Pooling design matrices for level balance

[feppink]

Hi all,

I have been tasked with analysing responses to a choice experiment with 4 (nominal) attributes with either 2 or 3 levels. Respondents saw 9 choice situations, each of which had 2 non-labeled alternatives and an opt-out. The odd thing, to me, is that 48 D-efficient design matrices were deployed in an effort to improve level balance. Respondents were randomly assigned to one of the CE versions and they are distributed evenly across versions. There are many thousands of respondents.

I have designed and analysed a few choice experiments in the past. With this data set, however, choice models (MNL, MIXL, ...) either do not converge at all or seem to converge but come with singularity warnings or errors that cast doubt on the results. I am familiar with my software (R), tried various approaches/packages to no avail, and have double-checked my data. I am confident the problem is not in these two steps of the analysis.

My question is: can the decision to pool design matrices be causing my estimation problems? I can imagine, for instance, that the 48 designs meet design theory requirements individually, but that collectively they do not. Maybe I'm just unlucky with this data set. I would greatly appreciate all thoughts and comments.

Regards,
Florian

[Michiel Bliemer]

I do not think that pooling data from different designs could cause this issue. Problems with estimation usually only occur if levels of certain attributes are (near-) perfectly correlated. While this could possibly occur within a single design matrix (although rare), it would be very unlikely to occur in all 48 design matrices such that attributes are perfectly correlated.

It is odd that even the MNL model would not converge, given that you have thousands of respondents. Maybe there is very little variation in the data? Or maybe the attributes are all not very relevant? It is difficult to see why it would not converge.

Michiel

[feppink]

Dank je, Michiel.

The responses are quite varied and I get seemingly logical results when I omit all data related to the opt-out alternative. So I suspect the problem comes from my data processing. But that is perhaps a discussion for another forum.

Regards,
Florian