Statistical methods for establishing the "insufficiency" of data cutoffs

research methods study interpretation
1

Encountered an interesting article here:

I have done a bit of reading about mitral regurgitation. Like MR, Tricuspid regurgitation classifies intervals of regurgitant volume and regurgitant area to create grades. In the clinical setting it is the grades that are used to render decision making: e.g. do nothing, medical management, surgical repair.

Any biostatistician can tell you: a continuous marker is more powerful than a cut-off, especially when the thresholds are based on univariate statistics like percentiles, and especially especially when the cut-offs are estimated from small, underpowered, and/or hypothesis generating samples.

The approach they use is fitting a sequence of models for mortality with possible adjustment for confounders in a Cox model. The base model adjusts for the 4 level grade as an ordinal variable, and then subsequent models adjust for the continuous markers, separately. The authors inspect the diminishing p-values and conclude the continuous markers are better.

My conclusions based on this research

  1. It’s not surprising to find statistically significant results in a sample of almost 400 participants.
  2. The participants with extreme values who belong in the grade 4 category are definitely going to be sicker than those with borderline grade 4 values.
  3. If I were a clinician I would probably not revise my strategy to care except to bear in mind the thresholds that contribute to grading and that intermediate values of grade might lead me to favor more or less aggressive therapy as needed.

My questions for the community:

  1. What is a more compelling, comprehensive, and applied approach to using data analysis to suggest that grading criteria should be revised or removed altogether?
  2. If grading and cutpoints are mandatory, what is a data-optimal way of identifying such endpoints when treatment is a mediating factor?
  3. What types of analyses would you use to predict the outcomes of a different therapeutic approach when the actual therapeutic approach is based almost entirely on established cut-points?
1 Like