Best CBC design for HCP study with 3 patient types

Question

Hopefully not too specific a question.  Am curious about the choice in design for a physician study with a new drug in clinical trials.  Estimated at ~10 attributes so there is a lot for physicians to process, 3 very similar patient diagnoses (but desire to understand differences in importance between them), and with all physician studies we have limited sample of about 250.

We ultimately decided to recommend Covariate CBC but curious if you would agree with the choice and/or if any of the pro/con considerations below do not really hold?  We debated 4 different designs.

Covariate CBC – show each respondent only 1 patient type diagnosis.  Run utilities all together with patient type as a covariate to take advantage of the larger sample but allow for and test differences between patient types.  Also limits consideration of one patient type per respondent since there are so many attributes.

Separate CBC – do as above but generate utilities for each patient type separately, so only n=100 for each.  Maximizes ability to allow patient types to be different but limits sample size.

CBC w/ Patient Type – Patient type diagnosis becomes an 11th attribute.  Didn’t think there was enough sample to run as MBC (or is there if there is only 1 patient attribute (diagnosis)?).  But as an 11th attribute I don’t like the idea of mixing in a patient variable as it then adds an even bigger burden on physicians to randomly change the patient frame each exercise.  But more importantly, patient type may interact with the importance of all drug attributes and that would be hard to account for.

Multi-Choice CBC Grid – Yields the most data since respondents select a best profile for each patient type in each exercise.  Would be a great option except concern for the number of attributes (10) and cognitive burden with considering each patient type on each exercise and also might drive too much correlation between choices (tendency to select the same profile for all 3 patient types).

Answer 1

Hi, Steve.

Note that your 4 options aren't mutually exclusive - for example, with the Multi-choice grid, you could choose to analyze it as 3 separate CBCs or as a single CBC with covariates.  The outlier here is your third option and like you I would NOT prefer to run the model with patient type as an attribute.  

I have done very many studies where we have the multi-choice grid and if the grid doesn't get too crazy (like if there are 10 patient types instead of the more manageable 3) the it works very nicely.

A white paper we've prepared on the subject of pharmaceutical choice experiments can be found here:  https://www.sawtoothsoftware.com/download/techpap/Choice_Experiments_for_Physicians_(Chrzan).pdf

To answer your specific question about analyzing as 3 separate CBCs or as one CBC with a 3-level covariate, a couple of years ago I fielded a study where we tested both options.  I had each respondent answer each CBC grid response question that included choices for each of 6 situations.  It turned out that whether I modeled this as 6 separate CBCs or as a single CBC with a 6-level covariate didn't matter much at all.  Both methods returned very similar utilities and simulations.  It just doesn't seem to matter much which route you go.

Comments

Keith, thanks for the insights.  A quick follow-up.  Is there actually enough base size to make MBC an option here, with the 1 patient type attribute representing that dimension?

And you mention not wanting to run patient type as just another attribute - is that because of the cognitive burden of shifting patient types to the physician and/or the troublesome interaction effects and/or something else?

And yes, that choice experiments for physicians paper comes in very handy on a regular basis!

Thanks!!!

---

Steve,

I'm confused how MBC relates to the discussion.  

For not wanting patient type, I was thinking of the issue of the interactions, yes.

---

In MBC, couldn't the patient type be set up in the prologue and the drug profiles that they choose from at the bottom, like shown in Figure 1.9 of  MBC (Menu-Based Choice) Documentation (2013)?  This may then just amount to running an interaction effect crossing patient type against all the drug attributes, which then would require larger sample sizes, more than 300, but not sure.

---

The MBC example works when you have a labeled CBC design and its analysis would produce an alternative-specific constant for each of the drug brands.  If you have a generic CBC design (instead of a labeled one) then I would not go the MBC route.