How to calculate the D-efficiency of experimental designs in conjoint analysis?

Specifically, how do you specify the $X$ and the number of $nBetas$ in this formula:

$$ D_e=\frac{|X'X|^{1/nBetas}}{nSets} $$

This question addresses the calculation of D-efficiency for simple experimental designs, but I don't fully understand how to apply it to conjoint designs (where there are different attributes and levels), both alternative-specific and not.

I would appreciate an example.


For those who want to learn about designing of choice experiments (CEs), I would strongly advise to read the documentation of the NGENE software (https://dl.dropboxusercontent.com/u/9406880/NgeneManual112.pdf), which is probably the best software to design CEs.

To answer your question, as indicated by the formulae, you need 2 pieces of information:
1/ What are the preferences (beta) for the attributes (X)?
2/ What is the design of the CE (i.e., content of the choice tasks)?
The tricky part here is to obtain this a priori knowledge about the betas.
By default, a conservative approach consists in assuming that the betas are null.
In this particular case there would be actually very little difference between a D-Efficient design and an orthogonal design.

You can relax this assumption in different ways:
- Run a pilot study to get a rough idea of what might be people preferences for (X).
- Look at the literature and try to find a comparable study - Unlikely to be a good idea.
- Make some pseudo-informed guesses (e.g., one would expect people preferences for product price to be negative - Therefore you could assume that beta_cost follows a negative log-normal distribution ...).

You obtain information about (X) by designing your experiment (Q.E.D.).

| cite | improve this answer | |
  • 1
    $\begingroup$ Thank you for the comment, but it does not answer the question. NGENE is great, I know. $\endgroup$ – k-zar May 21 '17 at 0:04
  • $\begingroup$ To compute measure of stat efficiency (not only D-Eff) you need an information about the betas (e.g. beta = [0.1 -0.2 0.5 -0.3]) and the design of your experiment (X) - More precisely (X) corresponds to your matrix of model predictors such that number of columns = length of beta vector. Then you simply need to apply the formulae and you will obtain the D-Eff of your design. Does it answer your question or am I missing something? $\endgroup$ – Umka May 24 '17 at 9:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.