choice-metrics.com

I am happy to write the first post on this forum.

I have a problem with generating design which would have the same status quo alternative in each choice set and use orthogonal design at the same time.

The design I would like to achieve will be clear from the syntax. I tried:

Code: Select all

Design
; alts = sq, law, feed
; rows = 12
; orth=ood ? I also tried seq here - same result
; model:
U(sq)   =      hh * H1[0]      + ge * G1[0]    + co * C1[0]      /
U(law)  = b2 + hh * H[0,1,2,3] + ge * G[0,1,2] + co * C[0,1,2,3] /
U(feed) = b3 + hh * H[0,1,2,3] + ge * G[0,1,2] + co * C[0,1,2,3]
$


and:

Code: Select all

Design
; alts = sq, law, feed
; rows = 12
; orth=ood ? I also tried seq here - same result
; model:
U(sq)   =      hh * H.ref[0]       + ge * G.ref[0]     + co * C.ref[0]       /
U(law)  = b2 + hh * H.piv[0,1,2,3] + ge * G.piv[0,1,2] + co * C.piv[0,1,2,3] /
U(feed) = b3 + hh * H.piv[0,1,2,3] + ge * G.piv[0,1,2] + co * C.piv[0,1,2,3]
$


They both generate the same error:

ERROR: Something went unexpectedly wrong. You may wish to contact Ngene's authors with a copy of the error.log file.
Aborted.

Interestingly, they both seem to work fine with e.g.

Code: Select all

; eff=(mnl,d)

instead of

Code: Select all

; orth=ood

.

My two questions are as follows:
1. Are the two forms of specification actually the same? I am confused, since Ngene would not accept specifying different number of attribute levels for different alternatives. I am not sure if providing them as a different attribute for different alternatives (but with the same parameter) is ok.
2. Would it matter (for orthogonality / efficiency) if in my case I simply left the status quo alternative out and added it to my generated design manually later?
3. Am I doing some obvious mistake here or reference scenarios would not work with orthogonal designs at all?

Thank you in advance for all your comments / suggestions.

Congrats on the first post!

1. The two specifications you have provided are equivalent. For sequential orthogonal designs specified with ;orth=seq, attributes must be of the same dimension across alternatives, as under this technique, only the first alternative is generated, and subsequent alternatives have levels assigned using modulo arithmetic. ;orth=seq2 also generates a sequential design, but finds a unique orthogonal table for each alternative. However, since the status quo alternative only has one level, no orthogonal or ood designs can be found (for ood designs, refer to Street et al, 2005, as listed in the references section of the manual. Sq is undefined for a single attribute level)

2. For orthogonal designs, you could remove the status quo alternative and just search for an orthogonal or ood design among the remaining alternative. For efficient designs, the choice probabilities are important, and so the status quo alternative (and any associated priors) will have an impact on the efficiency of the design, and should be specified in the utility expressions.

3. Yes, reference alternatives will not work with orthogonal designs, as the reference alternative will never vary, and the remaining alternatives can not be orthogonal with that alternative.

Regards,
Andrew