Research Methods used to provide evidence for vancomycin therapeutic range

In clinical practice, vancomycin is dosed to achieve either a trough target or an AUC target ( usually estimated using Bayesian PK software). The range for trough or AUC comes from evidence from observational studies with the following methodology:

  1. Using CART analysis find the threshold for some outcome( therapeutic failure ( various definitions, mortality, AKI ( without specifying what degree of AKI)
  2. The CART derived threshold is entered into a logistic regression model to estimate odds ratio adjusting for important clinical covariates.

The thing is that various studies reported various thresholds some were found to be statistically significant and others not in the multivariate logistic regression.
Will be glad to have your opinion on this method ?

This method is to be avoided at all costs. Recursive partitioning (CART) is essential data torture and will result in a major bias. More to the point there is unlikely to exist a threshold against an outcome. I’m sure it’s a smooth relationship so that a threshold search is futile.


Thanks for your reply. Both the current recommend therapeutic range and the current recommend time at which it should be achieved are based on studies which used CART analysis.
Another problem with that is that the same dataset used to find the CART threshold is then used to test the hypothesis that this threshold is a significant predictor of the outcome.

Of all the things to turn into an arbitrary discontinuous relationship, dose-response relationships are not good examples.

1 Like

So would you enter AUC0-24h for vancomycin into a logistic regression model as a continuous variable using restricted cubic splines?

Yes, if it’s an independent variable.