My experimental design includes two conditions, in which the trial number in one condition is more than that in the other, resulting in data unbalanced. We want to perform ‘cross-classification’ through setting 'cfg.design.unbalanced_data = ‘ok’ ',although the TDT analysis was completed smoothly, the dprime image of the analysis result was null, so that further analysis could not be made.I want to ask why does unbalanced data cause TDT to calculate the dprime as null? How to deal with unbalanced data in TDT?
Hi Xinling,
I would recommend running AUC_minus_chance instead. It’s more sensitive, because it uses the decision values rather than the classifier’s choices. In your analyses, you have unbalanced data, so the classifier will likely always prefer one option. In that case, you will end up with a dprime of 0. AUC_minus_chance gets you around this, because it is not dependent on the classifier’s bias. For strong imbalances, you might still run into issues. I would in general recommend crossnobis (which is implemented in our toolbox) which can easily deal with unbalanced data since it is only based on the multivariate means. This might be what you are looking for. So, if you would like to stick to what you know, try AUC_minus_chance, else go for crossnobis. We have a template for this, too.
Best,
Martin
Thank you so much for your answer. Thanks to your answer, I solved my problem.