Variable not changing among alternatives

[fherrero]

I'm not sure whether this topic goes here or not, so I ask the moderators to move it if necessary.
I am quite new in SC designs and I am finding myself stuck with some issues. My model pretends to evaluate the decision of making a transfer or not in public transport. There are two alternatives: a no transfer trip versus a one transfer trip.

1. We have thought to make a first pilot with a efficient d-error mnl model to estimate the priors, then a second pilot with a Bayesian eff model to improve the priors and finally the final Bayesian study. Is this approach correct?

2. In our model it is very important that the sum of the different variable times in the no-transfer alternative is higher than the sum of the time variables in the one-transfer alternative, so we need the ;require property and thus, the mfederov algorithm; but I do not understand what is the number of candidates for and what are the consequences of changing it and what are the recommended values for it.

3. Finally, our model has a variable “snow” which equals to 1 if it snows and 0 in other case. We have found that this variable may be very important. However, it makes no sense changing it among alternatives, because when the user makes the choice of which trip they are going to take, the weather is the same in both alternatives and they cannot choose a sunny trip versus a snowy trip.
The problem is that when we try to add a variable with the same value in both alternatives (“snow” variable), NGENE do not find any design (number of invalid designs increases continuously). I have found out that this is because the Fisher matrix is singular. We have tried to include the “snow” variable only as an interaction, as the manual suggests. However, when we do this, the number of invalid designs continues increasing, ant the few valid designs obtained have Fisher matrices with a determinant really close to zero (~10^-13), so we are not confident about these designs.
Is there any other way to include this “snow” variable?

Hoping you may shed some light onto this,

Fernando

[johnr]

Hi Fernando

1. There is no theoretical correct approach, however your suggestion makes sense. The priors represent, as in any Bayesian method, your beliefs as to the state of the world. Where your beliefs are derived from however is a non-theoretical question. There are lots of transport papers (not sure specifically dealing with transfers as the choice, but certainly including a transfer penalty) that could help inform your priors, however a pilot of the population of interest certainly makes sense.

2. The algorithm requires that candidate sets first be constructed, and then swaps these in and out of the design for testing. Ideally, you would want the full factorial (all possible candidate sets), or in the case of constraints, all choice tasks that meet the constraint criteria. Depending on the design problem, this may represent billions of possible candidates. Given finite computing resources, it may not be possible to generated all of the candidate sets, hence you may need to reduce the number constructed.

3. There is an option in Ngene to include designs in which the scenarios are also experimentally varied. This is discussed in the advanced section of the manual under the heading "Designs within designs: Designs with scenarios in Ngene". The current set-up requires you to set up interactions between these common attributes and the attributes of the alternatives. We are playing around with alternative utility structures to do this, but have yet to implement them. This should work for what you want to do for the moment however.

John

[fherrero]

Thank you so much for your response.

I have just read again the section “Designs within designs: Designs with scenarios in Ngene” as you indicated. With mfederov algorithm mimic variables are not allowed. However I have managed to make equal parameters through restrictions (;require).
I think I need to explain my model in further detail:
I actually have two weather variables: “rain” if it rains and “snow” if it snows. I also have many time related variables, such as access time (tacc) and egress time (tegr).
As you have suggested, I have introduced the rain and snow variables as interactions with tacc and tegr:
U(0T) = tacc0[-1]*tacc0[1,3,5] + tegr0[-1]*tegr0[1,3,5] + … + rain0a[-1]*rain[0,1]*tacc0 + rain0e[-1]*rain*tegr0 + snow0a[-1]*snow[0 ,1]*tacc + snow0e[-1]*snow*tegr
U(1T) = tacc1[-1]*tacc1[1,3,5] + tegr1[-1]*tegr1[1,3,5] + … + rain1a[-1]*rain*tacc1 + rain1e[-1]*rain*tegr1 + snow1a[-1]*snow*tacc + snow1e[-1]*snow*tegr
If I run this model, the program never finds any design, but it finds many invalid designs (current number of invalid designs increases continuously). However if I only include the interactions with tegr but not with tacc, the program finds many designs and current number of invalid designs remains zero and I don’t understand why this happens.
I have been suggested that I may introduce the interactions only in one of the utility functions, but I am not sure if this makes sense for interpreting the model later. If I introduce the snow variable, for example, only in the utility function of the one-transfer trip, am I not loosing information? Can I then obtain information about the influence of snow in people’s decision of using trips with transfers?

Thanks again for your respones

Fernando

[johnr]

Hi Fernando

What are the alternatives in your model? The problem with the syntax I can see is that

... + rain0a[-1]*rain[0,1]*tacc0 ...
... + rain1a[-1]*rain*tacc1 ...

rain can take the value 0 in the first alternative and 1 in the second (or vice versa). This suggests that it might be raining if you take the (train?) but not raining if you take the (bus?), unless you have handled this in the restrictions. Without the full syntax, it is not possible to locate what the exact issue causing the problem is. Can you post your full syntax?

John