Page 1 of 1

Software error_large number for attribute level

PostPosted: Tue Dec 02, 2014 3:13 am
by Elnaz
Dear Ngene team,

I'm trying to do a pivot design with a fixed level for the first alternative. One of my attributes has a fixed level of 22,000, but unfortunately any time I try to run my syntax I get an error and the software closes automatically, however when I try it with a smaller number such as 2,200, I don't receive any error. Could you please let me know how I could deal with this issue?
Below is an example of my syntax:

Design
;alts = alt1, alt2, alt3, alt4

;rows = 48
;block = 8
;eff = (mnl,d)
;rep = 1000
;rdraws = halton(500)

;model:

U(alt1) = b1[-0.2] * X1.ref[22000] + b3[-0.1] * X3.ref[513]/

U(alt2) = b1 * X1.piv[0%,50%] + b2[0.1] * X2[0,7.5,15] + b3 * X3.piv[-50%,-25%,0%]/

U(alt3) = b1 * X1.piv[0%,50%,100%] + b2 * X2[0,7.5,15] + b3 * X3.piv[-50%,-25%,0%]/

U(alt4) = b1 * X1.piv[0%,50%,100%] + b2 * X2[0,7.5,15] + b3 * X3.piv[-75%,-50%,-25%]$


Regards,
Elnaz

Re: Software error_large number for attribute level

PostPosted: Wed Dec 03, 2014 7:25 pm
by Michiel Bliemer
Your prior is way too large, your model does not make behavioural sense. Your attribute level and your parameter value are not consistent.

Note that U = -0.2 * 22000 - 0.1 * 513 = 4451.3

Note that the logit model evaluates the exponent of the utility, i.e. exp(U) = exp(4451.3), which is extremely large and the computer cannot calculate this.
Your priors should be something like -0.0002 and 0.001 to make sense. Please choose your prior values carefully, for example doing a little pilot study.

Re: Software error_large number for attribute level

PostPosted: Fri Dec 05, 2014 3:19 am
by Elnaz
Thanks so much. Large prior was the issue indeed.
We actually plan to do a pilot. Do you suggest to do orthogonal design without any priors for the pilot?

Thanks again,
Elnaz

Re: Software error_large number for attribute level

PostPosted: Fri Dec 05, 2014 6:27 am
by Michiel Bliemer
Instead of an orthogonal design, you can simply choose very small priors close to zero (but still having the correct sign so you could possibly also detect and remove dominant alternatives). A design with zero priors is quite similar to an orthogonal design.

Re: Software error_large number for attribute level

PostPosted: Tue Dec 09, 2014 7:07 am
by Elnaz
Right, my only concern is that at the end, some variables may not even have the sign that we expected. I know it's very unlikely, but for example in the choice of a vehicle we usually expect a negative sign for the price of vehicle but it could also be positive arguing that more expensive vehicles have better quality!