Experiment Design Checklist

Atal Gawande’s The Checklist Manifesto is a compelling account of how the simple checklist can scaffold expertise and teamwork in complex domains. Good checklists, he says, nudge our memories, prompting us to do what we already know we should do, but risk missing.

Good checklists are also short. 5-9 “killer” items, says Gawande.

So, I got to thinking, what would be on my checklist for experiment design? Additional suggestions welcome, by email or tweet, but this is what I got so far:

1. Did you discuss authorship with the research team?

Not sure who is an author: ICMJE Authorship definition. Bonus points: start filling in the Contributor Roles Taxonomy

2. Is your study adequately powered?

Most studies are underpowered, don’t be one of them. Useful: Understanding mixed effects models through data simulation. What is your recruitment plan? How will participants be enticed to take part? How will they be rewarded for taking part? Will they be brought into the research process in any way?

3. Should you include a manipulation check?

Probably you should. How will you know that participants were affected the way you think they should have been by your intervention?

4. Does your experiment produce data that you can analyse?

This means you should do a full dry run through the experiment, noting how long it takes, and checking that the measures you need are recorded. Bonus points: simulate participant data (some platforms like Qualtrics will do this for you) and check your proposed analysis runs on the output. Bonus points: Exercises for Lab Groups to Prevent Research Mistakes

5. How will you judge the size of any effect?

even a statistically significant result might be meaningless. How will you calculate an effect size, and how will you gauge whether it is important? Bonus points: maximal positive controls

6. Will you be able to interpret all possible results which aren’t in line with your predictions?

Maybe you made predictions. What will it mean if you get a null result? Or an intermediate result? Or any other unexpected outcomes.

Aim for your final data to be FAIR - Findable, Accessible, Interoperabe and Reusable

8. Have you checked prior work on this topic?

How systematic was that review of the previous literature?

9. How will the final result be criticised?

Imagine what your strongest critic will say when presented with your final results. Plan your defence. You might want to consider List #2


List #2 : Common criticisms

Standard flaws, and standard criticisms that you might hear about your result

-

placebo effect

Participants responding to the situation, not your intended treatment. Address with an adequate control

demand effect

Participants responding to what they think you want. Address by keeping partipants blind to which condition they are in. Advanced: your experiment will contain an incentive structure which participants respond to. It may be that errors are more costly than successes (so you are rewarding conservative behaviour), or it may be something as simple as that you can get through the experiment more quickly if you take a certain choice (so you are rewarding fast/inaccurate responding). You should know what the incentive structure of your experiment is. The incentive structure of your experiment may be driving behaviour, as much as any deeper psychological desires or biases. Even if this isn’t the case, how will you convince a critic that your experiment reveals something about human psychology beyond “participants do what they are rewarded for?”

experimenter bias

may not be deliberate. Could result from unintentional exploitation of researcher degrees of freedom, inadvertant signalling to participants. Partially address with preregistration.

selection and survivorship bias

Are the results distorted by how you recruited, or how participants differentially dropped out during the experiment? Partially address with (truly) random assignment.

regression to the mean

Particularly a problem with test-retest designs

common method variance

Particularly a problem if both your dependent and independent measures are of the same type (e.g. survey items)

so what?

“This was obvious”, says everyone after you’ve done the work to show it happens

false positive

You got lucky, they say. Address by replicating

confound

Something else you didn’t control for produced the effect


You may also enjoy

Created: 2021-07-10
Latest update: 2021-08-30
Repo (contains citation widget): github.com/tomstafford/psy-checklist
DOI

Creative Commons License
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.