choice-metrics.com

Dear all,
I have a question concerning abundance of levels in the design.

We are interested in the reasons for farmers to replace there current artificial fertilizer with a bio-based fertilizer. We have an experiment with 3 alternatives (2 alts+ status quo). Because we don’t have the ability to check for priors in various countries and because the hypothesis is that priors would be different in various countries, it was decided to stick with zero-priors. In order to specify the possibilities for the status-quo correctly and to avoid unrealistic combinations of levels, I added some reject conditions.

Because orthogonality does not work with the status-quo and conditions (reject as used in my design), we used an efficient design (;alg=mfederov(candidates=100)). From the NGene manual (NGENE 1.1.1 p 103) I understand that from the candidate a balanced design is produced?

Code: Select all

Design
;alts=alt1,alt2, alt3
;rows=12
;eff = (mnl, d)
;alg=mfederov(candidates=100)
;block=2
;require:
alt3.A=0, alt3.B=0, alt3.C=0, alt3.D=0, alt3.E=0, alt3.F=0, alt3.G=0
;reject:
alt1.E>alt1.G,
alt2.E>alt2.G,
alt1.B=1 and alt1.A=0 and alt1.C=0 and alt1.D=0 and alt1.E=0 and alt1.F=0 and alt1.G=0,
alt1.B=3 and alt1.A=0 and alt1.C=0 and alt1.D=0 and alt1.E=0 and alt1.F=0 and alt1.G=0,
alt2.B=1 and alt2.A=0 and alt2.C=0 and alt2.D=0 and alt2.E=0 and alt2.F=0 and alt2.G=0,
alt2.B=3 and alt2.A=0 and alt2.C=0 and alt2.D=0 and alt2.E=0 and alt2.F=0 and alt2.G=0
;model:
U(alt1)=b2[(u,-1,0)]*A[0,1,2,3]
+ b3[0|0|0].effects*B[0,1,2,3]
+ b4[(u,-1,0)]*C[0,1,2,3]
+ b5[(u,-1,0)]*D[0,1,2,3]
+ b6[0].effects* E[0,1]
+ b7[0].effects* F[0,1]
+ b8[(u,-1,0)].effects*G[0,1]/
U(alt2)=b2*A+b3*B+ b4* C+b5 *D+ b6* E+ b7*F+ b8* G/
U(alt3)=b1[0]+ b2*A+b3*B+ b4* C+b5 *D+ b6* E+ b7*F+ b8* G

$



The generated design is unbalanced, which I think should not be a big problem, however:

1) in block1 there is a certain level from an attribute not present, the same occurs once in block2, for a different attribute.

2) A second thing that I noticed, is that for example a level always occurs with a the same level from a different attribute. As a result you can’t gain information about these levels, because you don’t know on which attribute the respondent based its choice? For example, in case a fertilizer is always slow (a level that occurs 10 times in version A) when the advised volume is 'x4'(which occurs only 2 times in version A) and someone does not want a slow fertilizer, you will never be able to score volumex4? (this is the case in choice cards 1 and 3 from version A). But, in version B, the level 'volumex4' occurs 4 times with attribute 'nutrient release' altering slow and fast.

Did I do anything wrong in the design or will this generate enough information because we assume that respondents will be divided equally between versions A and B?


Thank you very much for suggestions upon this matter.
Best regards
Louise

Dear Louise,

There may not always exist a balanced design due to the specific constraints. In your case, I think rejecting alt1.e>alt1.g is very restricting, this means that most values for alt1.g will be equal to one, it is impossible to obtain an attrnibute level balanced design with this constraint. You can however try to make it as attribute level balanced as possible by adding the notation (lower-upper) to your attribute levels, in which lower is the minimum number of appearances of this level in the design, and upper the maximum number. For example, A[0,1,2,3](3,3,3,3) would mean perfect attribute level balanced, while A[0,1,2,3](2-4,2-4,2-4,2-4) would mean each level occurs between 2 and 4 times in the design. Note that blocking is merely aiming to make each block as much attribute level balanced as possible, but clearly with constraints this may be very difficult to achieve.

I am not sure which attributes you refer to, as you use A,B,C etc in your syntax, and I am not sure what you mean with version A and B (is that alt1 and alt2)? The design you generated will be different from my design, as Ngene uses random seeds, so I cannot check if there is a problem with the design, perhaps you can post the design.

Dear Michiel,

Thank you very much for your answer.

I tried you suggestion to add (lower-upper) like you suggested (2-4,2-4,2-4,2-4) or even (1-5,1-5,1-5,1-5) to the design for the 4level attributes and (5-7,5-7) for the 2level attributes, this gave me the error:

[Modified Federov] ERROR: Unable to find a design that satisfies the specified attribute level frequency constraints. You may need to relax some attribute level constraints, combinatorial constraints (if specified), or dominance checks. The candidate set has 100 rows.

Even when I only included these boundaries for one attribute, I got the same error. When I delete all restrictions (reject, require) then the model runs for a few level appearance boundaries (2-4,2-4,2-4,2-4). But when I add these for all attributes I get the same error as above. What is the problem? What is the consequence of having an unbalanced design?

Related to your post, with versions A and B I meant block 1 and block 2, sorry.
The attributes and levels:
A= price (same price, 20% cheaper, 40% cheaper, 60% cheaper)
B= form (liquid, granulate, semi-solid, a combination)
C= volume needed (same volume, 2x, 4x, 6x)
D= certainty about N-content (certain, possibly 25% variation, 50%, 75%)
E= organic carbon (no organic carbon, as much carbon as in straw containing stable manure)
F= pests and disease (made hygienic, not made hygienic)
G= speed of nutrient release (fast, slow)

Could I email you my design?

Thank you very much for your support.

Best regards
Louise