I am trying to model the visitation frequency of bird to a plant species along an elevational gradient and two season. Moreover I have the number of flowers each observed plant had. My response variable was calculated as follows Frequency= number of visits/(observed time * number of flowers). This, variable it is not normally distributed and plus it has an inflation of zeros (see histogram). To normalize the variable I log-tranformed it and a small constant to avoid getting "inf" values for the zeros. before the transformation and after exploring the dataset with a boxplot, I realised that the relationship between frequency and elevation might not be linear picking at the middle elevation. Thus, I thought of using the General Additive Mixed Effect Models (GAMM). For this purpose I am using the R package mgcv and the function gamm(). My model looks like this:
gam_elevation_all_plants|t|) (Intercept) -7.68425 0.32117 -23.926 < 2e-16 *** flowers 0.07511 0.01624 4.625 6.17e-06 *** seasonWet-dry 1.49145 0.42153 3.538 0.000485 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Approximate significance of smooth terms: edf Ref.df F p-value s(elevation) 3.373 3.373 40.08 %
ggplot(aes(x=elevation, y=log_frequency_f)) +
geom_point(alpha=.3) +
geom_ribbon(data=pred_elevation, alpha=.4, aes(ymin=fit-CI, ymax=fit+CI), show.legend = F, fill='forestgreen') +
geom_line(data=pred_elevation, aes(y=fit), show.legend = F, color='forestgreen') +
theme_classic()
; whereas this the one used to back transformed the estimates:
ggplot(hypericum_all_plants, aes(x = elevation, y = frequency_f) ) +
geom_point() +
geom_smooth(aes(y = exp(predict.gam(gam_elevation_all_plants$gam))), size = 1)
Am I misunderstanding something? is it okay if I report my results with the log_tranformed variable. Also I would to understand which distribution sould I used for my model. Also Can someone help out to interprete the model summary? Basically which of the summaries should I use to plot my estimations the ones of GAM or Lme?
Can someone help me out?
Thank you
Why would your data have to be normally distributed?
Why would you do any log-transformation?
What kind of relationship do you expect? Some kind of peak around a certain altitude?
You can try this software; it has many model functions that might be useful:
www.lerenisplezant.be/fitting.htm
If you can share your data, I can try it for you.
Hi Guillermo Uceda Gómez
a few comments.
1) Here are a few resources on GAMMs (perhaps, however, you know them already)
https://fromthebottomoftheheap.net/2021/02/02/random-effects-in-gams/
http://jacolienvanrij.com/Tutorials/GAMM.html
Article Generalised additive mixed models for dynamic analysis in li...
2) Your response looks like zero-inflated Poisson, however, due to your calculation scheme, its values are non-counts. Other canditates for error distributions could be Tweedie, Gamma or exponential. The mgcv package does not allow many distributions in a GAMM case; perhaps you take a look onto the gamm4 package instead.
I did not read the whole thread, however. Please consider attaching a small txt file with code and/or output as posting this in normal text field is hard to read :)
Good luck
Holger
- Badges
- Science topic
- Similar topics
- Psychology
- Mental Processes
- Learning