I generated an I-Optimal DoE with only numeric factors (4 of them) with responses. After conducting my experiments, I was getting some strange ANOVA R squared values for some responses that weren't lining up with the correlation that the actual results were showing me. Now one of these numeric factors was Raster angle (as my research is about 3D printing). Two of the four levels of the raster angle are alternating, which means that it's -45/45 and 0/90. (the other numeric levels being 0 and 45), thereby making the levels: -45/45, 0, 45, 0/90

Originally when I created the DoE, I simplified the comprehension of the framework and used -45, 0, 45, 90. However, while it worked for some responses and I got very favourable R squared values (for surface roughness etc.), the tensile test was not producing any values at all! (i.e. R squared values were literally zero). I was very confused, and intuition led me to changing the Raster Angle factor to Nominal Categoric in the DoE table.

By just doing that and re-analysing my responses, my Tensile Strength R squared values went from 0 to 0.86, 0.82, 0.76. After conducting backward elimination technique, most of the other responses I was targeting produced close to or better R squared values than before including their predicted R squared values. What I realised however is the predicted R squared values for some of them (e.g. Surface Roughness) changed from 0.642 (when Raster Angle was still numeric) to -0.007, until I reduced the model using Backward elimination, which led it to becoming 0.8357.

So the question I re-iterate: is this allowed regarding the legitimacy of my data after creating the DoE and already having my responses? The meaning behind what the values are on the surface level remained the same (as well as at the experimental level). The difference is the type of factor Design Expert perceives it to be now I'm assuming?