Hi Kay,

The Information Theoretic approaches are used to weigh evidence for each model in your candidate set, not to develop the candidate set.

Although different people use different approaches to develop a candidate model list, one authoritative source is the book "Model Selection" by Burnham and Anderson. They argue you should develop a small set of a priori models designed to address specific hypotheses, instead of running all possible models.

In the 2002 edition, they write (p. 16):

"Model formulation is the point where the scientific and biological information formally enter the investigation. Building the set of candidate models is partially a subjective art; that is why scientists must be trained, educated, and experienced in their discipline. The published literature and experience in the biological sciences can be used to help formulate a set of a priori candidate models. The most original, innovative part of scientific work is the phase leading to the proper question. Good approximating models, each representing a scientific hypothesis, in conjunction with a good set of relevant data can provide insight into the underlying biological process and structure."

Furthermore (p. 17):

"Development of the a priori set of candidate models often should include a global model: a model that has many parameters, includes all potentially relevant effects, and reflects causal mechanisms thought likely, based on the science of the situation. The global model should also reflect the study design and attributes of the system studied. Specification of the global model should not be based on a probing examination of the data to be analyzed. […] Models with fewer parameters can then be derived as special cases of the global model."

They contrast this with methods they recommend against (p. 19):

"This a priori strategy is in contrast to strategies advocated by others who view modeling and data analysis as a highly iterative and interactive exercise. Such a strategy, to us, represents deliberate data dredging and should be reserved for early exploratory phases of initial investigation."

Hi Kay,

The Information Theoretic approaches are used to weigh evidence for each model in your candidate set, not to develop the candidate set.

Although different people use different approaches to develop a candidate model list, one authoritative source is the book "Model Selection" by Burnham and Anderson. They argue you should develop a small set of a priori models designed to address specific hypotheses, instead of running all possible models.

In the 2002 edition, they write (p. 16):

"Model formulation is the point where the scientific and biological information formally enter the investigation. Building the set of candidate models is partially a subjective art; that is why scientists must be trained, educated, and experienced in their discipline. The published literature and experience in the biological sciences can be used to help formulate a set of a priori candidate models. The most original, innovative part of scientific work is the phase leading to the proper question. Good approximating models, each representing a scientific hypothesis, in conjunction with a good set of relevant data can provide insight into the underlying biological process and structure."

Furthermore (p. 17):

"Development of the a priori set of candidate models often should include a global model: a model that has many parameters, includes all potentially relevant effects, and reflects causal mechanisms thought likely, based on the science of the situation. The global model should also reflect the study design and attributes of the system studied. Specification of the global model should not be based on a probing examination of the data to be analyzed. […] Models with fewer parameters can then be derived as special cases of the global model."

They contrast this with methods they recommend against (p. 19):

"This a priori strategy is in contrast to strategies advocated by others who view modeling and data analysis as a highly iterative and interactive exercise. Such a strategy, to us, represents deliberate data dredging and should be reserved for early exploratory phases of initial investigation."

Thank you so much for the helpful responses, Philip and Matthew.  It reminds me to think about research questions that I am trying to answer instead of taking too much attention to put all the data (especially environmental parameters) that I have to the model and also to read more about occupancy modelling.

Hi Kay,

I think Philip and Matthew gave some good advice. As I was reading through this thread two things came to mind....

First, be cautious of overparameterizing your model. Evaluating 10 site covariates is a lot to ask of a 30-site data set. I would be thinking about what hypothesis each covariate represents and trying to narrow it down a bit.

The other thing was to not throw out the idea of comparing models of all possible combinations of covariates. If your covariates are well thought out and all possible combinations of them represent plausible hypotheses, then there is nothing wrong with evaluating all possible combinations.

Good luck with your study!

Dan

Well noted with thanks, Dan :)

Dear Kay Zin,

I am sorry for my late reply.

In my case, I use Genetic Algorithm to find the best model. GA can search the model relative with the target object during short-time. You can check about the GA in my published paper and conference paper.

Best Reagards,

Phyu

Hi Khaing Win Phyu,

I am doing species-habitat occupancy modelling of the sampled population which is assumed to be closed. I think but not sure that GA can be applied to determine the genetic structure of population and so far GA is not the feasible option to use in species occupancy modelling but I will be happy if you could clarify a bit more about that how Genetic Algorithm can be related to species occupancy of a site. Thanks for your response :)