4

I conducted an experiment with a system that allows to perform a task on a test-object using several methods. The quality of how the task was performed is measured by an error.

I measured the errors of 10 methods each conducted by 5 users on 2 test-objects whereas each user conducted 6 attempts per method and test-objects.

I would like to model this setup using a linear mixed effects model. In particular, I would like to evaluate whether there are (significant) differences

  • between users (intrauser reliability)
  • within users (interuser reliability)
  • across test-objects (across test-object reliability)
  • between methods

I think I have

  • fixed effects: method
  • random effects: user, test-object

Accordingly, the model would be

error ~ methods + (1|user) + (1|test-object)

Is that correct?

Furthermore, how can I know if I have a nested structure? Accordingly I would require a term like (1|User/Attempt)?

Thanks

user137589
  • 123
  • 4
  • 1
    If I understand correctly, there are only 2 levels of the `test-object` variable? If so then it is questionable to use `test-object` as a random effect - just retain it as a fixed effect – Robert Long Jul 08 '18 at 17:04
  • Yes that is correct, I can change it to a fixed effect. The question if I need a term like `(1|User/Attempt)` remains though? – user137589 Jul 09 '18 at 14:54

1 Answers1

2

From your description, to fit a linear mixed effects model to your data you could use:

error ~ methods + (1|user) + (1|test-object)

Nesting is a property of the experimental design itself. From your description, a test-object does not belong to a particular user, and neither does a particular user belong to a particular test-object in the sense that a child would belong to, say, a class in a school, in a nested design. So you do not have nested data.

Therefore you do not need (1|user/test-object) - though it is worth noting that the results should be the same, provided that the variables are encoded appropriately. See the accepted answer here for more details.

Finally, with only 2 levels of test-object, it doesn't really meet the requirements for random effects, so a better model may be:

error ~ methods * test-object + (1|user)
Robert Long
  • 53,316
  • 10
  • 84
  • 148