ROC analysis for repeated measures in Stata

[Horea Feier]

Hello Statalisters,

I have a dataset which contains repeated measures of independent variables (blood values) and I try to establish which of them is best suited for predicting the outcome (death=yes/no) using ROC analysis. I know that this has been proposed by Liu and Wu in 2003 (I have added the link to the bottom of this message) and they created a SAS macro for that purpose. Is there a way to perform this in Stata (I have Stata 15.1)?

Kind regards, Horea

[Leonardo Guizzetti]

I believe Roger Newson's -somersd- package handles clustered observations which could be applicable, but the method used it not the same as the Liu and Wu article. For that specific method, you may need to implement it yourself. There is a non-parametric version that is also very good by Obuchowski, NA. (1997) Nonparametric analysis of clustered ROC curve data. Biometrics. 1997: 567-578. Her method handles estimation of AUC and differences in AUC for correlated observations.

[Horea Feier]

Thank you for your reply Leonardo,

I am unfortunately not that proficient in programming as to be able to implement the Liu& Wu program in Stata.

This is a situation that is very often encountered in medical studies: the repeated measurements of several variables over time, and to discern which ones are most predictive of the desired outcome. I am quite surprised that it hasn't been programmed in Stata yet.

[Leonardo Guizzetti]

Thank you for your reply Leonardo,

I am unfortunately not that proficient in programming as to be able to implement the Liu& Wu program in Stata.

This is a situation that is very often encountered in medical studies: the repeated measurements of several variables over time, and to discern which ones are most predictive of the desired outcome. I am quite surprised that it hasn't been programmed in Stata yet.

I agree with you, Horea. It is my experience that clinical research does not often make use of repeated-measures designs which call for ROC curve analysis. As for what ends up in statistical software packages, it is a mix of well-established methods, those that are highly requested, or necessity when either of the first two options are found to be lacking. Also, academic statisticians are not always interested in translating their methodology into user-friendly code for obvious reasons.

If it is important to implement this in Stata, you may wish to take up the challenge of translating from SAS to Stata which will make you a better Stata programmer in the end.