Hello
I want to compare temporal trend of a population because it went through dry (2008-10) and wet phase (2011-13) for 3 years each. At dry phase data has lots of zeros (60%) so presumably be fitted with zero-inflated quasi-Poisson/negative binomial but the wet phase fits quasi-Poisson best with zeros < 5%. I work with GLMs only at R platform.
I wonder if I'm doing wrong by fitting two era with two different distribution. Please help
A problem with using different distributions (models) is that your parameter estimates will not be comparable. You might want to model using each of the distributions / approaches and determine whether a single approach has an acceptable trade-off (variance explained, extremes observations, leverage, goodness-of-fit in terms of predicted values etc) that can be applied to both data sets.
Well that is weird, I just posted an answer and wanted to expand on it.
Have a look at the publication in the link (if you are not aware of it already).
Hello,
Of course I am not familiar with your data, but I guess that there may be no need for splitting your dataset into two periods. At least I think that a fair alternative may be to fit a two-stage model consisting of two submodels: 1) the first one aiming at predicting the probability of occurrence of the attribute of interest in your population for different climatic conditions throughout the whole period (2008-2013) using for instance a logistic model, and 2) the second submodel aiming at predicting the amount of the attribute of interest in your population (also through the whole period) conditional on its occurrence (as predicted by the first submodel). Maybe this article can inspire you since it also deals with longitudinal series of (often) zero-inflated data: http://www.sciencedirect.com/science/article/pii/S0378112714004344. In case you cannot get access to it I can provide you with an electronic copy. At least, hope this helps.
Best regards
Thanks a lot for more insight. I will look up to them. I'm not aiming at model selection but I definitely concerned for good fit of the data. For ex. during dry phase Species A almost did not sample till early 2010 (2008-10) so lots of zeros are inevitable. but the same species was all over during (2011-2013). As in both cases the dispersion parameter were very high so I pick negative binomial. I'm confident that zero-inflation is required for 1st half but not for the 2nd. Which I'm doing when I separate the data set
but only when I'm trying to do these GLMs I'm bit confused
(Species Count~Phase (Dry/Wet), family = ?)
(Species Count~ Entire period (2008-2013), family = ?)
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R