choice-metrics.com

Hi,

I'm new to Ngene (which I really like) so please excuse any silly questions. I have several:

1. I saw in prior posts that using a Bayesian prior parameter distribution (say, the uniform) will work if nothing else is known about a parameter other than the sign. For example, for price it was suggested that a uniform from -1 to 0 be used. Is there any literature or general rules of thumb (ROTs) explaining/describing/discussing the different distributions for different types of variables such as price? Any other suggestions besides the uniform? Any general guidance?

2. I ran a simple design for a tablet computer (memory size in GB; display size in inches; battery life in hours; and price in dollars) which is:

Design
;alts = tab1, tab2
;rows = 16
;block = 2
;eff = (mnl)
;model:
U(tab1) = B0 + B1*memory[16, 32] + B2*display[9.7, 10.1] + B3*battery[10, 14] + B4*price[300, 500, 800] /
U(tab2) = B1*memory + B2*display + B3*battery + B4*price
$

The result was:

Design
Choice situation tab1.memory tab1.display tab1.battery tab1.price tab2.memory tab2.display tab2.battery tab2.price Block
1 32 10.1 10 500 16 9.7 14 300 1
2 32 9.7 10 300 16 10.1 14 800 2
3 32 10.1 10 800 16 9.7 14 300 1
4 16 10.1 14 300 32 9.7 10 800 1
5 32 9.7 14 300 16 10.1 10 800 1
6 32 9.7 14 800 16 10.1 10 300 2
7 16 10.1 14 800 32 9.7 10 300 2
8 16 9.7 10 300 32 10.1 14 800 1
9 16 9.7 14 500 32 10.1 10 500 2
10 32 10.1 10 500 16 9.7 14 500 2
11 16 9.7 10 500 32 10.1 14 500 2
12 16 9.7 14 800 32 10.1 10 300 1
13 16 10.1 10 300 32 9.7 14 800 2
14 16 9.7 10 800 32 10.1 14 300 1
15 32 10.1 14 300 16 9.7 10 500 1
16 32 10.1 14 500 16 9.7 10 500 2


Notice the last three lines. In the last two, no one would pick the second option; in #14, no one would pick the first. How can this be avoided?

3. When a design is created such as the above, Ngene keeps iterating. I apparently have to manually stop it. Why?? Shouldn't there be a setting somewhere that stops it after, say, 1000 iterations?

Any help for this Ngene beginner is appreciated.

Thanks,

Walt

They are all good questions. Some answers below.

1. Bayesian priors can come from a pilot study or literature, for example using normal distributions with the mean equal to the parameter estimate from literature or a pilot study, and the standard deviation equal to the corresponding standard deviation. If no literature or pilot study is available, but you still would like to use the information on the sign, you can indeed use a uniform distribution (between -1 and 0, but could also be -0,1 and 0 or -0.5 and -0.2, depending on what you think could be the upper and lower limit for the price parameter), but also a lognormal distribution could be used for example. There are no real guidelines for this I am afraid.

2. In the example you describe you have two generic alternatives. To avoid "silly" dominated choice alternatives, you can add a star (*) behind the alternative to indicate it is generic, such that Ngene will try to avoid these choice tasks. Clearly, Ngene can only do this if you provide priors for the parameters, because you have to tell Ngene which parameteres are positive and which are negative. So if you are not sure of the prior, but you are sure of the sign, just use a very small prior (for example 0.01 or -0.01 or so), then at least Ngene knows the sign and can avoid these choice tasks.
So, it would become:
;alts = tab1*, tab2*

3. Use can set a maximum number of iterations or a maximu time limit. Myself, I do not see why we should, so I just let it run as long as possible (for example run overnight and take the best design found the next morning). Most algorithms will keep running indefinitely. It is possible to force an algorithm to stop after a certain amount of time, or a certain number of iterations. For example,
;alg = swap(stop=total(10 mins))
will run the swapping algorithm for a total of 10 minutes, and
;alg = mfederov(stop=total(100000 iterations))
will run the Modified Federov algorithm for a total of 100000 iterations. I refer to the manual for more options.