Page 1 of 1
Replication of the Design
Posted:
Thu Feb 17, 2011 9:26 pmby kelvin
If I do a simple mnl design as below. Can I change/add settings so that it will give me exactly the same result each time?
Design
;alts = alt1, alt2
;rows = 4
;eff = (mnl,d)
;alg=swap(random=500, swap=1, swaponimprov=40, reset=10000, resetinc=5000,stop = total(10000 iterations)))
;model:
U(alt1) = b1[0] + b2[1.2] * A[0,1] + b3[1] * B[0,1] /
U(alt2) = b2 * A + b3 * B$
Re: Replication of the Design
Posted:
Fri Feb 18, 2011 5:25 amby miq
I suppose specifying the seed for random draws would help. Try ;rseed (or ;bseed for Bayesian priors) in the manual.
Re: Replication of the Design
Posted:
Fri Feb 18, 2011 6:38 pmby kelvin
Placing rseed equal to some number does not seem to leed to replication. bseed does not seem applicable.
Re: Replication of the Design
Posted:
Sat Feb 19, 2011 9:06 amby Michiel Bliemer
The ;rseed and ;bseed commands are used to ensure that a design will always be evaluated using the same draws.
For the search algorithms, there is no such command (as far as I know, Andrew may know some hidden features), as it does not seem very useful. If you want to keep a design, you can just save it to the project.
But perhaps you have a good reason for wanting to do this? I would be curious to know.
Re: Replication of the Design
Posted:
Sun Feb 20, 2011 3:20 amby kelvin
Point taken. I suppose is only really useful if the user wants to understand the impact of changes in other
settings such as modifications in the priors etc. i.e. make a change, then evaluate what happens to the design.
For example, one might expect that with a small change in a prior the design will remain the same.
I guess if the algorithm is able to find a truely globally optimal design then it would not matter. But in finite runs,
it would be nice to pin down whether what is happening is due to random influences, or whether it is
due to changes in the parameters governing the design.
Re: Replication of the Design
Posted:
Sun Feb 20, 2011 4:56 amby Michiel Bliemer
If you would be mainly interested in analyzing what happens to a specific design when the priors would change, then this can be done by saving the design, and use 'alg = eval' to assess the design with different priors. But from what I understand, you are more interested to see what happens to the optimal design. In that respect, I do not think that there is a way now in Ngene to really know (only if you let it run very long time).