During the analyses of distance sampling using Multiple-covariate distance sampling procedure I followed the step wise approach to incorporate different covariates into the distance software for estimating detection function. But I have doubt that to fit the model I used initially only on covariate and selected the one model with lowest AIC and then progress to add the remain covariate one by one by keeping the original one until all covariates are used to fit the models. The problem I faced here is that I do not have any prior knowledge whether I should stick only to the lowest AIC or to take into account the 2nd or 3rd stage analsysis at which the final fitted model with low AIC and multiple covariates exist. Below is the snapshot of the result section attached. How can I resolve this doubt. Thank you!
Take a look to the following links. It could be useful in line with this study.
https://doi.org/10.1111/2041-210X.12105
https://doi.org/10.1642/AUK-14-11.1
https://doi.org/10.1371/journal.pone.0118726
Dear Hamid,
Thank you so much. I will take a look at them and working on analysis accordingly
I don't think that's how you do it. Your own model needs to incorporate all variables, then you fit different types of distributions to that. Only then do you select the distribution with the lowest AIC. You cannot compare AICs between your different models having different covariates. You need only one big model to ensure you capture all interactions :)
Dear Dr. Pegman (PhD),
Thank you so much. My intention was not to compare AICs between different models having different covariates. I used the forward stepwise approach to see the interaction between the covariates using the model which has lowest AIC at initial. For instance I simply add one covariate (e.g. cluster size) to fits detection function using HN & HR key function. Then I compare the AIC btwn them and select HN because it has lowest AIC. To take look at the interactions of different covariates, I then proceed by incorporating the covariate one by one until all covariates are fitted the data (HN+Cluster size, HN+cluster size+site, HN+cluster size + site+ altitude). So, as I understand from your feedback I do not need to compare btwn the covariates interaction and select the one with lowest AIC? IS that right? or simply I use one big model with all covariates (HN+ALL COVAR, HR+ All covar) and select the one with lowst AIC. Thank you once again for your cooperation.
Of course, you can use AIC to compare MCDS models having different covariates. The only thing you need to watch out for is to make sure that the models being compared are the same dataset (same filter and same truncation) when using AIC to compare models. There is no cookbook for model selection. There are plenty of resources you can benefit from. Besides, always check if the model fits the data or not by examining qq-plots and goodness-of-fit tests such as CvM. You can check the link below to see the basic model selection process.
https://examples.distancesampling.org/Distance-lines/lines-distill.html#model-selection-is-not-a-cookbook
Dear Ugur (PhD),
Thank you so much. I will accommodate the feedback and rerun the model to fit the data by examining the lowest AIC. Thank you again!
- Badges
- Science topic
- Similar topics
- Geoscience
- Cartography
- Mapping