Sample size for blocking

[jgrant8]

Hello,

I am conducting an unlabelled DCE with health workers in Ghana that has 2 alternatives and a current policy opt-out, 5 attributes (3 at 2 levels, 2 at 3 levels), and 12 choice tasks. We are including all health workers in the district where the intervention is taking place, which will be around 100 participants (minimum this will be 80, maximum around 120). Because the intervention is only in this district, we don't have any opportunities to expand the sample size.

I am not sure if I should be considering blocking participants in the DCE into two blocks of 50 (which in a worst case scenario could end up as 2 blocks of 40)? I understand that blocking can be used to create larger/more complex designs that explore a wider range of choices, but I am not sure we have the numbers to do this. Using the rule of thumb, 42 participants are needed for the sample (/in each block), but I am cautious about blocking and potentially getting so close to this number, as this assumes 'perfect' participants.

Would you be able to share any guidance on this issue?

Thank you,
Jane

[Michiel Bliemer]

Your choice model seems to have at most 8 parameters to estimate (unless you are also considering interaction effects). With 3 alternatives (including status quo) this means that you need at least 8/(3-1) = 4 choice tasks, but it would be recommended to use at least 3*4 = 12 choice tasks. So you could estimate your model using just 12 choice tasks and show them to each respondent without blocking.

If you can give them more than 12 choice tasks, then of course you would capture more information.

You seem to suggest using 100 choice tasks in 2 blocks of 50, so each respondent is given 50 choice tasks. That would be a lot. Since there is no reason to use blocking, you could also consider a design with 50 choice tasks.

If you do use blocking, the main thing to consider is that you have some balance between completions in each block, so about half of respondents completes block 1 and the other half block 2.

Michiel

[jgrant8]

Thank you very much for your response with these considerations.

I realised my initial question wasn't written very clearly. From other DCEs in this setting and a pre-test, we have found that 12 choice tasks is the upper limit of what we would want to present to each participant. So with this in mind, and our sample size limitations, I am wondering about the choice between:
- having all 80-120 participants complete the 12 choice tasks
- having 24 choice tasks and blocking participants into 2 blocks so that each block completes 12 choice tasks. My understanding is that the benefit of this would be capturing more information as you say, but we would only have 40-60 participants in each of the two blocks which seems a bit low.

Ideally, I would rather just use 12 choice tasks and show them to each participant without blocking.


Best,
Jane

[Michiel Bliemer]

You should be able to estimate a model (with main effects only) with 12 rows in your design. However, you may not be able to estimate interaction effects if you only have 12 choice tasks in your design. Using 24 rows will create more variation in your data, giving you more flexibility in defining your utility functions after data collection. It does not matter if each block has only 40-60 respondents, that would still be plenty.

Michiel

[jgrant8]

Okay thanks very much.

Best,
Jane