Hi. I have count data from camera traps and want to assess if there is a relationship between capture rate and a covariable. When i standardize counts per number of trap nights, the variable gets continuous. Is it okay to run GLM with Poisson or negative binomial dist even my count data is now continuous? If not, is there a way around this problem?
You should consider a GLM of the gamma family. The gamma distribution is just modelling variables like "rates" (as it is a generalization of the exponential distribution).
Hi Jochen. Thank you so much for your answer. I tried that, however, I have 0 in my dataset. I could not find how to deal with that. Do you know if I can change the 0 to 0.01 for example? Or add 1 to all my dataset?
Thank you very much!
Best,
Ramiro
If there are few cases in your data, then this is ok. I would think of substutute these data to half of the lowest non-zero value. You may check if the occurrences of zeros is similar in your groups (as specified by the covariate; if the covariate is continuous you may check a logistic regression on the presence/absence of zero). If there is a strong imbalance, the substitution can bias the results. If this bias is in the same direction as the reported effect, a bit more care must be taken for the interpretation. If the bias is in the opposite direction, well, then the reported effect is already a "careful guess".
If there are many zeros in your data (don't know... more than 5-10%...), it might be better to use a zero-adjusted model (zero-inflated or hurdle model).
Thank you very much indeed for your response. You have been really helpful.
Regards,
Ramiro
Dear Ramiro,
If your raw data is count data with zeros included, I would not advice you to transform it into continuous data before analysis. Instead, as you say, the Poisson and negative binomial GLMs are excellent starting points for modelling your data. I would neither (at this point) start thinking about zero-inflated models or hurdle models - zeros are a common and perfectly sensible outcome for Poisson and neg. bin. processes with low expectation.
There most common way to properly go around the problem of varying effort between different data point (here counts) is the use of a so-called "offset variable", which by definition has an effect (regression coefficient) equal to 1. In log-linear models (with log-link; such as the Poisson and neg. bin GLM) you put the natural logarithm of a measure of effort as the offset. In your case it would be offset = log(trap.nights).
Hope this helps!
Best wishes,
Andreas
Did this ever end up working? I am attempting to do a similar analysis.
Thanks,
SA
Hi,
Andreas, I have not seen your post until now. So I want to thank you for your comment.
Sarah, I ended up changing the analysis to occupancy with presence/absence data. But Andreas advice sounds like a good option. I have not tried it though.
Best,
Ramiro
- Badges
- Science topic
- Similar topics
- Agricultural Science
- Agronomy
- Crop Science