A multivariable ROC AS SUCH is a nonsense, given it is related to the change in sensitivity and specificity (the axes of the ROC plot, to be more precise they are 1-specificity (FDR) and sensitivity but the point is the same) driven by the change in a single parameter (threshold).

So said, you can separate the multivariable (the logistic equation is an excellent choice, or you can use a canonical correlation as well) so to single out a single discrimnant variate (logidtic or canonical variate) to be used for a ROC analysis that now becomes de facto multivariable in meaning given the 'test variable' comes froma multivariable analysis and collates all the relevant information for discrimination.

Hi Cristina. When you say multivariate, do you mean a model with 2 or more outcome variables, or a model with one outcome variable and 2 or more explanatory variables? Thanks for clarifying.

I ask for the clarification above because there is a lot of variation in how people use this terminology. For statisticians, univariate vs multivariate describes the number of outcome variables in a model; and univariable vs multivariable describes the number of explanatory variables. (See the article linked below, for example.) But many users of statistics use multivariate to describe the number of explanatory variables, and so will say they have performed multivariate regression when they really have a multivariable univariate model (i.e., multiple explanatory variables, but only one outcome variable).

Article Multivariate or Multivariable Regression?

In this regard, the Graph Pad Prism 8 has progressed a lot. You can use it.

R package HUM might be a possibility; see https://cran.r-project.org/web/packages/HUM/index.html

Best Matthias

Thank you all for your answers.

Abolfazl Ghoodjani I will try with Graph Pad Prism 8. Thank you for your suggestion.

Matthias Kohl unfortunately I have no experience with R. I know it is a very useful tool in research, and I have tried to learn to use R by myself, but it is hard. R is one of my purposes for the future.

Bruce Weaver I think your answer is really illuminating. I looked at the paper you suggested and I think that I have used the wrong words. According to this paper, I think I am working with a multivariable model. I have two group of patients, one affected and one unaffected, and a set of biomarkers which I would like to use to diagnose the disease. Which is the best procedure for this kind of model? Thank you very much for your help.

Good question. I want to add to the question. Is statistical software package like SPSS not possible to run it?. I am reading about Graph Pad Prism 8 for the first time?.

Given that you have a multivariable model, not a multivariate model, I think you can get what you want very simply by estimating a logistic regression model (with condition, Yes/No as the outcome variable) and saving the predicted probabilities. Then use the saved predicted probability as the "test" variable when creating your ROC curve. From a quick glance, the file linked below should guide you. HTH.

https://case.truman.edu/files/2015/06/SPSS-Logistic-Regression.pdf

Thank you very much Bruce Weaver! I performed a binomial logistic regression analysis as you suggested and obtained the multivariable ROC curve.

I really appreciate your help. Thanks a lot again!

No worries, Cristina. Glad to hear it worked out.

Cristina Barbagallo , I guess that Alessandro Giuliani , can help you.

Alessandro Giuliani is there any way to help Cristina Barbagallo ?

A multivariable ROC AS SUCH is a nonsense, given it is related to the change in sensitivity and specificity (the axes of the ROC plot, to be more precise they are 1-specificity (FDR) and sensitivity but the point is the same) driven by the change in a single parameter (threshold).

So said, you can separate the multivariable (the logistic equation is an excellent choice, or you can use a canonical correlation as well) so to single out a single discrimnant variate (logidtic or canonical variate) to be used for a ROC analysis that now becomes de facto multivariable in meaning given the 'test variable' comes froma multivariable analysis and collates all the relevant information for discrimination.

Thank you for your help Chiara Bianca Maria Platania and Alessandro Giuliani !

È stato un piacere