Area under the ROC Curve for genetic data

Question

[HumeMarx]

Hi Biostar.

I have performed logistic regression on 3 independent sets of case/control cohorts. I now have the task of figuring out how well a collection of the SNPs implicated in the study predict disease/control status of the individual.

The package I am trying to use in R is ROCR however I cannot figure out how I should generate the prediction object and other files needed to start the process. The code looks very simple and straight forward but I am struggling with getting started.

Any idea how I can generate the prediction object and other files for several hundred SNPs? I have the statistical values generated from the logistic regression analysis for all the SNPs as well as the affliction status of the 3 cohorts.

I appreciate any help you can provide.

[HumeMarx]

Surely someone must have an answer to this...

Answer 1

[Shicheng Guo]

ROCR is the last step you should applied. But first you need to build the prediction model. In this step, the most important you need do is feature selection in which you will select all significant SNPs to be used to predict for the sample status. In your situation, you only have hundreds SNPs, I think you can first do the association study one by one for each SNPs/ or lasso based method and then collect all significant SNPs. Then build a multi-variate logistic regression model. In this step, you need split sample to training set and test test, using training set to built the prediction model and use test set to test the sensitivity and specificity and then you can get the ROC curve. Thousand R package can do this job. such as https://cran.r-project.org/web/views/MachineLearning.html

[HumeMarx]

Thank you, this is most useful. Thanks