Residuals vs. Predictor: quantile deviation DHARMa

Question

I have run a Gamma model where I try to predict the bioturbation potential by some sediment predictors. We sampled multiple stations over 10 years. Therefore I included station and year as random factors and added an autocorrelation term.

model <- glmmTMB(BPc ~ mudfraction + medsand + X250.500um + shellfraction + (1|StationCode) + (1|Year) + ar1(Year + 0|StationCode),family = Gamma(link = "log"), data = GLMMdataset) 

I get following summary output :

     Family: Gamma  ( log )
Formula:          BPc ~ mudfraction + medsand + X250.500um + shellfraction + (1 |      StationCode) + (1 | Year) + ar1(Year + 0 | StationCode)
Data: GLMMdataset

     AIC      BIC   logLik deviance df.resid 
  6398.5   6439.7  -3189.3   6378.5      444 

Random effects:

Conditional model:
 Groups        Name        Variance Std.Dev. Corr      
 StationCode   (Intercept) 0.05146  0.2269             
 Year          (Intercept) 0.04565  0.2137             
 StationCode.1 Year2010    0.16480  0.4060   0.35 (ar1)
Number of obs: 454, groups:  StationCode, 93; Year, 9

Dispersion estimate for Gamma family (sigma^2): 0.259 

Conditional model:
                Estimate Std. Error z value Pr(>|z|)    
(Intercept)   12.8665791  0.5721050  22.490  < 2e-16 ***
mudfraction   -0.0543623  0.0129689  -4.192 2.77e-05 ***
medsand       -0.0102631  0.0008794 -11.670  < 2e-16 ***
X250.500um    -0.0464859  0.0047010  -9.889  < 2e-16 ***
shellfraction  0.0092637  0.0027312   3.392 0.000694 ***

And residual diagnostic plots made with the DHARMa package:

I would conclude these plots look OK, but when I look at residual vs predictor I get a deviation for shellfraction at the 0.75 quantile.

How can I address this deviation for shellfraction in the residual plot so my residuals are evenly distributed?

Answer 1

The deviation is so minimal that I don't see any need to address it!

That being said, what the plot shows is that the residual dispersion is larger for low values of shell fraction, so if you want to address that (again, I see no need to do so), you should at dispformula ~ shellfraction to the glmmTMB model.