Predictions from Poisson GLMM (lme4) lower compared to GLM

Question

I am modelling visitor counts to a sample of sites in a forest in order to predict the number of visitors to the rest of the forest.

My predictor variables are time of day (categorical), day of week (categorical), distance of site from nearest access point (m) and number of households around access points (in three different buffer distances i.e. 'bands').

I am fitting a GLMM with Poisson error in lme4. I am using site as a random effect because sites were visited more than once (~ 4 times).

When I use the Predict function in lme4 I have to set re.form = NA because I am predicting for out of sample sites, thus the new data points do not have a site ID. I read that this is setting random effects to zero and predicting only at the population level.

My question is this:

I have found that a GLM (i.e. site is not included in the model) results in higher predictions (2 fold) (mean predictions from GLM = 0.2 visits cf GLMM = 0.1)

Why would this be? The coefficients for the fixed effects in the GLMM are not hugely different from the coefficients in the GLM.

The model outputs are as follows (sorry I can't provide data to reproduce these models):

GLM

Call:
glm(formula = dog.walkers.count ~ day5code + dog.time + wght_dist + 
    band1to4 + band5to9 + band10to11, family = "poisson", data = dog.data2)

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-1.9657  -0.6832  -0.4969  -0.3070   6.4859  

Coefficients:
                 Estimate Std. Error z value Pr(>|z|)    
(Intercept)     -1.914835   0.084291 -22.717  < 2e-16 ***
day5codeMon     -0.232330   0.110504  -2.102   0.0355 *  
day5codeSa-BH    0.646091   0.083244   7.761 8.40e-15 ***
day5codeTue     -0.081767   0.097487  -0.839   0.4016    
day5codeWe-Th   -0.152445   0.088925  -1.714   0.0865 .  
dog.time12/17   -0.535558   0.100841  -5.311 1.09e-07 ***
dog.time6       -0.192202   0.112955  -1.702   0.0888 .  
dog.time7/10/15  0.122536   0.065583   1.868   0.0617 .  
dog.time8        0.124831   0.094472   1.321   0.1864    
dog.time9        0.399156   0.076175   5.240 1.61e-07 ***
wght_dist       -0.720439   0.036177 -19.914  < 2e-16 ***
band1to4         0.278427   0.016545  16.829  < 2e-16 ***
band5to9         0.162052   0.026452   6.126 8.99e-10 ***
band10to11      -0.006501   0.028706  -0.226   0.8208    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for poisson family taken to be 1)

    Null deviance: 6531.8  on 6837  degrees of freedom
Residual deviance: 5387.2  on 6824  degrees of freedom
AIC: 7713.9

Number of Fisher Scoring iterations: 6

GLMM

Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) ['glmerMod']
 Family: poisson  ( log )
Formula: dog.walkers.count ~ day5code + dog.time + wght_dist + band1to4 +      band5to9 + band10to11 + (1 | SiteID)
   Data: dog.data2
Control: glmerControl(optimizer = "bobyqa")

     AIC      BIC   logLik deviance df.resid 
  6707.6   6810.1  -3338.8   6677.6     6823 

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-1.7131 -0.3860 -0.1928 -0.1154  7.2060 

Random effects:
 Groups Name        Variance Std.Dev.
 SiteID (Intercept) 1.991    1.411   
Number of obs: 6838, groups:  SiteID, 339

Fixed effects:
                Estimate Std. Error z value Pr(>|z|)    
(Intercept)     -2.76661    0.14187 -19.501  < 2e-16 ***
day5codeMon     -0.03030    0.12248  -0.247  0.80464    
day5codeSa-BH    0.54844    0.09099   6.027 1.67e-09 ***
day5codeTue     -0.06656    0.10631  -0.626  0.53126    
day5codeWe-Th   -0.22075    0.09758  -2.262  0.02368 *  
dog.time12/17   -0.48614    0.10772  -4.513 6.39e-06 ***
dog.time6       -0.11158    0.12296  -0.908  0.36414    
dog.time7/10/15  0.18912    0.07199   2.627  0.00861 ** 
dog.time8        0.29850    0.10667   2.798  0.00514 ** 
dog.time9        0.47665    0.08761   5.441 5.31e-08 ***
wght_dist       -0.67480    0.10080  -6.695 2.16e-11 ***
band1to4         0.39614    0.08732   4.537 5.72e-06 ***
band5to9         0.16443    0.09713   1.693  0.09046 .  
band10to11       0.03160    0.10094   0.313  0.75423

score 1 · Answer 1 · answered Mar 17 '21 at 14:34

Very late, but here's my answer in case others come across the question. In the GLM, since you do not provide the information that your data includes repeated observations (~4 per site), it assumes that even your same-site observations are independent and hence tends to be slightly overoptimistic. This is because conceptually, any event X happening twice independently is more "significant" than it happening twice at the same location. Since the model is overoptimistic, the predictions are as well.

As for the coefficients, I'm not sure how you determined that the two "are not hugely different". Your models have the Poisson family specification (with log link function) and thus the coefficients in the output are on the transformed log-scale. You need to transform the coefficients back by exponentiating (exp()) in order to make inferences on your original scale of counts. If you do not do the transformation, what you can interpret is how much change each predictor causes on log(dog.walkers.count).

> (-1.914835) - (-2.76661) # difference between intercepts (log scale)
[1] 0.851775
> exp((-1.914835) - (-2.76661)) # difference between intercepts on original scale of counts
[1] 2.343803
# same as above but in a different manner following the Quotient Rule for exponents:
> exp(-1.914835)/exp(-2.76661)
[1] 2.343803

The difference between intercepts of your GLM (-1.914835) and GLMM (-2.76661) is ~0.85 on the log-scale, i.e., ~2.34 on the original scale. This probably explains the two-fold change in predictions you observed from the two models, provided you used type="response" in the predict() call.

Links:

There are plenty of threads on this forum regarding the use of random effects, which might be relevant here:

Predictions from Poisson GLMM (lme4) lower compared to GLM

1 Answers1

Links: