Hi everyone!
I am investigating if different variables are associated with the choice of a nest box in a wild bird population.
Briefly, I am comparing the characteristics of multiple nest-boxes that were visited by 100 differents birds. However, from this bunch of nest-boxes, only 1 is chosen by each bird at the end. I wonder which is the best way to analyze these type of data.
For the moment I have built a logistic regression model in which my response variable is a binary outcome: a nest is (1) or not chosen (0). For this, I built a dataset with the 100 nest boxes that were chosen (y = 1) by each of the birds and another 100 nest-boxes that were not chosen by the birds (y = 0) (this 'not chosen' nests are picked randomly from the bunch of nest-boxes visited but not chosen of each bird) , coding the ID of the bird as a random factor.
I've repeated this procedure 1000 times and computed the mean estimates. But I am not quite sure this is correct and I do not find a lot of info on the internet. Is there any better procedure to analyze this type of data? Or does anybody know the name of this type of analyses?
Thank you in advance!
Iraida
Why not run a logistic model including all the nests? You could code the binary variable "chosen"=factor(0,1) and then include the different variables describing nest-box characteristics. You can later use re-sampling techniques to estimate robust confidence intervals around your model coefficients, but I don't see the reason for not fitting a model with all the data. Perhaps you can explain why you came with that idea?
Javier Abalos Hi! Thank you for the answer :-). I have run that model and I get a very similar result. It only seemed pretty odd/weird for me comparing 139 chosen nests vs. 2359 not chosen.
Hi Iraida,
for generalized linear regression with a binary response variable this is not an issue considering this is a "precision" question. It becomes an issue when dealing with count or "percentage" data overflooded with zeros and/or ones.
For classification purposes, however, this can also be an issue. At least, depending what part of the model you want to maximize/minimize (loss function). If accuracy is the goal assuming 139/2359 and predicting every thing absent resulting in an accuracy of 2359/(2359+139)*100%=94% would be perfectly fine. Unless you get a better prediction by tuning the model. Yet, if we would use any other loss function it would be suboptimal.
The following comes from the "Encyclopedia of Machine Learning" (Chapter Class Imbalance Problem
): Often the minority class [your 139] is very infrequent, such as 1% of the dataset. If one applies most traditional (cost-insensitive) classifiers on the dataset, they are likely to predict everything as negative (the majority class [2359]). This was often regarded as a problem in learning from highly imbalanced datasets. However, Provost (2000) describes two fundamental assumptions that are often made by traditional cost-insensitive classifiers. The first is that the goal of the classifiers is to maximize the accuracy (or minimize the error rate); the second is that the class distribution of the training and test datasets is the same. Under these two assumptions, predicting everything as negative for a highly imbalanced dataset is often the right thing to do.What you performed on your data by bootstrapping 1000 times is a form of "stratified sampling" (https://en.wikipedia.org/wiki/Stratified_sampling) by randomly "under sampling" of the majority class (https://www.analyticsvidhya.com/blog/2016/03/practical-guide-deal-imbalanced-classification-problems/). Under the assumption you work in R there are different ways of addressing this issue for classification https://www.r-bloggers.com/2020/11/undersampling-by-groups-in-r/.
Best,
- Badges
- Science method
- Similar topics
- Biological Science
- Ecology