CODEWITHSUNDEEP
rlme4mixed-modelslmer
user2027051
user2027051

Reputation: 153

Discrepancy between glmer and difflsmeans for poisson models

I am having trouble understanding some discrepancies in the results between glmer with a poisson model and difflsmeans. Both functions are from the lmerTest package. Basically, glmer tells me the two coefficients are significant at p < 0.05 but when I use difflsmeans it gives a different result. I am using a poisson link and a offset to model count data, and including a fixed effect for the treatments and random effects for the batch of the experiments.

In other analyses using a gaussian link both, lmer and difflsmeans, give the same result.

Is it valid to use difflsmeans for a mixed-model with a poisson link?

Is there another way to check the significance of the coefficients? (I know this has been asked before but I mean in the context of this analysis)

Thanks in advance

RESULTS FROM GLMER

Generalized linear mixed model fit by maximum likelihood ['glmerMod']
Family: poisson ( log )
Formula: EventType1ObjectCount ~ offset(log(ValidObjectCoun)) + Treatment +      (1 | Plate) 
Data: data.m2 

  AIC       BIC    logLik  deviance 
 3050.641  3061.986 -1521.321  3042.641 

Random effects:
Groups Name        Variance Std.Dev.
Plate  (Intercept) 0.787    0.8871  
Number of obs: 126, groups: Plate, 5

Fixed effects:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept) -0.93879    0.39800  -2.359  0.01834 *  
TreatmentB  -0.26771    0.01961 -13.650  < 2e-16 ***
TreatmentD  -0.06326    0.01937  -3.266  0.00109 **

RESULTS FROM difflsmeans

Differences of LSMEANS:
                 Estimate Standard Error     DF t-value Lower CI Upper CI p-value
Treatment F-B    3e-01       2.23e-01  3e+06    1.20   -0.170    0.706     0.2
Treatment F-D    1e-01       2.14e-01  3e+06    0.30   -0.357    0.483     0.8
Treatment B-D   -2e-01       2.22e-01  3e+06   -0.92   -0.640    0.231     0.4

Upvotes: 1

Views: 672

Answers (2)

IRTFM
IRTFM

Reputation: 263352

You ask: is it valid to use Poisson (log) link. I think a simple No is the correct answer. I think it's just getting ignored.

The testing of significance requires that you specify a hypothesis. I have yet to see one.

Upvotes: 0

user3393068
user3393068

Reputation: 11

The lmerTest package is supposed to work only with lmer objects, so it is not valid to use difflsmeans on glmer objects. Thank you for pointing at that - of course an error should come out when applying difflsmeans to other than lmer objects..

Upvotes: 1

Related Questions