Hi Frédéric,

I have been thinking about your problem, and I have one question and a few suggestions.

My question is: when you talk about "deception detection scores", do you mean that you are going to, somehow, extract a numerical success score? If so, it would mean that you only need to correlate numerical values (if you are working with Likert Scales as numerical data). Spearman or Pearson would fit here.

If not, and you are just comparing a categorical and a numerical variable, you could use t-tests (or, to be more precise, ANOVA, as your could group the people depending or their detection scores), or even their non-parametric version: Kruskal Wallis.

Finally, if you are assuming that Likert scores are categorical (it depends on the theoretical background you are relying), then you would have a comparison categorical vs categorical. In this case, we are currently using Chi Square/Fisher's F/McNemar.

If I have misunderstand anything, I would be happy to change my suggestions after you could provide new information.

Hope this helps!

Hi Javier,

Thanks for your answer, as it provides interesting insights.

This is how I computed the answers: two detection tasks were done before training, and two after training. Correct detection was classified as +1, incorrect as -1 and the "don't know" as 0. This gives 5 final scores ranging from +2 (both correct answers) to -2 (both incorrect answers). I believe that these could be considered as categories (terrible detector, bad detector, not a detector, good detector, excellent detector), and in that case, we could use an ANOVA.

What do you think?

Thanks again, and hope you are doing amazing in Madrid!

Fred

First, be very careful about adding categories. For example, with that logic, the person that guess the first try, but fails the second, shall be considered as "not a detector"... which is technically the same than a person which has not answered at all! That is the problem when approaching categorical values. I dont have a clear answer to you, but I can suggest you a couple of solutions.

The first could be to classify in (Both correct, both incorrect, one correct and one incorrect, one correct and one not answered, one incorrect and one not answered, both answers avoided).

The second is to go with the option you said, but maybe erasing the option "0" and calculating three options: one success, two successes, and no success at all. We had a similar problem than you in the dissertation of a friend, and we decided to use this second option.

About the ANOVA, if you are planning to combine the categories with the Likert (assumming the Likert as a numerical variable), it would be nice. But remember to analyze the normality.

Hope you are doing well on your PhD!!!