Support for glmmTMB in DHARMa

[florianhartig]

A user requested support for https://github.com/glmmTMB/glmmTMB

Status of this request (updated)

• glmmTMB is supported by DHARMA since https://github.com/florianhartig/DHARMa/releases/tag/v0.1.6.2.

• glmmTMB is fully supported (fixing the re.form problem and some other things, see below) since DHARMa 0.2.7 and glmmTMB 1.0

• A remaining limitation is the lack of support in glmmTMB to condition simulations on fitted REs, see Implement re.form option in predict() also for simulate() and possibly also for residuals() glmmTMB/glmmTMB#888

A simple example is

m <- glmmTMB(count~ mined, family=poisson, data=Salamanders)
summary(m)

res = simulateResiduals(m)
plot(res, asFactor = T)

More examples in the vignette or in https://github.com/florianhartig/DHARMa/tree/master/Code/DHARMaPackageSupport/glmmTMB

[florianhartig]

glmmTMB has now implemented a simulate function, and I have added this to DHARMa. You can test this by installing the development version of DHARMa (see https://github.com/florianhartig/DHARMa).

Note: I had problems with the CRAN version of glmmTMB (crashed with some models). I would therefore recommend installing the development version of glmmTMB before installing DHARMa via

devtools::install_github("glmmTMB/glmmTMB/glmmTMB")

Main limitation is currently that glmmTMB doesn't support the reform argument in either predict or simulate. Predict() is conditional on all random effects, corresponding to lme4 re.form = NULL. Simulate() is unconditional, i.e. all random effects will be re-simulated, corresponding to lme4 re.form = 0. That means that all predictions and simulations are conditional on REs, which can sometimes create a positive correlation between res and predicted , see #43

Other than that, it seems to me that the glmmTMB interacts fine with DHARMa. A simple example is

m <- glmmTMB(count~ mined, family=poisson, data=Salamanders)
summary(m)

res = simulateResiduals(m)
plot(res, asFactor = T)

More examples in the vignette

[florianhartig]

Merged into master with ed836dd

[florianhartig]

Things that are solved

• works with the development version of glmmTMB, but not yet with the CRAN version [SOLVED new CRAN version glmmTMB 0.2.1 works fully with DHARMa]

• pearson residuals don't work with zi terms (not implemented), which limits some of the overdispersion tests - solved, simply not fix this (probably not a real problem, see glmmTMB residual checks with DHARMa  glmmTMB/glmmTMB#313)

Things that are still unsolved

• glmmTMB doesn't support the reform argument in either predict or simulate. Predict() is conditional on all random effects, corresponding to lme4 re.form = NULL. Simulate() is unconditional, i.e. all random effects will be re-simulated, corresponding to lme4 re.form = 0. That means that all predictions and simulations are conditional on REs, which can sometimes create a positive correlation between res and predicted , see Include random effects when plotting residuals vs. predicted? #43

• glmmTMB residual function fails with factor response (already fixed in glmmTMB develoment branch)

• Check / created tests for structured RE behavior

[florianhartig]

added instructions about how to calculate unconditional predictions by hand

[mebrooks]

Any chance you could test the re.form in predict sometime soon? Will this resolve the warning from simulateResiduals which refers to this issue?

[florianhartig]

Hi Mollie, yes, I have tested this in https://github.com/florianhartig/DHARMa/tree/2.4.2-glmmTMBfix and it works like a charm, at least as far as the integration in DHARMa and the problem I had with #16 is concerned. As soon as you push this on CRAN, I will do an update to DHARMa!

[florianhartig]

Notes to users: if you want to make use of these recent changes

• install glmmTMB from GitHub
• install https://github.com/florianhartig/DHARMa/tree/2.4.2-glmmTMBfix from GitHub (this is the DHARMa branch which fits to the updated glmmTMB)

[florianhartig]

Limitations (old)

This problem was fixed with 32cbadc and glmmTMB 1.0

====

The main limitation is currently that glmmTMB doesn't support the reform argument in either predict or simulate. Predict() is conditional on all random effects, corresponding to lme4 re.form = NULL. Simulate() is unconditional, i.e. all random effects will be re-simulated, corresponding to lme4 re.form = 0. That means that all predictions and simulations are conditional on REs, which can sometimes create a positive correlation between res and predicted , see #43

NOTES TO USERS (Dez 2019): this issue is fixed in glmmTMB, this fix is not yet on CRAN. You can

• install glmmTMB from GitHub
• install https://github.com/florianhartig/DHARMa/tree/2.4.2-glmmTMBfix from GitHub (this is the DHARMa branch which fits to the updated glmmTMB)

Alternative way to bypass this problem: calculate unconditional predictions by hand - follow this example

testData = createData(sampleSize = 200, randomEffectVariance = 2, family = poisson())

m1 <- glmmTMB(observedResponse ~ Environment1 + (1|group), 
                     family = "poisson", data = testData)

summary(m1)

res <- simulateResiduals(fittedModel = m1)

# this plot shows the residual pattern, despite the model being correct
plot(res)

# recalculate predictions by hand - see help ?predict.glmmTMB
# To compute population-level predictions for a given grouping variable (i.e., setting all random effects for that grouping variable to zero), set the group value to NA.
newdata=testData
newdata$group = NA
pred = predict(m1, newdata = newdata)

# Now perform the plot
plotResiduals(pred, res$scaledResiduals)

[florianhartig]

Since DHARMa 0.2.7 and glmmTMB 1.0, it seems all problems are solved, and I will close this now!

[mebrooks]

Hurray!

[florianhartig]

What remains open is to ability to condition on REs when simulating, see glmmTMB/glmmTMB#888