Pivot design with changing alternative based on responses

[Chaoh]

Dear Administrators,

Right now I'm trying to create a design for stated preference questions, and I have a couple of questions about the design that you may help answer. Below is a description of the issue that I’m having:

Background information
In the revealed preference (RP) section, I will ask information about the respondent’s most recent trip, such as Origin and Destination (OD), primary mode (auto, bus, LRT, walk, cycle, etc.) for that trip. Based on the OD, respondents will be categorized into multiple segments (Segment 1 to N). Note: available travel modes in each segment might be different. In the RP, in addition to the primary mode, we will also ask the respondent to provide two preferred alternate modes pretending the actual primary mode is not available. Then these three alternatives (primary, two alternates) plus other two (No Travel, and Change Destination) will be used in the following stated preference (SP) questions with attribute levels pivoted around those in the revealed preference. For example:

a. A few respondents (Group 1) in Segment 1 indicate that their REVEALED primary mode is Auto, LRT as the Alternate Mode 1, and Local Bus as the Alternate Mode 2. For Group 1 in SP questions, respondent will be given three alternatives (Auto, LRT, Local Bus) plus other two (No Travel, Change Destination).
b. Another bunch of respondents (Group 2) in Segment 1 with LRT as the revealed primary mode, Local Bus as the Alternate Mode 1, and Walk/Cycle as the Alternate Mode 2. For Group 2 in SP questions, respondent will be given three alternatives (LRT, Local Bus, Walk/Cycle) plus other two (No Travel, Change Destination).

In my understanding of creating stated choice experiments, three steps need to be followed (quoted from the NGene Manual):
Step 1: a complete model specification with all parameters to be estimated has to be determined.
Step 2: based on the model specification, an experimental design type (orthogonal, efficient, etc) has to be selected, then the SP design can be generated.
Step 3: a questionnaire (either paper, internet, CAPI, etc) is created based on the underlying experimental design.

Following the steps above, for Group 1 and Group 2, their model specifications are different because combinations of alternatives are different, and therefore I will need one design for each group. If this is true, then I will end up with multiple designs for each of the N segments. This is technically doable but practically infeasible, because I may end up with too few responses for each combination in a given segment (bearing in mind as a rule of thumb 200 is a minimum for each design/model), so it would be difficult to have multiple design for each segment.

Proposed Solutions
My current preliminary plan is to use orthogonal design (fractional factorial) for each of the N segments. I’m thinking of possible solutions:
1. In the SP questions, list all available alternatives for a given segment, in this way, I then could have an all-inclusive single design for each segment. But this will create additional burden on the respondents facing too many alternatives (up to 7) to choose in one choice set.
2. In the SP questions, list only Primary Mode, Alternate Mode 1 (probably Mode 2 as well) plus the other two (No Travel, Change Destination), but still use just one all-inclusive design for each segment. I can do this because the property of an orthogonal design. An orthogonal design “is said to be orthogonal if it satisfies attribute level balance and all parameters are independently estimable”. Following this property of an orthogonal design, in theory for a given segment in our case (say Segment 1), I can create a single factorial design including all possible alternatives (Auto, Bus, LRT, etc.) – in the design only, but not in the SP questions. I need to specify nested logit models for each combinations (for example, Model 1 for Group 1, Model 2 for Group 2) because this will not change the orthogonality because the parameters can be estimated INDEPENDENTLY. But this approach might create problem for modelling purpose - I may end up with getting two few responses for a particular group.

My questions can be summarized as:
Q1: Would the above two proposed solutions work? Which one would you think is better? If solution 1 would work, I can also use efficient design in addition to orthogonal design, right?

Q2: If neither proposed solutions would work, could you give me some hints on handling the pivoted design with changing alternatives?

Q3: I would construct nested logit models for the survey data, is NGene 1.1.2 updated with the ability generating design for nested logit model? I noticed that Michiel mentioned in his reply to another post indicating that "a design that is optimised for the MNL model will also work for the nested logit model, it will just not be as efficient.... So using Bayesian priors, I would not be afraid at all to use an MNL optimised design for estimating a nested logit model later on."

Q4: How should the two alternatives (No Travel, and Change Destination) be included in the design? I won’t specify utility functions for the two alternatives – reasonable assumption? In the NGene Manual (Page 118), I read that “Independent of the form of the no choice alternative, one or more no choice alternatives can be accounted for in generating a design by naming an alternative in the alts property but not specifying a utility function after the model property”. Does this mean that I can simply add two alt in the syntax ;alts as below for example.

Design
;alts = alt1, alt2, alt_NoTravel, alt_ChangeDestination
;rows = 12
;eff = (mnl,d)
;model:
U(alt1) = b11[-0.2] + b2[1.2] * A[0,1,2] + b3[2.5] * B[0,1] /
U(alt2) = b12[0.3] + b2 * A + b4[1.1] * C[2,4,6,8] $


This post is too long, and I’m sorry to bother you. I appreciate for any advice and suggestions. Thanks in advance.

Chao

[Chaoh]

Hi all, I found answers for my questions (#1, and #2). But #3 and #4 still remain unresolved. Thanks.

[Michiel Bliemer]

To answer your last two questions:

3. Ngene 1.1.2 does not have the capability to optimise for nested logit models, so it is best to optimise for the MNL model.

4. It depends on how you would like to implement no travel and other destination in your demand model, but I assume you just want to estimate a model with constants.

So in your case it would be something like:

design
;alts = mode1, mode2, notravel, changedest
...
;model:
U(mode1) = asc_mode 1 + ... /
U(mode2) = asc_mode2 + ... /
U(notravel) = 0 /
U(changedest) = asc_dest
$

Since you have 4 labelled alternatives, you can include 3 alternative specific constants (asc). In this case I have chosen to set the notravel to zero, such that the other constants are relative to not travelling. The constant for change destination will determine the relative preference of choosing another destination over not travelling.

[mariareese]

Good information you have shared to us, thanks