I am working with camera-trap data on mammals. My data looks like this:

Zone   Point   Phase   SurveyLength   ProjectDay   Species1   Species2   
A      A1      Before  21             1            0          0          
A      A1      Before  21             2            1          0            
A      A1      Before  21             3            0          1            
B      B1      Before  21             1            2          0            
B      B2      After   21             1            0          0            
B      B3      After   21             1            0          1 

There are ten species in total. I want to compare detection rate between zones and phases, as this is a BACI experiment. Detection rate is measured as the number of observations of a given species divided by effort; effort = camera days. so essentially, detection rate = average for the entire dataset for each species).

I have separated the data by phase, then analyze using a glm. However, even with the most abundant species, the model fails the goodness of fit test. Here is an example of the code I'm using:

before <- subset(Counts, Counts$Phase=="Before")
ZoneB <- factor(before$Zone)
DurationDaysB <- as.numeric(before$SurveyLength)   

glmB <-glm(Species1~ ZoneB, family=poisson(link=log), offset=log(DurationDaysB), data=before)
count.covB <- cbind(ZoneB)
chsq<-sum(residuals(glmB, type = 'pearson') ^ 2)
gofB <- POIS_GOF(mu = glmB$coefficients, sigma = vcov(glmB),
            sims = 1000, chsq.obs=chsq, count.cov = count.covB, 
            offset.count = log(DurationDaysB))

There are a LOT of zero's in my data (histogram included below). After trying the species with the top three detection rates, p=0 in every case, indicating lack of fit with the model.

enter image description here

I've read on the UCLA website about a way to compensate for overdispersion by scaling the data in STATA, but I have not found a way to do this in R. I also tried using the "quasipoisson" distribution, which should allow for overdispersion in models where the dispersion parameter is not fixed. I have not fixed this in my model, but the quasipoisson model returns the same results as the poisson model.

Is there another way I can compensate for overdispersion in my data? Or is there a different model which I should use for this type of data?

I also ran the analysis using an anova followed by a Tukey test from the raw data (as shown below), but then remembered that I would need to use a Poisson distribution for this data. While we're here, I want to verify that this is an incorrect way to analyze the data - thoughts?

species1countB <- aov(before$Species1~before$Zone)

Thank you in advance for your help.


It looks like you don't necessarily have overdispersion in general, but a specific kind of overdispersion, namely too many zeros.

Two very useful approaches are zero-inflated models (e.g., zero-inflated Poisson regression - this page gives an introduction in R), or hurdle models (e.g., using the countreg package). Googling for these two words will yield a bountiful harvest.

Whether one or the other model is more appropriate should depend on just where your overabundance of zeros comes from. Think about your data generating process.

  • 1
    $\begingroup$ note the pscl package can handle both zero-inflated and hurdle variants $\endgroup$ – Ben Bolker Nov 10 '15 at 17:57
  • 1
    $\begingroup$ The important point is the rationale for using these models other than "we saw it in the data". The question that the OP is trying to answer is: "what is the detection rate of our capture-release system?" By allowing for zero inflation in a poisson model, you are implying some sampled sites are simply not speciated. Heavily relying on the assumption of Poisson distribution for counts, you can jointly estimate the proportion of sites which were not speciated. Be sure to report this with the results! $\endgroup$ – AdamO Nov 10 '15 at 18:41
  • $\begingroup$ Thank you, Stephan, your answer is very helpful. My data seems to fit zeroinfl Poisson regression better. My next question is how to determine the differences between all groups. I have three phases (before; during; after) and three zones (control; T; F) - reference categories are listed first here. If I'm interpreting correctly, when I use the interaction term of Phase:Zone I can only see if each combination of non-reference categories is different from the reference combination of before:control. Is there a way I can compare all categories similar to using an ANOVA? $\endgroup$ – abmiller8 Nov 10 '15 at 21:59
  • $\begingroup$ Once you have a significant interaction, you can look at submodels defined on subdata, e.g., constraining your data and your model to the "before" phase and comparing only "control" and "T" zones. This is in fact exactly what post hoc $t$ tests do in ANOVA, basically because the $t$ test is the same as an $F$ test if you have only two groups in your ANOVA. $\endgroup$ – Stephan Kolassa Nov 11 '15 at 7:49
  • $\begingroup$ Thanks again! Another question is in interpreting the output. I would like to plot the detection rates of each zone and phase (9 points for each species) and determine which are different from each other. Can I use the estimates from the ZIP output to plot these? a statistics student told me if I add the estimate for the intercept plus each zone/phase this will indicate the detection rate, but I'm not sure that's correct. Would it make any sense to plot the raw detection rate and then indicate significant difference between groups using the t-test as described above? .... $\endgroup$ – abmiller8 Nov 11 '15 at 15:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.