@bdclark23 : Multiple model approaches with a preponderance of evidence that consistently show the same effects across different models, each with their own biases and assumptions.
Question 2: “Covariate Approach” of applying constraints
Covariates.mp4
Video Player is loading.
Current Time 0:00
/
Duration 13:49
Loaded: 0%
00:00
Stream Type LIVE
Remaining Time -13:49
1x
Chapters
descriptions off, selected
captions settings, opens captions settings dialog
captions off, selected
en (Main), selected
This is a modal window.
Beginning of dialog window. Escape will cancel and close the window.
End of dialog window.
This is a modal window. This modal can be closed by pressing the Escape key or activating the close button.
@Uwe_Weber : Use the Profiler and they will understand
@anchildress : We have the most success at Eastman when we focus on relationship building with customers/clients, starting small to build trust with you and your techniques. Once they see some wins or successful DOEs with colleagues perhaps they are more trusting of DOE.
Results and cost savings are the most convincing technique
@DB1 : Changing a culture: Patience. Work around the edges. When you can convince one scientist or engineer, they can be more convincing of DOE use to their teammates than us. As Bea says, one really good DOE dataset in a company can convince a lot of people if used diplomatically and audience-aimed.
Some comments to elaborate on points discussed yesterday :
Stepwise selection
Stepwise regression can be problematic in designed experiments, as it :
ignores the experimental structure of the DoE : hierarchy principle, sparsity principle, and sometimes heredity principle, but this last principle can be forced in Generalized Regression platform,
creates biased/inflated parameters estimates, as these terms are primarily selected because of their magnitude/effect size,
increases the risk of false positives (probability of detecting an active effect that is not active) due to multiple testing without adjustment,
provides unreliable p-values, as they can be both used for terms selection and estimates statistical testing,
doesn't adress multicollinearity, which can be a problem with strong aliases/correlations between terms in the design, and/or in the case of Mixture designs,
completely ignores the assumed model related to the generated design.
There may be other reasons, some litterature and discussions on this topic can be found easily :
Stepwise selection can still be helpful as a benchmark model, and if the number of factors is not too high, the The All Possible Models Option can be interesting to see how different models agree and disagree about the terms selection and estimation. But the model comparison and selection should be guided by the assumed model behind the design, and by domain expertise.
Some discussions and posts in the Community related to Stepwise approaches on DoE data :
The choice of adding centre points is often more related to the objectives and practical reasons : detecting curvature, ensuring measurement stability process during experiments, having an estimate of pure error/variability, having an estimate in the centre of the design space (in the context of robustness analysis with regards to factors space/Quality by Design), ... Depending on the assumed model and the type of design, centre points may not be required ; classical ones like Central Composite Designs and Definitive Screening Designs have centre points, whereas computed optimal designs may not require them.
@maria_astals As a suggestion for next meetings, it would be great if some of the topics could be voted and prepared in advance : as the DoE club is a quaterly meeting, we would have enough time to create a poll with different topics, and vote for them before the next club meeting. This would ensure more in-depth discussions as well as more variety in the topics (and perhaps more participation/engagement). A mix between prepared/voted topics and spontaneous ones/questions could be a good idea for future meetings ?
It was great to have you in the meeting Bill @statman. Really enjoy learning from you.
Victor GUILLER
"It is not unusual for a well-designed experiment to analyze itself" (Box, Hunter and Hunter)
