Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

Conjoint design for Pharma research study

For the SCE conjoint that we were discussing I need your guidance on designing the conjoint.

We have 5 attributes with below details:
3 attributes with 3 levels
1 attribute with 4 levels and
1 attribute with 2 levels

Total possible combination 216

Sample size 150

We'll provide 1-2 option to choose from 14 available drug profiles.

We are planning to show each level at least 3 times to 1 respondent and the total 12 cards with 1 product in each.

We need both overall and individual level utility for this study.

Please let me know if 12 cards would be enough to get robust utilities on both overall and individual level.  I really appreciate all your help on this so far.

Thanks & regards,
asked Jul 28, 2017 by kshitijkumarsingh (160 points)

1 Answer

0 votes
Most of this makes sense to me, except "We will provide the respondents up to 2 options to choose from 14 available drug profiles."  I don't know what that sentence means.  

So your attributes and levels describe patient variables, right?

And you plan to show 12 patient profiles per respondent?  So far so good.

And there are 14 levels in your dependent variable?  If so that spreads the sample quite thin, because you'll be estimating 11 x 13 =143 parameters on just 12 x 150 = 1,800 observations.  The standard reference for sample size in logit models (Peduzzi et al 1996) recommends a minimum number of observations for models like this is 10 times the number of parameters divided by the (minimum) likelihood of choice.   Likelihood of choice is, AT BEST 100% divided by the number of levels of the dependent variable, so it's 1/14 or 0.0714.  Plugging that into the equation would suggest that you need 10 x 143 / .0714 = 20,020 observations, not the 1,800 that you have.  With such a small number of observations relative to what you need, you may have utilities that do all sorts of strange things.  If you can cut down on the number of levels in your DV (perhaps by collapsing across categories of drugs into a smaller number of levels) that could help quite a bit.  

That's just for an aggregate unconditional logit model.  Because of the sparse data problem these models are usually NOT run at the individual level and you have more reason than most not to do so here.
answered Jul 28, 2017 by Keith Chrzan Platinum Sawtooth Software, Inc. (92,075 points)