I'm presently working on the methodology of a psychological experiment, and I'm having trouble to decide of the best way to choose the stimuli.

I got an image database in which different models are doing two different facial expressions (neutral and angry), and each of these images are rated for valence, intensity, genuineness etc...

What I'd like to do is to select every model for which the valence of the angry expression is significantly lower than the valence of his neutral expression.

Here's an exemple of the data I own :

Model Neu. An.
2,00 3,04 1,92
3,00 3,30 1,68
4,00 3,17 1,65
5,00 3,05 2,09
7,00 3,50 2,00
8,00 2,96 2,04
9,00 2,87 2,22
10,00 2,95 2,26
So, I got two different distributions of valence scores, and what I want is to be able to say for which model the valence of the Angry expression is really lower than the valence of the Neutral expression. I think that means that the difference between Neu. and An. expressions for a model should be greater than the average difference for the same expression between models, but I'm not certain I'm understanding this right.

My teacher told me to use a paired t-test with models as subjects, but I don't really understand this because the t-test only tells me if the two distributions (Neu. and An.) are different or equals, and not what's the minimal difference between the two scores of one model to be significant.

Do you think my first intuition is good ? In that case, I don't know if a t-test really is necessary.

If not, which output value from the t-test should I consider to be the significance threshold ?

I'm using Excel 2010, and tried different type of analysis but none was really satisfying...

If you have another idea which could lead me to the result I want I'd be very interested too.

Thanks by advance !