ApolloChoiceModelling forum

Hi,

thanks a lot for privode this amazing package apollo. I have two questions about ICLV model.

Question 1: I am trying both sequential estimation and simultaneous estimation. sequential estimation is by lavven package and simultaneous estimation is by apollo. The estimated parameters of latent variables are quite different. There are four latent parameters.
Apollo: lambda.TRAIN1 6.3731 lambda.SM1 1.7476 lambda.TRAIN2 8.6003 lambda.SM2 3.4354
Lavven: lambda.TRAIN1-0.503092 lambda.SM1-0.575931 lambda.TRAIN2 0.532848 lambda.SM2 0.341701
The results of apollo looks wired. the attachment is my data and my programming. could you please tell me if I make some mistakes?
I cannot upload any attachment. the system tell me "Invalid file extension: HCM contious2.R" everytime when I try.



Question 2: Apollo takes a long time. lavven needs several minutes. i think maybe it is because lavven doesn't take panel effect into consideration. I also want to decrease the panel effect of apollo. so I change
"randcoeff[["LV1"]] = gamma_AGE1*AGE + gamma_GENDER1*GENDER + gamma_INCOME1*INCOME + eta1"
to
"randcoeff[["LV1"]] = gamma_AGE1*AGE + gamma_GENDER1*GENDER + gamma_INCOME1*INCOME"
and also delete "P = apollo_avgInterDraws(P, apollo_inputs, functionality)"
the results become
gage.TRAIN -0.2121 NaN NaN 163.7251 0.00
ggender.TRAIN -1.3568 NaN NaN 296.4883 0.00
gincome.TRAIN 0.2102 NaN NaN 265.9663 0.00
gage.SM -0.6425 NaN NaN 61.7211 -0.01
ggender.SM -1.0098 NaN NaN 111.6463 -0.01
gincome.SM 0.6646 NaN NaN 100.2392 0.01
lambda.TRAIN1 -0.3786 NaN NaN 453.2218 0.00
lambda.SM1 0.1751 NaN NaN 167.6733 0.00
lambda.TRAIN2 -0.7047 NaN NaN 300.0220 0.00
lambda.SM2 -1.1982 NaN NaN 115.2265 -0.01
What should I do to decrease the panel effect of apollo?

Hi

1. Sequential and simultaneous estimation should not be expected to give the same results. WIth sequential estimation, you are optimising the latent variables only for the indicators, while, with simultaneous, you're optimising it for the indicators and choice at the same time.

2. I don't know Lavven so wouldn't want to compare speed. But a sequential model would for example in general be expected to be quicker. But what you're doing is in fact removing the random component of the latent variable, so it's no longer a latent variable at all.

Best wishes

Stephane

Thanks a lot!

Could I ask how to do sequential estimation in Apollo? I tried but I am not good at programming. I don't know how to calculate the values of the latent variables. Attached is my failed attempt. Thanks again for your patience and support.

Best wishes!

Hi

what you need to do is to first specify a model that contains only the structural equation for the latent variable and your measurement model. Then estimate the parameters for those two components. And then as a next step, you specify the choice model that also includes the latent variable (i.e. the structural equation) but not the measurement models, and you use the results from the first model as parameters for the structural equation, and keep them fixed, and only estimate the others parts.

Stephane