How to add a status quo alternative in Ngene designs?

[Shan]

Hi,

I am doing a stated preference experiment designs in Ngene. In my choice experiments, there is a status quo alternative (altcar) out of three alternatives, whose attribute values are determined by answers extracted from previous questions. I use an alternative specific constant in utility functions. Thus, I only got Ngene designs for the other two alternatives (altbus, altmetro). Could you please tell me how to achieve a Ngene design containging all three alternatives? Since I would like to import the design containing all three alternatives into Qualtrics later. I put my codes here for your reference.

Design
;alts=altcar,altbus,altmetro
;rows=72
;orth=sim
;block=8
;model:
U(altcar)= bcar /
U(altbus)= bbus
+bbus,1.dummy[0]*fuel[0,1]
+bbus,2.dummy[0|0]*access[0,1,2]
+bbus,3.dummy[0|0]*freq[0,1,2]
+bbus,4.dummy[0|0]*ttime[0,1,2]
+bbus,5.dummy[0|0]*seat[0,1,2]
+bbus,6.dummy[0]*quiet[0,1]
+bbus,7.dummy[0|0]*fare[0,1,2] /
U(altmetro)= bmetro,2.dummy[0|0]*access[0,1,2]
+bmetro,3.dummy[0|0]*freq[0,1,2]
+bmetro,4.dummy[0|0]*ttime[0,1,2]
+bmetro,5.dummy[0|0]*seat[0,1,2]
+bmetro,6.dummy[0]*quiet[0,1]
+bmetro,7.dummy[0|0]*fare[0,1,2]
$

[Michiel Bliemer]

You are referring to a status quo alternative that is not fixed but rather is respondent-specific. This is also referred to as a reference alternative. Since the reference alternative is different for different respondents, you cannot use a constant for altcar. Further, you have to ask yourself, are the levels of the attributes of altbus and altmetro the same for everyone or do they depend on the levels of altcar provided by the respondent? For example, if one respondent reports a car travel time of 60 minutes and another one reports a 10 minute travel time, then it is not appropriate to use the same bus travel times for all respondents. You can pivot the attribute levels around the reference levels, although this may be more difficult with dummy coded variables.

The Ngene manual has a section on so-called pivot designs around a reference alternative, I suggest you have a look there. In order to optimise your design, you will need to know the attribute levels of the car, which you do not know until you do the survey. Therefore, it is best to assume "average" attribute levels or create a set of experimental designs for different segments of respondents (e.g. high travel time and low travel time respondents).

See also this paper:

Rose, J.M., M.C.J. Bliemer, D.A. Hensher, and A. Collins (2008) Designing Efficient Stated Choice Experiments in the Presence of Reference Alternatives. Transportation Research Part B, Vol. 42, pp. 395-406.

Michiel

[Shan]

[quote="Michiel Bliemer"]

Hi Michael,

Many thanks for your detailed reply. The travel time of bus and metro in my survey design are not fixed either. Their attribute levels are pivoted based on each respondent's reported travel time in bus and metro, respectively. However, I didn't know how to code under this circumastance because I don't know how to assign their reference levels in utility functions. Therefore, I just simply assigned three pivoted levels for travel time of bus and metro (e.g. -10%, -20%.-30%).

I am a little bit cofused with your reply. Do you mean I should pivot the travel time levles of bus and metro's based on the reference levels of altcar? If so, I understand what to do next. I should either assume an "average" attribute levels of altcar or create a set of choice experimental designs for different segments of respodents based on the travel time of altcar.

Shan

[Michiel Bliemer]

Yes you understood correctly. For example, you could create designs with for "short", "medium", and "long" travel times, based on what is reported for the travel time for car, where car, bus and metro travel times are different in each of these designs.

Note that for continuous variable, such as travel time, you may not want to use dummy coding, since for short trips you may get travel times of 10,15,20 minutes, for medium trips 25,35, 45 minutes, and for long trips 60, 90, 120 minutes, which would generate many dummy coded variables in model estimation. If you use travel time as a continuous variable, you can simply estimate a single coefficient for travel time. If you are interested in nonlinearities, you could choose to estimate transformations, e.g. log(traveltime).

Michiel

[Shan]

[quote="Michiel Bliemer"]

Hi Michiel,

Many thanks for your reply. Could I ask another simple question? One of my attributes is frequency of service, whose levles were assigned as every 5, 7.5, or 10 min. My question is that whether this type of attributes is continous variable. I am confused with what kind of code I should use for this type of attributes whose attribute levels are fixed time.

Best,
Shan

[Michiel Bliemer]

You can choose which type of coding you would like to use. While you may end up using linear coding and estimate one parameter for frequency, you can always test and estimate nonlinear effects using dummy coding for frequency. Therefore, designing for dummy coding may be appropriate since it will make sure that you can always estimate a dummy coded effect, even though you may end up estimating a linear effect. If you are interested in calculating willingness-to-pay or elasticities, using dummy coding makes this much more complex. If you are going to do forecasting, dummy coding may not be useful if some modes go every 6 minutes, since this would not fit within your dummy coding, but would be perfectly fine for linear coding. So it all depends on what you want to do with the model.

Michiel