DHARMa's temporal autocorrelation plot differs from acf plot about dharma HOT 6 CLOSED

Hello,

this is expected (however, admittedly, not ideal behaviour). The reason behind this is the following:

• DHARMa residuals are not expected to change when including REs, in particular REs with covariance structures, because glmmTMB does not (yet) allow to condition simulations and predictions on the fitted REs. Basically DHARMa residuals are calculated around your fixed effect structure (though they re-simulate the REs). See #16 and just opened issue glmmTMB/glmmTMB#888.

• glmmTMB residuals are calculated conditional on fitted REs, so in this case you see the residual correlation minus the RE structure.

How to fix this:

First of all, the fact that your autocorrelation structure disappears with glmmTMB residuals (either unstructured or with a structured RE) means that the REs are temporally autocorrelated, so you should fit an AR1 structure.

Ifs you test a glmmTMB model with covariance in the REs (such as AR1) in DHARMa, you can currently either ignore the autocorrelation, or use the rotation argument, see #364. I hope that the glmmTMB developers will include an option to condition simulations on the fitted REs in the future, which should produce similar residuals as you produce with act(residuals(model)).

from dharma.

Hi Florian,

Thanks for the quick and helpful response! I didn't realize that the DHARMa residuals for glmmTMB are not conditioned on the RE structure, but then it makes complete sense that the autocorrelation would still be present.

I'm fine with ignoring the autocorrelation in DHARMa until glmmTMB #888 is addressed.

But I did try the rotation argument and am still not getting the same results as acf. That's also to be expected, right? Because as you said in #301, "The problem, however, is that this will likely only work perfectly for symmetric linear homogenous cdfc, such as the multivariate normal." So, hypothetically, since I'm seeing a decrease in the autocorrelation structure by including the rotation argument, that's indicative that the AR1 structure is effective, even though the autocorrelation isn't completely solved by DHARMa's calculations?

This is mostly out of curiosity at this point, because as I said, I'm fine with continuing to use glmmTMB's residuals and acf to detect autocorrelation structure.

Thanks again!
Michelle

from dharma.

Hello Michelle,

you tried this on your AR1 model, right? My comments referred to real GLMMs - in your case, you essentially fit an LMM, so the rotation should be exact, provided your nSim is large enough.

By exact, I mean that the autocorrelation part of the AR1 model is correctly removed, and if you see autocorrelation remaining in the DHARMa residuals with rotation, you should probably interpret this as some kind of signal remaining.

There are still differences between the residuals, because DHARMa does not condition on your other REs, so the residuals will differ in their exact numeric values. It could be that some of the other REs absorb autocorrelation. In particular, in your model, you have both

(1 | date)
ar1(date + 0 | id)

The (1 | date) will likely absorb autocorrelation that does not fit the form of the AR1, which would show up in DHARMa but not in the glmmTMB plots. You could test this by calculating the acdf acf on the fitted REs of (1 | date)

Cheers,
Florian

from dharma.

from dharma.

typo, I meant acf

from dharma.

OK, I will consider this closed for the moment, in case you have further questions feel free to re-open the issue!

from dharma.