choice-metrics.com

Dear fellow researchers,
I am conducting a choice-based conjoint analysis in my research. Within the experiment, I´d like to compare two alternatives to a fixed status quo per choice task.
The status quo is modelled as a constant "none-option" and will be estimated as a constant. I am wondering, if I can use the levels which define the status quo also as levels for the alternatives in order to give the respondents the chance to select between options which are not completely different from status quo but only differ in some attributelevels but not over all attributelevels.

A possible choice task could look like this:
Status Quo via constant:
Attr. 1 status quo level = red
Attr. 2 status quo level = 300mph
Attr. 3 status quo level = Electric engine

Alternative 1:
Attr. 1 level =blue
Attr. 2 level = 300 mph (status quo level)
Attr. 3 level = diesel engine

Alternative 2:
Attr. 1 level = red (status quo level)
Attr. 2 level = 150 mph
Attr. 3 level = gasoline engine

Do you see any problem with regards to design or analysis that could come from this?
Any advice is highly appreciated.
Thanks and best,
Julia

Modelling a "no choice"/"optout" alternative is NOT the same as modelling a "status quo" alternative. In an optout alternative, you would merely use a constant, but in a status quo alternative (at least in an unlabelled experiment) you actually specify the utility function where the attribute levels are simply fixed. In an unlabelled experiment the status quo alternative therefore has exactly the same utility function as the other alternatives, but the attribute levels are fixed values. For example:

Code: Select all

design
;alts = alt1, alt2, sq
;rows = 12
;block = 2
;eff = (mnl,d)
;alg = mfederov(candidates = 500)
;require:
sq.COLOUR = 0,
sq.ENGINE = 1
;model:
U(alt1) = colour.dummy[0|0] * COLOUR[1,2,0] ? 0 = red (base), 1 = green, 2 = blue
        + speed             * SPEED[100,150,300]
        + engine.dummy[0|0] * ENGINE[1,2,0] ? 0 = diesel (base), 1 = electric, 2 = gasoline
        /
U(alt2) = colour * COLOUR + speed * SPEED + engine * ENGINE
        /
U(sq)   = colour * COLOUR + speed * SPEED_SQ[300] + engine * ENGINE
$

Such a model can be estimated as long as it does not contain a dummy variable where the level for the status quo ONLY appears in the status quo and not in the other two alternatives. If a dummy coded level only appears in the status quo alternative then this attribute is captured via a constant in the status quo alternative (while keeping the other attributes in the utility function).

Michiel

Dear Michiel,

thank you very much for your prompt and very helpful reponse.
In order to make sure that I get you right:
In the design of the DCE, I could treat the status quo levels as "normal" levels when generating a balanced overlap design. My respondents would then choose between three alternatives containing the status quo levels. Only in the analysis (after data collection), I would then model the status quo via fixed attribute levels?

Thanks for your support!
Julia

The model you use in model estimation and in the design would be identical. In both cases, the status quo alternative has fixed attribute levels, while the other two alternatives have varying attribute levels.

Michiel

Ok, I understand.
Can you explain, why you are saying: "at least in an unlabeled experiment". Why would there be a difference für labeled experiments where there would be something like "status quo", "alternative A", "alternative B".

Thank you very much for your support!

With a labelled experiment I mean here alternatives where the atttributes are different across alternatives or where coefficients are different across alternatives.

For example"

U(apple) = b1 + b2* screensize + b3 * harddisksize + b4 * speed + b5 * price
U(dell) = b6 + b7 * screensize + b8 * harddisksize + b9 * speed + b10 * price
U(sq) = b11 * screensize + b12 * harddisksize + b13 * speed + b14 * price

In this case, coeffients are all alternative-specific and screensize, harddisksize, speed, and price are fixed for the status quo alternative, which means that it is not possible to estimate b11, b12, b13, and b14. So only when coefficients are generic across alternatives you can include the attributes of the status quo, otherwise it will end up in the constant. In you case, I assume all coefficients are generic so that it is essentially an unlabelled experiment where the status quo may have a constant.

Michiel

I see, thank you for the comprehensive explanation.
In my case, I have one attribute of the five attributes which is only shown when another attribute takes a certain levels (alternative specific design). Does it in this case make sense take the status quo as a constant and not involve the status quo levels in the alternatives?

Thank you very much!

In that case you would still include the other attributes. If that specific attribute is added as a continuous variable, i.e. b1 * X1, then you include all five attributes in the status quo alternative. If that specific attribute is dummy coded, then there are two cases: (i) the status quo level does NOT appear in the other alternatives, then you exclude this attribute and replace it with a constant but do include the other four attributes, or (ii) the status quo level DOES appear at least once in the other alternatives, then you can include all five attributes in the status quo alternatives. This is all a matter of parameter identifiability.

Michiel

Dear Julia and Michiel

I think I have a similar question but I would like to verify I correctly apply Michiel's recommandation

I have an unlabelled experiment, with 2 alternatives + opt-out option (= no insurance).
For only one of the attribute (A), the level of the status quo is also a level possible for the alternatives. For all the other attributes, the level for the status quo ONLY appears in the status quo and not in the other two alternatives (no insurance means coverage = 0% and premium=0%).

U(GA) = b1.dummy[0|0|0]*A[3,2,1,0] ? A Commitment to reduce fungicide: A=3 TFI -50% ; A=2 TFI -30%; A=1 DST; A=0 no commitment to reduce fungicide
+ b2.dummy[0|0|0]*B[100,95,90,75] ? B coverage (% of losses) 100,95,90,75
+ b3.dummy[0]*C[1,0] ? C guaranteed yield : C=0 reference in the area , C=1 olympic average
+ b4.dummy[0]*D[1,0] ? D dammages evaluation: D=0 expert; D=1 index
+ b5[0]*E[0.1,0.5,1,2,4,7] / ? E insurance premium % of insured capital 0.1,0.5,1,2,4,7

U(GB) = b1*A + b2*B + b3*C + b4*D + b5*E /

I am not sure what is the correct way to specify U(SQ)
Option 1: U(SQ) = b0[-1]
Option 2: U(SQ)= b1*A[0] ? A=0 no commitment to reduce fungicide
+b0[-1]

Any opinion on whether it's option 1 or 2 for U(SQ) ? or may be something else !

Thanks a lot
Marianne

An opt-out alternative typically only has a constant while a status quo alternative has attributes. So you need to decide what type of alternative you are adding. For an opt-out it is typically clear what you are getting (namely, nothing), which can be captured by a single constant.

Michiel