Hi Michiel,
I'm creating an explicit and implicit partial profiles design for the first time and have a couple of questions. Of the ten attributes in my design, six are dummy-coded side effects (presence/absence) attributes, the rest are effectiveness and mode of administration attributes.
Implicit Design: I generated a large design in R. I filtered for rows where exactly three out of six side effect attribute pairs had a value of (0,0), and the other three pairs varied ((0,1) or (1, 0)). I then randomly selected 5,000 rows and then used it in ngene.
Explicit Design: Using the same large design, I filtered for rows where exactly three out of six side effect attribute pairs showed overlap (values of (0,0) or (1,1)), with the other three pairs varying. I then randomly selected 5,000 rows from this filtered set and used it in ngene.
My questions are:
1) Is it okay to leave attributes 1-4 'as normal' and only apply the filtering to a subset -the 6 side effects attributes? I'm only asking because, in the manual and in your excel candidate set generator file, the omitted attributes or the attributes with overlap are applied to the full set of attributes.
2) I have priors from a previous study for attributes 1-4, but not for the side effect attributes. Is it better to use a mixture of real priors and directional priors?
Thanks,
Tara
explicit and implicit partial profiles
Moderators: Andrew Collins, Michiel Bliemer, johnr
-
Michiel Bliemer
- Posts: 2039
- Joined: Tue Mar 31, 2009 4:13 pm
Re: explicit and implicit partial profiles
1. Yes certainly, that is up to you. The most common is that you create overlap across all attributes, but this is not necessary.
2. Mixing informative priors and uninformative priors is usually not ideal because the choice probabilities affect the efficiency and your uninformative priors would not affect the probabilities whereas the informative priors would.I would probably generate an efficient design using only uninformative priors (zero indicating preference order only), then conduct a pilot study, and then use informative priors for all attributes in the main study. But if you are not going to do a pilot study, then I guess mixing the types of priors would be a second-best option.
Michiel
2. Mixing informative priors and uninformative priors is usually not ideal because the choice probabilities affect the efficiency and your uninformative priors would not affect the probabilities whereas the informative priors would.I would probably generate an efficient design using only uninformative priors (zero indicating preference order only), then conduct a pilot study, and then use informative priors for all attributes in the main study. But if you are not going to do a pilot study, then I guess mixing the types of priors would be a second-best option.
Michiel