choice-metrics.com

Dear members,

Now I am trying to find a better D-efficient design with 0 prior because of the pilot study.

However, I can't check the OOD button in the properties instead it shows an error message "not enough memory to build OOD covariance matrix" when I run the following code and then put into the OOD button.

(CODE)
Design
;alts = alt1*, alt2*, alt3* ?unlabelled experiment (alt4 is opt-out)
;rows = 24
;eff=(mnl, d)
;block = 4
;model:
U(alt1)
=ta.dummy[0.00001]*ta[1,0]
+ki.dummy[0.00001]*ki[1,0]
+ei.dummy[0.00001|0.00002]*ei[2,1,0]
+oy.dummy[0.00001]*oy[1,0]
+sh.dummy[0.00001|0.00002]*sh[2,1,0]
+es.dummy[0.00001|0.00002]*es[2,1,0]
+ho.dummy[0.00001]*ho[1,0]
+pl.dummy[0.00001|0.00002]*pl[2,1,0]
+pr[-0.00001]*pr[9.9,19.8,29.8,39.8]/
U(alt2)
=ta.dummy*ta
+ki.dummy*ki
+ei.dummy*ei
+oy.dummy*oy
+sh.dummy*sh
+es.dummy*es
+ho.dummy*ho
+pl.dummy*pl
+pr*pr/
U(alt3)
=ta.dummy*ta
+ki.dummy*ki
+ei.dummy*ei
+oy.dummy*oy
+sh.dummy*sh
+es.dummy*es
+ho.dummy*ho
+pl.dummy*pl
+pr*pr
$

I can push this button when I run the code which decreases a level of an attribute like es.dummy[0.00001]*es[1,0] in the code above, for example.
I do not employ OOD though I am wondering the reason.
I was wonder if you could comment in my questions.

My setting:
I put Ngene in C drive.
I have installed new version 1.2.1.

Thank you for your time in advance.

Best regards,
Keiko Aoki

I have forwarded the question to the ChoiceMetrics technical support.

In short, it has to do with the the necessary computations described in the papers of Street and Burgess, which are quite involved. I would agree that memory should not be an issue here, but I do know that the matrices that need to be constructed in order to compute the D-efficiency value are huge. Once I receive an answer I will post it here.

Note that OOD and D-efficiency percentages are really only relevant for optimal orthogonal designs.

Michiel

Dear Professor Bliemer,

Thank you for your reply and forwarding my question to the technical support.

As further findings about this, I could check the OOD button and it showed "undefined" in the D-optimality, not figures.

By the way, I have understood that 0 priors in the D-efficient design is almost similar to the orthogonal one.
But, D-error was not always proportional to D-optimality when I employ 0 prior by the generating several codes.

As you explained before, I have understood that D-optimality is for orthogonal design, not D efficient design.
However, I would like to find the reason why I found a few phenomenon above.
So, I look forward the reply from the support.

Thank you for your time.

Best regards,
Keiko Aoki

A D-optimal design in the terminology of Street and Burgess refers to an orthogonal design for an unlabelled experiment with minimal D-error assuming equal choice probabilities (i.e., zero priors).

A D-efficient design refers to a design with low D-error, it does not constrain the design to be orthogonal or have zero priors and can be applied to labelled and unlabelled experiments.

A D-optimal design as defined by Street and Burgess is therefore a very special case of a D-efficient design. If you use zero priors in an unlabelled experiment and generate a D-efficient design and constrain the design to be orthogonal, then it will essentially create a D-optimal design (if it exists). If you do not impose the orthogonality constraint, then it will find a design that may be similar to a D-optimal design but it will not be the same because orthogonality is not imposed.

Michiel

Dear Professor Bliemer,

Thank you for the quick response.

I have misunderstood about zero priors in a D efficient design.
Although I use zero priors in an unlabelled experiment and generate a D-efficient design, I don’t constrain the design to be orthogonal.
So, my D efficient design with zero priors is not the D optimal design (orthogonal design).

I appreciate you so much again.

Thank you for your time.

Best regards,
Keiko Aoki