choice-metrics.com

Dear Ngene experts:
Thank you very much for your previous reply. Now I still encountered some questions, If I can get your help, I would appreciate it very much.

This is my syntax:
Design
;alts=alt1*,alt2*,alt3
;rows=20
;block=4
;eff=(mnl,d)
;model:
U(alt1)=b1.effects[0.001]*A[1,0]+b2.effects[0.003|0.002|0.001]*B[3,2,1,0]+b3.effects[0.001|0.001|0.001]*C[3,2,1,0]+b4[0.001]*D[3,2,1,0]+b5[-0.001]*E[3,2,1,0]/
U(alt2)=b1*A+b2*B+b3*C+b4*D+b5*E/
U(alt3)=b0[0]
$

And my questions are:
1 My variable D and variable E are continuous variables. Do I handle the syntax correctly (i.e. b4[0.001]*D[3,2,1,0]+b5[-0.001]*E[3,2,1,0])? After reading the “Ngene user manual”, I have a question: why does the continuous variable described in Chapter 9.2(Design with continuous levels in Ngene) of the manual use a more complex syntax? Is there any difference between contents described in Chapter 9.2 and the syntax above I used to describe continuous variables D and E(i.e. b4[0.001]*D[3,2,1,0]+b5[-0.001]*E[3,2,1,0])?

2 There is a paragraph in section 7.1.9(Generate efficient design) of the manual:” In general column based algorithms offer more flexibility and can deal with larger designs, but in some cases(for unlabeled designs and for specific designs such as constrained designs ,see Section 8.2) row based algorithms are more suitable.” As shown above, my design is an unlabeled design. Do I need to change my current syntax to a row based algorithms are more suitable.”
As shown above, my design is a unlabeled design. Do I need to change my current syntax to a row based algorithms? Is it ok if I continue to use column based algorithms?

3 As shown above, variable A only has two levels. Is it feasible to use effect coding for variables with only two levels?

4 After running the above syntax. I get a D error of 0.157381 and a S estimate of 1130252.377918. Is this D error ok? What range of D error can we say that this design is efficient? In addition, does the S estimate mean that I need to investigate 1130252 samples? Is there any way to reduce it?

5 There is a “segment number” in “Insert into title” of “Configuration scenario formatting” I have never understood what “segment number” means? So I want to ask for your help.

6 In efficient design, if the iteration history of the last line is not updated after half an hour, can it be regarded as the final design?

Thank you again for your help.

1. A continuous variable can have discrete levels or continuous levels. For example, if price is assumed to be between $1 and $5 then we typically present levels such as $1, $2, $3, $4, and $5. This would be represented by X[1,2,3,4,5] and considers five levels. But since it is a continuous variable, we could actually consider ANY value between 1 and 5, e.g. $1.01, $1.02, $1.03, ..., $4.99, $5, this would be represented by X[1:5] and considers an infinite number of levels. I would recommend using discrete levels.

2. No you do not have to. You only have to change from the default swapping algorithm (which is column based) to the modified Federov algorithm (which is row based) if (i) the default algorithm cannot find a design, or (ii) if you apply ;require or ;reject constraints.

3. Yes, you can apply dummy or effects coding with 2 or more levels.

4. There is no guidance on D-error since this is case specific, sometimes a value of 0.1 is very good, sometimes this is bad. The lower the better is all we can say. Generally values above 1 indicate possible problems with model estimation and if the D-error is infinite/NaN then the model cannot be estimated. A value of 0.15 sounds fine to me. You should ignore the S-error since you have not used informative priors. Only when you use priors that are reasonable, i.e. that come from estimating a model using pilot data, the S-error and S-estimates will be meaningful.

5. A segment could be "male" and "female", see Section 8.4.1 of the manual. You could create separate designs for different population segments.

6. For a simple MNL model (with or without Baysesian priors), yes that sounds more than reasonable. If you are generating a Bayesian efficient design for the panel mixed logit model, which is extremely slow to run, then it may need to run for many days or even much longer.

Michiel