Wrong priors and wrong block.. Need help..

[Benjamin]

Hi!

I posted in the other forum what maybe should have been posted here..

We have done a 3000-respondent survey including a DCE.
But now we have discovered that wrong choice-sets have been used ...

So:

1. We had 7 attributes, and runned a eff(mnl, d) with signs as priors (-0,00001 / 0,00000) on 4 of the attributes, the others we did not no sign of.
But in the DCE every binary is opposit. This means that the 4 attributes we put a prior-sign on, was wrong, minus when pluss, and pluss when it should have been minus.
What problem will this do? Totally unusable?


2. We had 4 blocks, 4 choice sets in each block = total 16 choice sets. 3 of the blocks were correct (correct opposit as from question 1), but the last block were totally another block from another model in DCE (really wrong).
What problem will this do?

What would you suggest to do with the data? Garbage?


-Ben

[Michiel Bliemer]

1. See my post on the other forum. A prior of -0.00001 and a prior of 0.00001 are both essentially zero and hence do not differ in efficiency. The only influence that a wrong sign can have is failure in detecting strictly dominant alternatives (in which all attribute levels are better than in the other alternatives). Given that you have 7 attributes, it is unlikely you will have problems with dominancy, so I would not worry about it.

2. I do not understand the issue. Why is the 4th block wrong? Does it have silly choice tasks? Or dominant alternatives? As long as there are no strictly dominant alternatives in the data, choice tasks cannot be wrong. Choice tasks with strictly dominant alternatives should
be removed from the dataset.

You should be able to use the data as you normally would, there should not be any problems with the data.

Michiel

[Benjamin]

Nice!

Thanks for replying!

The choice-sets/scenarios generated from Ngene is correct in the 4th block, however we mixed up and used some old generated scenarios, which makes the 4th block just totally different...

So the question is: when you have 16 scenarios, divided in 4 blocks, 3 of the blocks (12 scenarios) are right, but the last is wrong (missing). Does this make the first 3 blocks unusable? Do you need all the blocks to be able to use them? Or can you use some separated from the other?

[Michiel Bliemer]

It is not a problem if one block is missing, although you will have less variation in the data and hence leads again to less information for parameter estimation (especially if you have many parameters to estimate; if you are estimating 7, then I think 12 choice tasks will provide sufficient information).

[Benjamin]

Thanks again!

So then another problem appears.
When we try to run regressions with only one block, then half of the parameters gets omitted because of collinearity.. Is this because we have to little variation with only 4 scenarios?

[Benjamin]

One more question.

We have 4 blocks (4 scenarios in each) generated in Ngene. We mixed up, and only used 3 of the blokcs, the last block (Block D) we mixed with an earliger generated outcome (however same syntax and design), just another "run".

Can we include the last block? And use all the results? Or do we have to drop the last block? (If you get the question)

[Michiel Bliemer]

1. I am not sure why you would want to perform regression on one block. First of all, the data should never be analysed per block, but all blocks should be pooled together to get enough variation in the data as otherwise you lose too much information. You can likely afford losing one block, so you have to pool the three remaining blocks. Further, the data is generated for estimating logit models (which are nonlinear and multicollinearity is unlikely to be an issue), so I am not sure why you would want to run (linear?) regression.

2. Your last question confuses me as earlier you mentioned the 4th block is completely wrong and should be considered missing, but now you are asking if you can include the last block? I do not think I can assist any further since I do not know your data. I suggest you consult an econometrician that can have a look at your data.