choice-metrics.com

Hi all,

I am new to experimental design, so am still coming to grips with what’s required.

My question relates to the best method to complete segmented analysis of consumer groups. I believe there to be two major methods as follows:

1.Two model analysis – simply divide the market into segmented sections ( ie 2 groups for gender male/female) and then evaluate a MNL model for each section. The Beta’s etc and their relative influence for each attribute ( or MRS) can then be compared.

2.Introduce the desired variable as an interaction term ( ie gender as a dummy variable)- then by analysing the interaction term, we can see if the variable is significant.

Is the above correct and when would you opt for one approach over the other??

I am keen to hear everyone’s thoughts.

Thanks for all the assistance.

James

The use of respondent characteristics when generating an efficient experimental design is rare, as in many cases it is difficult to obtain priors and more complex to create different surveys for different segments.

Clearly, in estimation you typically include such variables into your model. There are many ways to include them, and you identified two.

1. Estimate separate models and compare ratios of parameters
2. Estimate a single nested model and compare scale parameters of each nest
3. Estimate a latent class models in which the segment variables enter the class probability model
4. Estimate a single model where segment variables do not appear in all alternatives (suitable for labelled experiments only)
5. Estimate a single model where segment variables appear as interaction effects

Ngene cannot generate efficient designs for nested models (option 2) or latent class models (option 3).

Hi David

I would suggest it depends on sample size. Estimating a single model and testing for differences (via interaction effects) will tend to produce more robust results as it uses the entire sample for the model. If you have a small number of respondents available in per segment, then this is perhaps the way to go. It also allows for tests of scale differences, however as you are interested in MRS, this is less of an issue. Separate models work well if you have large samples to play with for both models.

As Michiel suggests, using a latent class model is a different approach you could explore. It allows for discrete distributions (think separate MNL models), where you can estimate a class assignment model. In the class assignment model, you can use socio-demographics to assign individuals to different classes. From there, you can obtain individual specific estimates which vary by socio-demographic segment.

John