Skip to content
Prev 19262 / 20628 Next

Model specification/family for a continuous/proportional response with many zeros

Hi Thierry,

Thank you for your advice and speedy response.

Most of the data is closer to the lower bound (0). e.g. the mean time for
group A in zone A = 15.1 seconds and group A in zone B = 3.8 seconds.
However there are a very small number of outliers near the upper bound, the
largest being 294 out of the 300 seconds (see the attached file if you want
to look at the data).

I have taken a stab at running a Zero-inflated Beta GLMM using glmmTMB in R
like so:

betta_mod <- glmmTMB(prop_time ~ group*zone + (1|id),
                             family = beta_family(),
                             ziformula=~1,
                             data = glmm_zone_data)

summary(beta_mod)

*Family: beta  ( logit )*






*Formula:          prop_time ~ group * zone + (1 | id)Zero inflation:
      ~1Data: glmm_zone_data     AIC      BIC   logLik deviance df.resid
-763.6   -736.3    388.8   -777.6      359Random effects:Conditional
model: Groups Name        Variance  Std.Dev. id     (Intercept) 2.386e-09
4.885e-05Number of obs: 366, groups:  id, 14Overdispersion parameter for
beta family (): 13.1Conditional model:                  Estimate Std. Error
z value Pr(>|z|)    (Intercept)        -2.7685     0.1031 -26.844  < 2e-16
***groupB             -0.4455     0.1498  -2.975 0.002932 **zonezone_B
    -0.4179     0.1524  -2.741 0.006124 **groupB:zonezone_B   0.8443
0.2190   3.855 0.000116 ***---Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*?
0.05 ?.? 0.1 ? ? 1Zero-inflation model:            Estimate Std. Error z
value Pr(>|z|)    (Intercept)  -1.1804     0.1233  -9.575   <2e-16
***---Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1*

Does this look like the correct way of specifying the model? I am a little
confused about specifying and interpreting the zero-inflation component - I
have only just begun reading about this.

I noticed that the dispersion parameter is quite high at 13.1. I'm not sure
if this matters for beta models?. I tried running DHARMa simulateResiduals
on the model output and got significant deviations in the dispersion
(<2.2e-16) and KS tests. e.g.

DHARMa::testDispersion(beta_mod)

*DHARMa nonparametric dispersion test via sd of residuals fitted vs.
simulated*

*data:  simulationOutput*
*ratioObsSim = 1.3612, p-value < 2.2e-16*
*alternative hypothesis: two.sided*



Many thanks,
Mike

On Mon, 17 May 2021 at 13:22, Thierry Onkelinx <thierry.onkelinx at inbo.be>
wrote: