Excluding implausible combinations will introduce some correlation into your experimental design (i.e. it won't be perfectly orthogonal), but if you minimize these correlations by pre-specifying your model at the design state, the results will be perfectly valid. Even though correlation between attributes limits statistical efficiency, failing to exclude implausible combinations or dominated alternatives also has an adverse effect on efficiency.

Hi Chris,

When you say pre-specifying my model at the design stage, do you mean I need to exclude levels of attributes that has little effect on correlations? May problem actually is that I have 11 attributes and 4 levels each attribute. so my showcards will be plenty and it is not viable for the attention span of my respondents if they will be asked to rank it. But there are attributes that has similar levels (at least two) so maybe i'll be expecting a fewer (but still plenty for a short attention span respondents) of showcards. I am considering of limiting my attributes based on its relevance in the competition. I am planning to exclude those with two or more similar levels. Maybe I can have fewer attributes but has important marketing implications. Is this what you mean by pre-specifying my model at the design stage?

Thanks. Hope to get your feedback.

Hi Seda,

Thanks for sharing your thoughts on my query. That's really my problem, too many attributes. I am considering right now dropping those that are not really a concern in the competition between players.

Hi Seda,

Thank you for the kind offer, yes I will send you the product features. Just send me your email add to : [email protected].

Thanks.

Hi,

As follow-up to your second question, if you are not using an orthogonal design (i.e. one that ensures there is no correlation between attributes, so that all effects can be estimated independently of each other), you need to specify what effects you want to estimate in your model, including interaction terms. The design algorithm (such as Kuhfeld's SAS macros, or AlgDesign in R) will generate a design that minimizes correlation between the effects you specified, but allow correlation between the effects you didn't specify as relevant. You are not necessarily excluding attributes, you're just making sure that the attribute levels you present don't move together. To illustrate, if price and quality, foe example, always move in the same direction in your choice sets, you won't be able estimate the effect of one independently of the other. By specifying those as effects of interest in your model, the design algorithm will ensure that their levels are not correlated, although they may be correlated with a third attribute that you didn't specify as relevant (such as the interaction between price and quality, for example).

As Seda says, you should be very sure that an attribute isn't relevant before you decide to exclude it.

Chris.