choice-metrics.com

Dear all,

I would like to create a pivot design using the two-step method based on the information I found in:

(Rose, J.M., Bliemer, M.C.J, Hensher, D.A., and Collins, A.C. (2008) Designing Efficient Stated Choice Experiments Involving Respondent Based Reference Alternatives, Transportation Research Part B, 42(4), 395-406).


Consequently, this has brought about the following questions:

I understand from the reference that such a pivot design leads to a single design.
Consequently, my current interpretation is that with this approach everybody receives his or her own reference alternative while the pivot design is identical for all.
Is this correct?

Consequently, assuming the above is correct, one is required to design a homogeneous pivot with as many 'segments' as there are individuals.
Is this correct?

Additionally, for such pivot designs, is there a way to do model averaging? Or is the best I can do: estimate the RPEC panel data model and see how it fits for the other model types?
And finally, using such a design you don't get an S-estimate seeing that you are determining N yourself. Any suggestions on how to determine the necessary sample size?

Kind regards,
Sebastien

One of my colleagues is likely better able to answer your questions, but the answer to your first question is correct. However, in your second question, you can do a pivot design for each segment, thereby creating a heterogeneous design for different segments. Model averaging should work as well I believe, and you should be able to get an S-estimate, as you can define the Fisher matrix in which you use equal weights for each segments. I have asked my colleague to look at your question (probably next week), as he may have some other suggestions (and perhaps correct some of my answers).

Tnx!

To give an update, following the above, I am stuck on making the homogeneous pivot for 78 respondents.
I continually receive the notification:
Error: A ';fisher' property contains fixed weights that do not sum to 1.
While: 77*0.01282 + 1*0.01286 = 1 are the weights that I am using (and I have made the syntax on a single line).

Strangely, when I use 77*0.01 and 1*0.23 it works perfectly...
Does Ngene have difficulties with decimal numbers? and how can I solve them????

So I assume you are using something like (according to the numbers you provided):

;fisher(F) = des1(segment1[0.98714]),segment2[0.01286])

Is that correct? Or could you otherwise post the syntax? I am not quite sure what the problem is.

Hello Ngeners,

When I use:
;fisher(Fish) = des1(N1[0.01], N2[0.01], N3[0.01], N4[0.01], N5[0.01], N6[0.01], N7[0.01], N8[0.01], N9[0.01], N10[0.01],N11[0.01], N12[0.01], N13[0.01], N14[0.01], N15[0.01], N16[0.01], N17[0.01], N18[0.01], N19[0.01], N20[0.01],N21[0.01], N22[0.01], N23[0.01], N24[0.01], N25[0.01], N26[0.01], N27[0.01], N28[0.01], N29[0.01], N30[0.01],N31[0.01], N32[0.01], N33[0.01], N34[0.01], N35[0.01], N36[0.01], N37[0.01], N38[0.01], N39[0.01], N40[0.01],N41[0.01], N42[0.01], N43[0.01], N44[0.01], N45[0.01], N46[0.01], N47[0.01], N48[0.01], N49[0.01], N50[0.01], N51[0.01], N52[0.01], N53[0.01], N54[0.01], N55[0.01], N56[0.01], N57[0.01], N58[0.01], N59[0.01], N60[0.01], N61[0.01], N62[0.01], N63[0.01], N64[0.01], N65[0.01], N66[0.01], N67[0.01], N68[0.01], N69[0.01], N70[0.01], N71[0.01], N72[0.01], N73[0.01], N74[0.01], N75[0.01], N76[0.01], N77[0.01], N78[0.23]);

It works perfectly. I get a homogeneous pivot design for all my respondents (n=78). But these weights are wrong as I am favoring the last respondent.

