Be it logistic or survival analysis/cox regression, there is utility in determining cutoff points to categorise a continuos risk factor into various risk strata.
However there is a plethora or methods to go about defining the 'optimal' value depending on the nature of the model outcome. Generally they fall into methods that look at the model's sensitivity and specificity or methods that look to maximise the p value of significant difference between the survival curves of the resulting stratas.
Currently, i'm modeling the risk of body mass index on diagnosis of certain medical conditions through survival analysis to identify at risk individuals since the current bmi risk groupings prove inadequate to stratefy risk groups. For my purposes, i'm thinking of going with a method to maximise sensitivity with maintaining a certain threshold of specificity ,instead if going eith the standard procedure to maximise the youden index , since the cost of misclassifying in the model isn't a great deal.
The issue im facing is how to justify the specificity threshold chosen? Or is there a better method of determining the best cut off to use for catorising bmi risk groups. Also , how to determine the number of categories?
The implementation of this in R is also shrouded in difficulty and the packages i found are unclear in the innerworkings of the functions.
Anyone able to shed some light on this grey situation?
Don't categorise a continuos risk factor into various risk strata.
Simply don't do that. It's bad, it's medival. Even for clinical practice, the decision (about what to do with a patient) depends on many variables. If the information from these variables should be formally integrated into a decision process, the physician will need help of a software anyway and "simple-to-remember" risk classifications are not required. If the information is used more informally, risk-classes are not taken literally (and anyway being related to what is known else about the patient), so any kind of "optimality criterion" for the classification is actually irrelevant. I am convinced that even then a simple (and "easy-to-remember") quantitative rule to get a risk (ideally really as a conditional probability value) is of more practical use.
Thanks jochen , I think i see where you are coming from. However i think there could still be utility of the risk stratification for the public use. Especially since the curent WHO categories do not stratefy the risk groups well in my population (Hazard ratios are very similar between groups) . Having readjusted risk groups would provide a better rough guage for the public to set bmi goals... if implemented in healthcare it will be used more informally to recommend patients to programs aimed to modify lifestyle choices.
I need to ask about what is minimal number of cases needed to perform multivariate cox analysis
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS