Two choice experiments choice task generation

[bobby1994]

Best professor,

I have a choice experiment that I want to carry wchich consist of a access (to the train station) and a egress (from the train station) part.

The access consist of walking, public transport, bike, e-scooter and shared modes.
The egress consist of walking, public transport, bike, e-scooter and shared modes.

What I want to do is do two choice experiments, one for access and one for egress and then present these options to the respondent.
So the respondent, during a choice task, a respondent has to answer a question first on access mode choice and then on egress mode choice.

My question is when I determine the choice task for acesss and egress, can I just randomly show an access and egress choice task from the design?
for example there are 9 rows for access and 9 rows for egress. So for the first choice task, I show one row from the egress and one row from the access. Is that oke?

There is another problem. The e-scooter can only be used for both access and egress. So I wanted first to show the respondent a access mode choice task consisting of all the mentioned modes above without the e-scooter and the same for egress and then ask the respondent again if they keep their current choice for access and egress or choose e-scooter for both access and egress. Is this approach oke?

[Michiel Bliemer]

I guess what you are proposing is fine, but you will need to think about how you will estimate the model. Is it a single model for access and egress, or are they two separate models. If two separate models, then it should be fine, but if a single model you will need to think how you will merge the two choices.

Yes you can randomly select choice tasks, that is an acceptable strategy and the default of most survey software. There is some risk that the respondent will not see all access/egress modes, but since you are pooling the data from all respondents this is usually not a real issue.

Your last question, I am not sure. If they make a choice and then you ask whether they want to choose e-scooter for both access and egress, how would you code this in your data set when you estimate the choice model? What is the choice set? And what is the choice, as there seem to be two choices. So you will need to think about the implications for model estimation.

Michiel

[bobby1994]

Best professor,

Thanks for your reply.

I am also thinking of how to program this.

So I have two choice experiments. One for access and one for egress with the options e-scooter, bike, walking and BTM for both access and egress.
With Ngene I generate two experiments.

I explained earlier. During the choice task people will have to answer three questions:
1) for access with the alternatives bike, walking and BTM
2) for egress with the alternatives bike, walking and BTM
3) a question asking respondents if they want to keep their current choice for access and egress or choose the private e-scooter to be used for both access and egress.

During estimation of the model I want to make utility functions of the possible combination of the modes and then estimate the parameters. Combination of modes because respondents choose an access and an egress option. So 3*3 is 9 possible options for access and egress plus 1 for the private e-scooter option. So 10 utility functions and then estimate the parameters.

What do you think?

[Michiel Bliemer]

So you would generate the two designs separately and then merge the data into a single database with 10 choice alternatives. I think that would work.
It would not be possible in Ngene to optimise for the joint model with 10 choice alternatives, but that is okay. You could consider using an orthogonal design for the access and egress choice tasks.

Michiel

[bobby1994]

Best professor,

Thank you for your reply.

So the access and egress choice tasks can be generated separately.

But how to ensure that the context variables are the same for both access and egress choice tasks?

I can include the context variables in the access mode alternatives utility functions and then generate the choice tasks. But how to have the same context variables also for the egress choice tasks?

[Michiel Bliemer]

If you would use orthogonal designs, you could generate three separate designs:
1. Orthogonal design for access mode
2. Orthogonal design for egress mode
3. Orthogonal design for context/scenario

Each choice task then is simply a combination of a row of each of the three designs.

Michiel

[bobby1994]

Best professor,

Thank you for your reply.
Because I have few constraints, I opted for the d-efficient design. Below I have 3 programs as you suggested, access, egress and the context. I have few questions.
1) Do you have any comments on improvement for the codes?
2) I choose to do 12 rows with 3 blocks. With the amount of parameters included could you maybe indicate whether this is fine?
3) In the context program, I have three utility functions. Is it coded right? The third utility function U(purpose) is without the asc. Is that correct?
4) For making combinations can I just select randomly a row from each program? for example to make a choice task can I just select the 1st row from access + the 1st row from the egress + 1st row from the context?
5) Is it allowed to have a bayesian prior for only one attribute and the rest non-bayesian priors?
6) Is it allowed to have priors for all and for one attribute a zero prior?
7) So I have three blocks. I need to implement this in the survey, so three versions. I can just show one block per respondent right, but eventually I should show all three blocks right to different respondents right?

Access

design
;alts = PE, BTM, BIKE, WALKING
;rows = 12
;block = 3
;eff = (mnl, d)

;cond:
if (BIKE.access_BIKE_TT = 8, PE.access_PE_TT < 8),
if (BIKE.access_BIKE_TT = 12, PE.access_PE_TT < 12),
if (BIKE.access_BIKE_TT = 16, PE.access_PE_TT < 16)

;model:

? PE = Private e-scooter, BTM = Bus, Tram, metro, BIKE = Bicycle, WALKING = walking
? for PEPE, the asc and the dummies are removed, because this is the reference alternative

U(PE) = asc1
+ a1 * access_PE_TT[7, 11, 15] ? TT = travel time
+ a2.dummy[0] * com_PE[1,0] ? com = comfort in train, 1 = stored, 0 = holding
/

U(BTM)= asc2
+ b1 * access_BTM_WKT[4,6,8] ? WKT = walking time
+ b2 * access_BTM_TT[4,7,10]
+ b3 * access_BTM_WTT[6,9,12] ? waiting time train
+ b4 * access_BTM_TC[1, 2, 3] ? TC = travel cost
/

U(BIKE)= asc3
+ c1 * access_BIKE_TT[8,12,16]
+ c2 * access_BIKE_PST[1,4,7]
/

U(WALKING) = e1 * access_WALKING_TT[25,30,35]

$




Egress

design
;alts = PE, BTM, BIKE, WALKING
;rows = 12
;block = 3
;eff = (mnl, d)

;cond:
if (BIKE.egress_BIKE_TT = 8, PE.egress_PE_TT < 8),
if (BIKE.egress_BIKE_TT = 12, PE.egress_PE_TT < 12),
if (BIKE.egress_BIKE_TT = 16, PE.egress_PE_TT < 16)

;model:

? PE = Private e-scooter, BTM = Bus, Tram, metro, BIKE = Bicycle, WALKING = walking
? for PEPE, the asc and the dummies are removed, because this is the reference alternative

U(PE) = asc1
+ a3 * egress_PE_TT[7,11,15]
+ a4 * egress_PE_PST[1,4,7] ? PST = parking search time
/

U(BTM)= asc2
+ b5 * egress_BTM_WTB[6,8,10] ? WTB = waiting time BTM
+ b6 * egress_BTM_TT[4,7,10]
+ b7 * egress_BTM_WKTD[2,5,8] ? WKTD = walking time to destination
+ b8 * egress_BTM_TC[1, 2, 3]
/

U(BIKE)= asc3
+ c3 * egress_BIKE_TT[8,12,16]
+ c4 * egress_BIKE_PST[1,4,7]
/

U(WALKING) = e2 * egress_WALKING_TT[25,30,35]

$




Context

design
;alts = weather, tod, purpose
;rows = 12
;block = 3
;eff = (mnl, d)


;model:

U(weather) = asc1 + w1.dummy[0] * weather[1,0] ? 1 = sun, 0 = rain
/

U(tod) = asc2 + y1.dummy[0] * tod[1,0] ? tod = time of day, 1 = daylight, 0 = dark
/

U(purpose) = z1.dummy[0|0] * purpose[1,2,0] ? 1 = education, 2 = leisure, 0 = work

$

[Michiel Bliemer]

That is a lot of questions

1) It looks fine.
2) Yes that should be fine.
3) For context, the full factorial is only 12 rows, so you can use ;fact instead of ;eff, it should give you the same design. The constants you have added are not needed because you are not going to estimate these utility functions.
4) Yes you can randomly select a row from each design. Survey software should be able to do this for you.
5) Yes.
6) Yes, but it is not ideal as the choice probabilities will be less accurate and therefore the design will be less efficient.
7) I am not sure why you use blocks, why not randomly select one row from each of the 12 choice tasks in each design? That way, you have more variation in the data. But if you insist, you can create 3 blocks for each design, but that does not mean that you need to use the first block of the access design with the first block of the egress design, you still have 3*3=9 different combinations of blocks. I would remove the blocks and just randomly choose X design rows from each, where X is the total number of choice tasks you would like to show to each respondent.

Michiel

[bobby1994]

Best Professor,

I see that many posts have been removed.

I still have a question.

So once again the design:
access choice experiment with 12 rows
egress choice experiment with 12 rows
context experiment with 12 rows.

Every respondent have a row from each design.

You suggested last time to not include blocks in order to have more variety.

So I want to randomly select a row from access, a row from egress and a row from context and show that to a respondent.
I want to have 4 scenarios, with each scenario consisting of access, egress and context) shown to each respondent. That is possible right instead of showing 12 or so?

Last time you mentioned checking on correlations between access and egress. THe thing is that both experiments have different attributes, expect for 2 attributes. SHould I still check on correlation?

WHen already looking at analysing the data with an MNL or ML model, is it possible to analyse the access and egress component combined besides analysing them separately?

[Michiel Bliemer]

Yes we were having some issues with a corrupt database and had to revert back to a backup from a week ago.

Since you wanted to use blocking, I suggested checking correlations is exactly because you want to analyse the access and egress components combined. If you also add the scenario to the utility function, you need to check correlations across access, egress, and scenario since this is how the data will be set up. This is to avoid that one of your access attributes is perfectly correlated with one of the egress or scenario variables within the block. Of course even if attributes are different, there can still be correlations between these attributes.

If you would randomly assign profiles from the access, egress, and scenario designs then multicollinearity would likely not be an issue. This would increase variety and would avoid multicollinearity. But you said you could not do random selections and if you stick to specific blocks then you need to confirm that there is no multicollinearity when you combine the data sets.

I hope that we can close this topic?

Michiel