How Do You Compute the Most Successful Recipe Using Multivariate Testing?

[Chad0123]

Homework Statement

My goal is to have enough info that I could sit down with a piece of paper and actually compute a pretend test from start to end.

All the examples and docs I have found online are not nearly in layman's terms.. oh, please excuse my math-lameness btw. :-)

If I have 5-7 factors with a variable number of possibilities per factor (finite at the start of the test), how do you actually compute which recipe will be the most successful in generating an outcome?

What does one go through to determine the best recipe?

ex:

1) is it possible to use multivariate testing if you have a variable number of possibilities per factor? (could be 1-1200?)
2) how do you pick the recipe to test?
3) how do you know if you've tested enough (believe this is standard deviation?)
4) how do you consolidate the test results to evaluate recipes you haven't tested?
5) how do you put it all together?
*bonus question) Which type of equation is best suited for this?

A simple explanation would be super awesome...

Thanks a lot everyone.

Homework Equations

Multivariate
Taguchi
Discrete choice
etc

The Attempt at a Solution

1,1,1,1,1 = 32 yeses
1,2,1,1,1 = 1 yes
1,1,3,1,1 = 5 yeses

so...

*shrug* I have no idea :-l

 

[D H]

1) is it possible to use multivariate testing if you have a variable number of possibilities per factor? (could be 1-1200?)
2) how do you pick the recipe to test?
3) how do you know if you've tested enough (believe this is standard deviation?)
4) how do you consolidate the test results to evaluate recipes you haven't tested?
5) how do you put it all together?
*bonus question) Which type of equation is best suited for this?

Chad,
1) It sounds like you are wanting to use multivariate analysis of variance (MANOVA) techniques. (Not quite sure; the description of the problem was a bit vague.) I suggest going to a library and finding a book on the subject.

2) I'll assume you have a moderately hard problem: You have a number of candidate independent variables you are trying to use to explain a non-linear process. That can be not just moderately hard, but downright vexing. The problem in a nutshell: Some of the candidate independent variables are highly correlated with one another, some are completely irrelevant to explaining the data at hand, and the process is non-linear. I've dealt with these kinds of problems in two ways. One way is to start with a null model (the dependent variable(s) are pure noise). Determine which is the most statistically significant independent variable and add it into the model. Keep doing this until the most statistically significant variable that is not in your model fails to meet some test of significance. The reverse approach is to start with a full model and remove parameters one at a time until you find you have to remove a parameter that is statistically significant.

3) One key metric is the residual error. Any good book on multivariate analysis will discuss residuals at length. There are lots of other tests such as Student's t-test (a cornerstone of univariate ANOVA), chi square tests, ... Once again, go to a library.

4) This almost sounds like you are moving into the realm of machine learning. Learning how to do that properly requires taking multiple classes in college, several of which are at the graduate level. One last time, I strongly suggest you go to a library.