sample size in orthogonal design

[Anat Tchetchik]

Hi all,

I have ran the following orth desing (see (2) below) for a pilot phase of a new study (I had few priors from previuos study and for some param. I knew only the sign)
I received only 1 design which looks fine but what bothers me is that
in the MNL efficiency measures (see (1) below) I recieve a huge sample size: S estimate equals: 54,432,463
is it an indication that the design is infeasible?
MNL efficiency measures

D error 0.272616
A error 2.635671
B estimate 99.676321
S estimate 54432463.607656

Prior b2 b3(d0) b3(d1) b3(d2) b4 b5(d0) b6
Fixed prior value -0.1 0 0 0 0.1 0 -0.001
Sp estimates 1474.660861 Undefined Undefined Undefined 21.259181 Undefined 54432463.607656
Sp t-ratios 0.05104 0 0 0 0.425092 0 0.000266

(2)Design
;alts = alt1, alt2, alt3, alt4
;rows=128
;orth=sim
;block=16
;model:
U(alt1) = b2[-0.1]*price[.091,0.117,0.143,0.169,0.196,0.222,0.248,0.274] +b3.dummy[-0|-0|-0]*airline[0,1,2,3]
+b4[0.1]*direct[0,1] +b5.dummy*jet[0,1] +b6[-0.001]*risk[0.1, 0.02, 0.01, 0.0067,0.005,0.004,0.002,0.001] /

U(alt2) = b2*price + b3*airline+ b4*direct+ b5*jet+ b6*risk/
U(alt3) = b2*price + b3*airline+ b4*direct+ b5*jet+ b6*risk/
U(alt4)= b1
Best,
Anat

[johnr]

Hi Anat

This is a common issue with poor prior selection. Take for example, the average of your price attribute levels (0.1825) and calculate the marginal contribution to overal utility for that attribute (0.1825 * -0.1 = -0.01825). Now do the same for the risk attribute. The average of the levels is 0.018588 and the marginal contribution to overal utility for that attribute is (0.018588 * -0.01 =) -0.000185875. The contribution to utility of the price attribute is 98 times larger than that of the risk attribute. This means that relative to the price attribute, the risk attribute has to try much much much harder to have any influence on utility. To make any difference, it is going to require a substantial sample size which is what you are finding in your design.

In selecting priors, these should be scaled to the magnitudes of the attribute levels as would occur in estimation (i.e., the parameters are scaled to the magnitude of the levels - if you multiply the data by 10, the parameter will be divided by 10). If you don't scale them, then you will run into this problem all the ime.

Hope this helps

John