Constraints on attributes levels with an orthogonal design

[romlbc]

Hi all,

I would like to have an orthogonal design with some constraint on the attributes such as:
alt1.A > alt2.B [1]

However I noticed that the 'reject' and 'require' functions do not work when using an orthogonal design. Does someone know how to handle this?
I could remove the choice profiles that do not satisfy but I might loose the orthogonality of the design...

Many thanks in advance for your response,

Best,

Romain

[Andrew Collins]

Hi Romain

The problem is that the contraints will almost certainly break the orthogonality.

A compromise solution might be to generate an efficient design with some priors, use the reject, require or cond properties, then of the best of the solutions, choose the design which has the lowest correlations (which can be found in the Design window, under Design in the tree on the left). Setting ;store=100, or some large number, will give you access to, say, the 100 most efficient designs.

It would be nice if a value such as the sum of the correlations, or the maximum correlation value, could enter the ;eff property, and be optimised on. This is something we may add in a future point release.

Andrew

[Michiel Bliemer]

Is there an important reason to hold on to orthogonality?
I agree with Andrew that it is almost impossible to find orthogonal designs in the presence of constraints, hence Ngene cannot do that.

However, it is possible to find near-orthogonal designs with Ngene. If you have an unlabelled experiment with generic parameters, you can generate a D-efficient design using zero priors, and add your constraints. A D-efficient design with zero priors will be close to a sequential orthogonal design in this case. If you have a labelled experiment with all alternative-specific parameters, the D-efficient design generated with zero priors will be close to a simultaneous orthogonal design. With D-efficient designs you have a great flexibility in adding constraints, so I hope this will be a solution to you. If you require strict orthogonality (but I do not see why, really), then maybe manual construction may be possible, but the design just may not exist that you are looking for.

Michiel