Similarly, when I use:
;fisher(Fish) = des1(N1[0.01282], N2[0.01282], N3[0.01282], N4[0.01282], N5[0.01282], N6[0.01282], N7[0.01282], N8[0.01282], N9[0.01282], N10[0.01282],N11[0.01282], N12[0.01282], N13[0.01282], N14[0.01282], N15[0.01282], N16[0.01282], N17[0.01282], N18[0.01282], N19[0.01282], N20[0.01282],N21[0.01282], N22[0.01282], N23[0.01282], N24[0.01282], N25[0.01282], N26[0.01282], N27[0.01282], N28[0.01282], N29[0.01282], N30[0.01282],N31[0.01282], N32[0.01282], N33[0.01282], N34[0.01282], N35[0.01282], N36[0.01282], N37[0.01282], N38[0.01282], N39[0.01282], N40[0.01282],N41[0.01282], N42[0.01282], N43[0.01282], N44[0.01282], N45[0.01282], N46[0.01282], N47[0.01282], N48[0.01282], N49[0.01282], N50[0.01282], N51[0.01282], N52[0.01282], N53[0.01282], N54[0.01282], N55[0.01282], N56[0.01282], N57[0.01282], N58[0.01282], N59[0.01282], N60[0.01282], N61[0.01282], N62[0.01282], N63[0.01282], N64[0.01282], N65[0.01282], N66[0.01282], N67[0.01282], N68[0.01282], N69[0.01282], N70[0.01282], N71[0.01282], N72[0.01282], N73[0.01282], N74[0.01282], N75[0.01282], N76[0.01282], N77[0.01282], N78[0.01286]);

I get:
"Error: A ';fisher' property contains fixed weights that do not sum to 1." (while it (77*0.01282 + 0.01286) does sum to 1).

Hence, my question: does Ngene has trouble dealing with decimal numbers?
I have had a similar problem with Stata before, and there it could be solved by using the function float(#).

Kind regards,
and thanks for the quick response!

Sebastien

Hi Sebastien

I can look into this, but I will need your full syntax so that I can attempt a replication of your problem. If you don't want to post it here, send it to contact@choice-metrics.com.

Thanks
Andrew

Thanks for sending through the syntax.

Generally, Ngene has no problem with decimal numbers. The code that parses the fisher weights has a fairly artificial constraint, which limits the precision to three decimal places. This was for historical reasons, so I will seek to relax this constraint in the next point release.

I successfully ran your design with N1-N64 having weights of 0.013, and N65-N78 having weights of 0.012. The syntax is below for your convenience. The small difference in these weights should not be a real concern.

Code: Select all

;fisher(Fish) = des1(N1[0.013], N2[0.013], N3[0.013], N4[0.013], N5[0.013], N6[0.013], N7[0.013], N8[0.013], N9[0.013], N10[0.013],N11[0.013], N12[0.013], N13[0.013], N14[0.013], N15[0.013], N16[0.013], N17[0.013], N18[0.013], N19[0.013], N20[0.013],N21[0.013], N22[0.013], N23[0.013], N24[0.013], N25[0.013], N26[0.013], N27[0.013], N28[0.013], N29[0.013], N30[0.013],N31[0.013], N32[0.013], N33[0.013], N34[0.013], N35[0.013], N36[0.013], N37[0.013], N38[0.013], N39[0.013], N40[0.013],N41[0.013], N42[0.013], N43[0.013], N44[0.013], N45[0.013], N46[0.013], N47[0.013], N48[0.013], N49[0.013],
N50[0.013], N51[0.013], N52[0.013], N53[0.013], N54[0.013], N55[0.013], N56[0.013], N57[0.013], N58[0.013], N59[0.013],
N60[0.013], N61[0.013], N62[0.013], N63[0.013], N64[0.013], N65[0.012], N66[0.012], N67[0.012], N68[0.012], N69[0.012],
N70[0.012], N71[0.012], N72[0.012], N73[0.012], N74[0.012], N75[0.012], N76[0.012], N77[0.012], N78[0.012]);


Tnx Andrew.

I hope this is helpful for other practitioners as well.

Could you also take a look at my first post? Most important is:
Is there a way to do model averaging next, as I am planning on using double layered questions (free choice-foreced choice)?
Or is the best I can do fit both models separately for the ECRPL?

Kind regards,
Sebastien