Design of Experiment - Test for Curvature

Hi,

I have performed a DOE with 2 factors, each at 2 levels and one center point. When I analyzed the results the p-value for Center point is 0.611, which indicates that the curvature is not significant at 5%. But when I made the 'Main effect' plot with center point, I can visually see that the center point is 'quite far' from the linear graph. My understanding is that if the center point is not significant then it should lie close to the linear graph.

Do you have any good explanation to explain why the 'red dot' is far from linear line and still not-significant.

Thanks

None of the factors are statistically significant. Therefore, the factorial plots are entirely meaningless, and you should not try to interpret them. If this lack of significance is surprising due to prior process knowledge, you may have had an underpowered design. I would recommend adding some replicates and additional center points. You could add these to the existing experiment by blocking on the different days you ran these experiments.

There are many reasons why a DoE will result in insignificant results for all of he factors. One is that you have too few replicates for the amount of variation (as Miner indicated) injected into your study. Another is that you didn’t inject enough variation in the controlled factors (ie. You didn't set the factors hi and lo levels far enough apart). Arguably teh most common reason is that the factors that you chose are NOT the drivers o f the variation You see or want to control.

In teh end it is never the statistical calculation that matters bu the study design itself. Why are you running the D0E? What output characteristic are you trying to control or reduce? how did you select the factors? How did you select the levels of teh factors? How did you ensure that the underlying process variation of all of the other factors were included in the study?

Bev D said:

Another is that you didn’t inject enough variation in the controlled factors (ie. You didn't set the factors hi and lo levels far enough apart).

Click to expand...

A common mistake is to set the hi and lo values at the spec limits for these process parameters. The problem is that these spec limits were chosen specifically to prevent variation in the product characteristic. Therefore, to see changes in a product characteristic, you must select hi/lo values for process parameters beyond the spec limits.

Miner said:

None of the factors are statistically significant. Therefore, the factorial plots are entirely meaningless, and you should not try to interpret them. If this lack of significance is surprising due to prior process knowledge, you may have had an underpowered design. I would recommend adding some replicates and additional center points. You could add these to the existing experiment by blocking on the different days you ran these experiments.

Click to expand...

Thanks for the reply. I have already added two center points in the study so I think it is not an underpowered design. This is a new process design so we do not have any prior process knowledge about so I can assume it is the factors which are not significant. I have also two other 'response variables' and there I can see some of the factors being significant so I guess it is only this response variable which cannot be predicted by same control factors.