Overdispersion tests from DHARMa and sjstats: conflicting results?

Question

I ran some models for my count data, and did some diagnostics to check for overdispersion.

Here is a dharma graph, which as I understand, indicates no overdispersion.

And this is the result I get when running overdisp(model1)

dispersion ratio = 1.2987
Pearson's Chi-Squared = 496.1125
p-value = 0.0001

Overdispersion detected.

the model looks like this:

model1 <- glmmTMB(species~ var1 + var2 + var3 + var4 + var5 + var6 + var7 + var8 + (1|randomeffect), family = "poisson", data = plants)

Why does it happen? Which method should I trust?

Answer 1

I'm the developer of DHARMa. First of all: note that results are not actually conflicting - a non-significant test doesn't mean that there is no overdispersion, it just means just that the respective test is not sure there is.

Specifically, DHARMa and overdisp or the newer performance::check_overdispersion perform different tests, and what you see is that one test shows that there is evidence for a problem, while the other one doesn't. I have just made a few simulations and can confirm that performance::check_overdispersion is more sensitive in some situations.

Thus, in doubt, I would treat your results as a problem, although I would also note that you will practically always get some significant deviations as soon as you have enough data. A dispersion parameter of 1.3 is relatively small, but given that a correction is easy (moving to nbinom), I don't see why you wouldn't do that.