AUC below 0.5?

[Torbjorn Skodvin]

Hi,

I am calculating area under the curve (AUC) values for the individual independent variables of a conditional logistic regression model. I am using the rocreg command.

Half of the ten variables show AUC values below 0.5, specifically values as 0.32, 0.38 and 0.45.

Do AUC values below 0.5 indicate data error? If not, can the values simply ble inverted (so that I present AUC of 0.32 rather as 0.68), or must they be reported as they are?

The study is a matched case control study with 12 cases and two matched controls for each case.

Thank you in advance.

[Joseph Coveney]

I'm curious—how do you do conditional logistic regression with rocreg? Are you fitting models with clogit first and using one or another of its predictions, or something?

I'm not familiar with the application of the area under a receiver operating characteristic curve as an index of discrimination in the context of a fixed effects logistic model, and so I can't really answer your questions, sorry. But my interpretation of a value less than 0.5 (assuming that it's really different from 0.5) is that when the forecast says "sunny and warm", you had better bring an umbrella and wear an overcoat.

[Steve Samuels]

I like Joseph's analogy: AUC<.5 can occur if higher values of a variable are associated with lower probabilities of event.

Note that conditional logistic regression and prediction don't go well together. See post 5 in this thread.

[Torbjorn Skodvin]

Thank you for your replies.

After starting the thread, I have understood elsewhere what you say, Steve, about clogit and prediction. We have decided to remove AUC from the table it was intended to be in. Still, thank you!