Hi all,
I am researching on consumers’ WTP for woods sourced from different sources. I wanted to generate an efficient design for three unlabelled alternatives with an opt-out. However, the design could NOT be generated. I appreciate if anyone gives me a hand.
Here is the code
Design
;alts = alt1*,alt2*,alt3*,alt4
;rows = 36
;block = 6
;eff = (mnl,d,median)
;rdraws=halton(5000)
;bdraws=halton(5000)
;model:
U(alt1) = b1.effects[(n,0.1,0.05)| (n,0.09,.05)|0]*SFWA[3,2,1,0] + b2.effects[(n,0.1,0.03)|0]*LOGOA[2,1,0] +b3[(n,-0.08,0.01)]*DFWA[50,100,200, 300] + b4[(n,-0.02,0.01)]*COSTA[220,240,260, 280]/
U(alt2) =b1*SFWA+b2*LOGOA+b3*DFWA+b4*COSTA/
U(alt3) = b1*SFWA+b2*LOGOA+b3*DFWA+b4*COSTA/
U(alt4) = b5(0.1)$
Thanks in advance
Efficient design
Moderators: Andrew Collins, Michiel Bliemer, johnr
-
Michiel Bliemer
- Posts: 2055
- Joined: Tue Mar 31, 2009 4:13 pm
Re: Efficient design
There are several issues:
1) Dominant alternatives appear a lot in your model because you have 3 alternatives each with only a small number of attributes and a low number of levels. Ngene finds is very difficult to find 36 choice tasks without any dominant alternatives. For such a heavily constrained design (dominance checks are constraints) you can use the modified Federov algorithm, i.e. add something like ;alg = mfederov(candidates = 1000). Note that this algorithm will not create attribute level balanced designs, so you may wish to impose some attribute level balance constraints to your syntax.
2) The prior in alt4 is not properly defined, it needs to be b5[0.1] using square brackets.
3) Your priors are problematic, please use appropriate priors, preferably from a pilot study, or use very small positive or negative priors to only indicate the sign. Ngene will not be able to find any designs with a finite D-error because of these priors since alt4 is dominant (the other alternatives will have a large negative utility and will never be chosen according to these priors). Your priors for DFWA and COSTA are too large, they should be 0.008 or 0.002 or smaller I guess.
Please note that this forum is for general questions regarding stated choice experiments. For Ngene specific questions like these, please use the Ngene forum.
Michiel
1) Dominant alternatives appear a lot in your model because you have 3 alternatives each with only a small number of attributes and a low number of levels. Ngene finds is very difficult to find 36 choice tasks without any dominant alternatives. For such a heavily constrained design (dominance checks are constraints) you can use the modified Federov algorithm, i.e. add something like ;alg = mfederov(candidates = 1000). Note that this algorithm will not create attribute level balanced designs, so you may wish to impose some attribute level balance constraints to your syntax.
2) The prior in alt4 is not properly defined, it needs to be b5[0.1] using square brackets.
3) Your priors are problematic, please use appropriate priors, preferably from a pilot study, or use very small positive or negative priors to only indicate the sign. Ngene will not be able to find any designs with a finite D-error because of these priors since alt4 is dominant (the other alternatives will have a large negative utility and will never be chosen according to these priors). Your priors for DFWA and COSTA are too large, they should be 0.008 or 0.002 or smaller I guess.
Please note that this forum is for general questions regarding stated choice experiments. For Ngene specific questions like these, please use the Ngene forum.
Michiel