Questions to (pilot study) choice design

[Michiel Bliemer]

You do not need to adjust anything for Ngene. Not in the MNL model estimates and not in the RPPANEL model estimates.

MNL models do not account for panel effects. In the literature, some people adjust the standard errors of MNL models manually, but this is just for reporting purposes in papers, it does not come out of the model by itself. I am not sure how many people actually do this.

RPPANEL models automatically account for panel effects and need no further manual adjustment.

Michiel

[greenvanilla]

Thank you very much!

I have now tried to use the parameter estimates and the (non-adjusted) standard errors from the pilot study as priors and standard deviation for the model.
Following advice provided in the Ngene manual (p. 108), I tried to specificy the model as a random parameter model in the utility functions but optimize it on the MNL model in the eff property,
in order to then get the efficiency measures for the rppanel model (what I am acutally planning to use) based on the estimated MNL model.
(I am finally planning to do do a latent class analysis of panel data; each respondent will answer several choice sets.)

My problem now is that it does not seem to work out, as I get the design and efficiency measures for the MNL model, but when I check the box for RP Panel (or RP), the results for the efficiency measures are reported as "undefined".
I don't know the reason for that. Maybe you have an idea what I did wrong?
Here is the code I used:

Code: Select all

Design
;alts = alt1*, alt2*, alt3
? alt1 and alt2 are unlabeled alternatives, alt3 is a generic no choice option
;rows = 16
;block = 2
;eff = (mnl, d)
;rep = 500
;rdraws = halton (100)

;model:
? priors AND standard errors taken from the results of the pilot study (conditional logit, no cluster-adjustment)
? if coefficients were insignificant, I use zero priors but the estimated standard errors as standard deviation
u(alt1) = b1[n,-0.691,0.087]*x1[0.79,1.49,2.49,3.99] + b2.effects[n,0,0.149|n,0,0.143|n,0.490,0.134]*x2[1,2,3,0] + b3.effects[n,0.163,0.073]*x3[1,0] + b4.effects[n,0.569,0.073]*x4[1,0] + b5[n,3.050,0.693]*x5[0.2,0.4] /
u(alt2) = b1*x1 + b2*x2 + b3*x3 + b4*x4 + b5*x5 /
? alternative specific constant for the opt-out option
u(alt3) = b0[n,-1.546,0.319]

$


Thank you very much in advance!

[Michiel Bliemer]

It seems to be a numerical accuracy problem in the calculation of the Fisher information matrix, computed by simulating a sample of 500 respondents and 100 Halton draws, which consists of a large number of computations with very small values that are summed. In the calculations, one or more of the effects coded standard deviation parameters gets a small negative value on the diagonal of the Fisher information matrix. Since negative values should not occur, Ngene returns an undefined D-error. This sometimes happens when parameter values are close to zero. There is no easy fix I am afraid. This is not an issue with your model, you should be able to estimate a panel mixed logit model.

Michiel

[greenvanilla]

Thank you for the explanation!

So if I understood you correctly, you would just use the MNL design that resulted from the Ngene code presented above and report the D-error from the MNL design,
and later still estimate a panel mixed logit model (I am planning to estimate a latent class model)?

Thanks!

[Michiel Bliemer]

Yes exactly. It is common to optimise for MNL and later estimate a mixed logit or latent class model. It is difficult to optimise for the latter two models, but luckily a survey optimised for estimating an MNL model is also reasonably efficient for estimating other logit models. We discussed this in Bliemer and Rose (2010).

Bliemer, M.C.J., and J.M. Rose (2010) Construction of experimental designs for mixed logit models allowing for correlation across choice observations. Transportation Research Part B, Vol. 44, No. 6, pp. 720-734.

Michiel

[greenvanilla]

Thank you very much!
I will read the paper you are referring to!