Hi,
I'm very new to SAS and don't have a basis in stats so struggling to get my head around a lot of the language in the various documentation, so I apologise if this is obvious or I'm not giving enough information. I'll describe the experiment I've done just so I'm not making incorrect assumptions.
I have 100 images which have been watermarked using 1 or 6 different methods, at 1 of 5 different levels (0.1,0.2,0.3,0.4,0.5bpp), so a total of 100*6*5 images. Each of these has then been analysed at 4 different thresholds to produce a value(%) indicating the similarity of the watermarked image to the original. So my dataset has 12000 observations and 5 variables: image(cat) watermarker(cat) bpp(interval) threshold(ordinal) PCM_value(interval)
I have found that the PCM_values are non-normally distributed for a given watermarker/bpp/threshold (actually in some cases, normal, in other cases very non-normal). As the same set image of images are watermarked for each set these are correlated. This correlation and non-normality is thwarting my attempts at understanding from other forum posts and blogs!
I was trying to perform regression using GENMOD with bpp as the IV, PCM as the DV, classes image watermarker and threshold, with repeated on image to account for the fact they are correlated, but I'm not convinced this is valid given non-normality.
Ultimately I'm looking to get values for regression coefficients that allow me to compare between watermarkers and thresholds i.e. does the choice of watermarker have a significant effect on the PCM values, similarly for threshold, as well as the general effect of bpp on PCM. But I'm really not sure how to go about this, whether I'm correct in using GENMOD, or should be using MIXED or if I'm way off base.
Will be grateful for any help anyone can give!