Due to its own nature mixture design models (mainly with interactions) presents multicollinearity problems, some authors declare that a combination of forward stepwise and step backward stepwise regression until they, converge in a unique model is enough.
Some authors go farther and declare that it is needed to employ regression techniques that do not assume mathematical independence of the factors, for example partial least square regression (PLS regression).
Are both theory right or is needed to employ PLS regression? Or exist other(s) way(s) to eliminate this problem?
Of those two options, I'd strongly advocate for PLS to increase the likelihood that your results will generalize.
Without knowing a lot more about your domain, I'd be hesitant to consider alternatives that are theory-based.
Article Multicollinearity and Leverage in Mixture Experiments
Article How Collinearity Affects Mixture Regression Results
I recommend this paper by a real expert at this found on Scholar.Googe:
Multicollinearity and leverage in mixture experiments
DC Montgomery, SR Voth - Journal of Quality Technology, 1994 - Taylor & FrancisThe impact of leverage, influential observations, and multicollinearity in mixture experiments is discussed. By example, it is shown that consideration of these factors is important in the practical selection of a mixture design. Both standard simplex-type mixture designs and computer-generated designs for constrained hyperpolyhedron regions can have very non- uniform distributions of leverage, and consequently, some design points may exert disproportionate influence on the least squares fit. Generally, multicollinearity is not a … Cited by 40 Related articles All 2 versionsShowing the best result for this search. See all results
Best, D. Booth
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression