Include random effects when plotting residuals vs. predicted? about dharma HOT 5 CLOSED

That is something I have been wondering quite some time now. But the answer still leaves me a bit puzzled. Which is now the type of plot that I should use to examine the fit of my model?

If I use the DHARMa standard, I will miss out on plotting the RE completely. Why is that the standard, when the RE is an essential part of my model?

from dharma.

Thanks a lot Florian!
I consider my self a begginer, but I also can't get this conditional/unconditional on REs issue, and would appreciate more explicit info in the vignette, both about reasons for use and the use itself. DHARMa is the first time I have ever read about it.

Also, I found some models show exactly the same output plots (QQ and res vs. pred) conditional and unconditional on REs, while other models show very different ones (e.g. one good and one bad)? Why is that?

Best,
Diego.

from dharma.

Sorry, I realise I also forgot to answer to @AmeBol, so this is to both

a) @nabaesjodar: for a perfect model, both conditional / unconditional plots should look fine, so in doubt just check both.

b) I have chosen the unconditional as a default for DHARMa, because this checks the RE + the final distribution at the same time (so the entire model structure is re-simulated). If you simulate conditionally, you check the final residuals, conditional on the fitted REs, so you check e.g. only that the residual distribution is Poisson, conditional on the values of the fitted REs.

c) However, as noted in the help about dispersion tests, in particular dispersion tests are less powerful with unconditional simulations. So I would suggest to do dispersion tests based on conditional simulations (if you regression package permits this)

d) @AmeBol, you can of course additionally plot the RE distribution, as advised in many textbooks, and test for normality (take out REs, make a shapiro.test on them). I have no special functions for this in DHARMa because this is a test not on the residuals, but on the fitted values, so it's completely independent of the quantile residuals that are calculated in DHARMa. It may also be useful to look at REs against predictors or space, as they can absorb misfit (this is again one of the reasons, by the way, to use unconditional simulations as a default, because then you see this misfit in the DHARMa residuals).

from dharma.

I will add a few clarifications about this in the vignette.

from dharma.

Looking at this, this is all explained in the vignette already, so I'll close this!

from dharma.