Hello,
I have come up with a DOE, ran experiments and I want to predict within the design space.
But after analysing my samples and putting it into JMP I see that I do not have normal distribution of my data. My question is, is it still possible to continue with my model building if I do not have the normaly distributed data even though the r2 value of the actual vs predicted is really high?
Personally, I think that it is still a viable model due to r2 value. Also, as this is a follow up DOE, meaning that I have data that covers broader space of the design and previous data set had normaly distributed data.
With the first DOE I wanted to know what are the preferred conditions, whereas with the 2nd DOE (attached file) I wanted to further investigate the design space and generate as much desired conditions as possible, therefore the data seems to be skewed.
Also, I wanted to ask, is it possible to join 1st and 2nd DOE (different designs), 1st DOE is fractional factorial and the 2nd DOE is RSM
In general I wanted to ask professionals what do they think and what could I do to improve my model further
Thank you in advance
When you say you concerned about non normal data are referring specifically to the distribution of raw response data? If that's the case, you don't need to worry, at least initially. Think about a designed experiment...you are almost always TRYING to elicit a response that is a result of intentional changes to factors, chosen at specific combinations, at specific levels in a way that shows how the responses behave to these intentionally run combinations.
Frankly, I'd be suspect if my raw response data IS normally distributed...in other words...maybe I didn't get a signal to rise above the noise. Normality of other DOE results like residuals etc. are a different story...or if you are unwilling to assume the normality of errors assumptions that make the undelying mathematics of ordinary least squares regression viable...again a different story. But for you raw response data...proceed.
