Forum: help

RE: Comparison of Multimodel (multiSAR) - Output with linear regression model [ Reply ]
By: François Guilhaumon on 2013-09-24 09:12

[forum:39835]

AIC and other information theoretic criteria are only comparable for models dealing with the same Y (explanied variable). Comparing models of the form "S~" with models of the form "log(S)~" makes no sense. To compare linear regression with non linear mmSAR models it's mandatory to work with S and not log(S) for the linear regression. If the normality of the residuals (and.or homoscedasticity) cannot be achieved with "S" for the linear regression then the linear model cannot be fitted to the data and should be excluded from the set of plausible models for the given SAR.

Hope this help,

cheers,

François

Comparison of Multimodel (multiSAR) - Output with linear regression model [ Reply ]
By: Bernd L on 2013-09-24 08:35

[forum:39834]

I used the multiSAR-function to fit various models to my dataset. Besides that I as well fitted a linear model to the data. I log-transformed both (richness and area) values to reach normality of residuals and assure variance homogeneity.
My problem now is that the corrected AIC values derived from multiSAR-function are negative (between -55 and -56) but the AICc value of the linear fit is at 181.
Inspecting the model fit visually it is observable the the linear and the exponential model almost identically fit the data (with as well very similar R² values!).

How do I interpret the negative cAIC values and how can I compare the nonlinear model fits with the linear one based on information theoretical considerations?

Thanks for any help and suggestions.