model using restricted cubic splines to avoid linearity assumptions.
The advantage of the PO model is its parsimony in dealing with an ordered outcome.
Therefore, we strongly recommend incorporating ordinal methods in the analysis of future trials when an ordered outcome measure is considered.
Thus, the simpler univariate model was chosen to show the magnitude where, given a specific percentage of bone loss, the probabilities of 0, 1, 2, or 3 complications were estimated as an ordered outcome variable.
We chose this simple variable because ordering by counting the number of events (for example, when using an ordered outcome variable as 0 v 1, 2, 3, 4, or more medication errors) is questionable because of the different characteristics and severity of events.
If the goal of the analysis is to assess the magnitude of the treatment effect on this ordered outcome, then an appealing approach is to assign numeric scores to the ordered categories and then to compare means between groups using conventional linear regression methods.
