choice-metrics.com

(posted on behalf of Nadine)

I am currently planning my first choice experiment on transport mode choice. In total, we include 10 means of transport in the experiment. To not overwhelme participants we came up with the following idea:
We plan to show each participant two blocks of five situations. 4 means of transport are fixed in every choice situation, the 5th option depends on what the participant stated as option, that he or she uses most frequently.
This means 5 choice options are permanent in one block. The remaining five alternatives should alternate in every choice situation so that every participant sees every option at least ones.

Now I would like to know, how I can estimate the size of the sample needed for a design like this.
I read in the user manual that adding high numbers to alternatives allows a „subset framework“, so I came up with this code so far :

design
;alts = car, pt, cs, bs, fuss, rad, rp, taxi, scooter, carmf
;rows = all
; eff = (mnl,d)
;model:
U(car) = const_car + b1_car*tt[10,14,18,9999] + b2_car*cost_car[2,4,6,9999] +b3_car*access_car[1,3,5,9999] + b4_car*dep_car[2,4,6,9999] /
U(pt) = const_pt + b1_pt*tt_pt[1,3,5,9999] + b2_pt*cost_pt[2,4,6,9999] +b3_pt*access_pt[1,3,5,9999] + b4_pt*dep_pt[2,4,6,9999] + b5_pt*wt_pt[2,3,4,9999] /
U(cs) = const_cs + b1_cs*tt_cs[10,14,18] + b2_cs*cos_cs t[2,4,6] +b3_cs*access_cs [1,3,5] + b4_cs*dep_cs [2,4,6] /
U(bs) = const_bs + b1_bs*tt_bs [14,18, 22] + b2_bs*cost_bs [2,4,6] +b3_bs*access_bs [1,3,5] + b4_bs*dep_bs [2,4,6] /
U(fuss) = const_fuss + b1_fuss*tt_fuss [25,30,35,40,9999] /
U(rad) = const_rad + b1_rad*tt_rad [20,25,30,9999] /
U(rp) = const_rp + b1_rp*tt_rp [1,3,5] + b2_rp*cost_rp[2,4,6] +b3_rp*access_rp[1,3,5] + b4_rp*dep_rp[2,4,6] + b5_rp*wt_rp[2,3,4] /
U(scooter) = const_scooter + b1_scooter*tt_scooter[10,14,18] + b2_scooter*cost_scooter[2,4,6] +b3_scooter*access_scooter[1,3,5] + b4_scooter*dep_scooter[2,4,6] /
U(carmf) = const_carmf + b1_carmf*tt_carmf[10,14,18,9999] +b5_carmf*wt_carmf[1,3,5,9999]
$

One can get an estimate of the size of the sample by looking at the S-estimates of the design. These indicative values, however, are only meaningful when using informative priors (e.g. from a pilot study). Without informative priors it is not possible to tell what the sample size should be since this varies from study to study.

In order to select a subset of mode alternatives you are following the procedure outlined in the manual, which requires the generation of a candidate set, so you will need to add ;alg = mfederov(candidates = .... ) into your syntax. You also need to specify the sign of some of the attributes in order to let the level 9999 (I would prefer to use a larger value, eg 99999) result in a large negative utility, e.g. b1_fuss[-0.001].

Partial choice set designs are somewhat complex to generate. Creating a candidate set where only a subset of mode alternatives appear in a choice task (alternatives with attributes that have levels 99999 can be omitted from the choice task) is something that you will need to create yourself. If you find this difficult, there is another way of creating a partial choice set design without the need for an external candidate set, which involves reformulation of your utility functions. This is a new method and has not yet been published, but I am willing to explain.

Michiel

Hey Michiel,

thank you so much for your answer.
I am still not quite sure what you mean with "I will have to do this by myself":
I know which combinations will be allowed in the experiment. However I don't know yet how to set the parameters for the different alternatives in one choice task.

To recap: In total we have 10 possible options for means of transport. But we only want to include 6 at any time. 4 modes should be constant and appear in every situation, 1 mode that is the "most frequently used" which also appears in every situation but depends on the respondents choice (to have a more realistic situation for him/her) and the 6th option will change in every choice task. If one respondent gets 5 tasks he sees every mode at least ones.

In total, there are 15 combinations possible to "pair" 2 out of 6. I already designed all the 15 combinations in seperate choice experiments. [e.g. foot and bike, foot and pt etc. ]
What I would need in the end is a design that combines different rows out of different experiments, e.g. if I walk most often I would need the following choices:
-foot & bike
-foot & public transport
-foot & car
-foot & car passenger
-foot & cab
Each represents one line out of the different choice experiments. But can I combine those lines randomly?

Thanks

I refer to Section 8.11 of the Ngene manual where the external candidate set is explained. Figure 8.31 on page 182 shows the Excel spreadsheet that you will have to construct yourself.

I am not sure what you mean with "different rows out of different experiments", I am not sure Ngene can do what you want.

Michiel

Hello,

thank you so much again. By now, I managed to create a candidate-set-file and to finally import it.
I still struggle with a design - problem while blocking the choice situations.
Out of 2 variable alternatives, one should appear in every situation that a respondent answers and the other should alternate.
So far, the blocks Ngene creates are random. Is there a command to specify the block-design, in a way that every block looks like this:
-foot & bike
-foot & public transport
-foot & car
-foot & car passenger
-foot & cab
or
-bike & foot
-bike& public transport
-bike& car
-bike& car passenger
- bike & cab etc.

Or do maybe know the effects of changing the blocs manually on the experiment or the model evaluation afterwards?

The aim of blocking is to achieve some level of attribute level balance. Note that changing the blocking has no statistical impact, the efficiency of the design remains the same, but of course you should avoid that a participant only sees "high levels" and no variation in levels within a block. You can usually change the blocking of the design without any problems.

An alternative is to create 2 separate designs, one for foot and one for bike, and in your survey pick choice tasks from each design.

Michiel

Great, that's what I already did to generate the candidate set. So then I will just assign the blocks manually. (Keeping the variation in mind, of course)
Thank you so much for your support!