Comparing levels of factors after a GLM in R

Question

Here is a little background about my situation: my data refer to the number of prey successfully eaten by a predator. As the number of prey is limited (25 available) in each trial, I had a column "Sample" representing the number of available prey (so, 25 in each trial), and another called "Count" which was the number of success (how many prey were eaten). I based my analysis on the example from the R book on proportion data (page 578). The explanatory variables are Temperature (4 levels, which I treated as factor), and Sex of the predator (obviously, male or female). So I end up with this model:

model <- glm(y ~ Temperature + Sex + Temperature*Sex, 
          data=predator, family=quasibinomial) 

After getting the Analysis of Deviance table, it turns out that Temperature and Sex (but not the interaction) have a significant effect on the consumption of prey. Now, my problem: I need to know which temperatures differ, i.e., I have to compare the 4 temperatures to each other. If I had a linear model, I would use the TukeyHSD function, but as I am using a GLM I can't. I have been looking through the package MASS and trying to set up a contrast matrix but for some reason it doesn't work. Any suggestions or references?

Here's the summary I get from my model, if that helps making it clearer ...

y <- cbind(data$Count, data$Sample-data$Count)
model <- glm(y ~ Temperature + Sex + Temperature*Sex, 
             data=predator, family=quasibinomial) 
> summary(model)
Call:
glm(formula = y ~ Temperature + Sex + Temperature * Sex,
   family=quasibinomial, data=data)


Deviance Residuals:
Min       1Q   Median       3Q      Max
-3.7926  -1.4308  -0.3098   0.9438   3.6831
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept)                             -1.6094     0.2672  -6.024 3.86e-08 ***
Temperature8                             0.3438     0.3594   0.957   0.3414
Temperature11                           -1.0296     0.4803  -2.144   0.0348 *
Temperature15                           -1.2669     0.5174  -2.449   0.0163 *
SexMale                                    0.3822     0.3577   1.069   0.2882
Temperature8:SexMale                    -0.2152     0.4884  -0.441   0.6606
Temperature11:SexMale                    0.4136     0.6093   0.679   0.4990
Temperature15:SexMale                    0.4370     0.6503   0.672   0.5033
---
Signif. codes:  0 ‘*’ 0.001 ‘’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
(Dispersion parameter for quasibinomial family taken to be 2.97372)
Null deviance: 384.54  on 95  degrees of freedom
Residual deviance: 289.45  on 88  degrees of freedom
AIC: NA
Number of Fisher Scoring iterations: 5

Answer 1

Anne, I will shorty explain how to do such multiple comparisons in general. Why this doesn't work in your specific case, I don't know; I'm sorry.

Edit: Nowadays, I'd recommend using the emmeans package to do pairwise comparisons of the marginal means. Another possibility is the multcomp package and the function glht, which compares the coefficients of the main effects and not the marginal means. In the presence of an interaction, this makes less sens in my opinion. Here is an example:

mydata      <- read.csv("https://stats.idre.ucla.edu/stat/data/binary.csv")
mydata$rank <- factor(mydata$rank)
my.mod      <- glm(admit~gre+gpa*rank, data=mydata, family=quasibinomial)
summary(my.mod)

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -4.985768   2.498395  -1.996   0.0467 *
gre          0.002287   0.001110   2.060   0.0400 *
gpa          1.089088   0.731319   1.489   0.1372
rank2        0.503294   2.982966   0.169   0.8661
rank3        0.450796   3.266665   0.138   0.8903
rank4       -1.508472   4.202000  -0.359   0.7198
gpa:rank2   -0.342951   0.864575  -0.397   0.6918
gpa:rank3   -0.515245   0.935922  -0.551   0.5823
gpa:rank4   -0.009246   1.220757  -0.008   0.9940
---
Signif. codes:  0 ‘*’ 0.001 ‘’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

If you wanted to calculate the pairwise comparisons of the marginal means between the ranks using Tukey's HSD with emmeans, you could do that in this way:

library(emmeans)
em <- emmeans(my.mod, "rank")
contrast(em, "pairwise", adjust = "Tukey")
contrast      estimate    SE  df z.ratio p.value
 rank1 - rank2    0.659 0.323 Inf   2.038  0.1739
 rank1 - rank3    1.296 0.358 Inf   3.620  0.0017
 rank1 - rank4    1.540 0.426 Inf   3.611  0.0017
 rank2 - rank3    0.637 0.291 Inf   2.190  0.1261
 rank2 - rank4    0.881 0.372 Inf   2.370  0.0830
 rank3 - rank4    0.244 0.401 Inf   0.608  0.9296
Results are given on the log odds ratio (not the response) scale. 
P value adjustment: tukey method for comparing a family of 4 estimates

The function emmeans prints a warning that there are interactions present which may make the results misleading. For more information, the emmeans package offers a lot of tutorials (called vignettes).

Now let's compare the coefficients of the main effect for rank with multcomp:

library(multcomp)
summary(glht(my.mod, mcp(rank="Tukey")))
Fit: glm(formula = admit ~ gre + gpa * rank, family = quasibinomial, 
    data = mydata)
Linear Hypotheses:
           Estimate Std. Error z value Pr(>|z|)
2 - 1 == 0   0.5033     2.9830   0.169    0.998
3 - 1 == 0   0.4508     3.2667   0.138    0.999
4 - 1 == 0  -1.5085     4.2020  -0.359    0.984
3 - 2 == 0  -0.0525     2.6880  -0.020    1.000
4 - 2 == 0  -2.0118     3.7540  -0.536    0.949
4 - 3 == 0  -1.9593     3.9972  -0.490    0.960
(Adjusted p values reported -- single-step method)
Warning message:
In mcp2matrix(model, linfct = linfct) :
  covariate interactions found -- default contrast might be inappropriate

All pairwise comparisons of the main effect for rank are given together with a $p$-value. Warning: These comparisons only concern the main effects. The interactions are ignored. If interactions are present, a warning will be given (as in the output above). For a more extensive tutorial on how to proceed when interactions are present, see these additional multcomp examples.

Note: As @gung noted in the comments, you should - whenever possible - include temperature as a continuous rather than a categorical variable. Concerning the interaction: you could perform a likelihood ratio test to check whether the interaction term significantly improves the model fit. In your case, the code would look something like that:

# Original model
model <- glm(y ~ Temperature+Sex+Temperature*Sex, data=predator, family=quasibinomial)
Model without an interaction
model2 <- glm(y ~ Temperature+Sex data=predator, family=quasibinomial)
Likelihood ratio test
anova(model, model2, test="LRT")

If this test is not significant, you may remove the interaction from your model. Maybe glht will work then?