Bernd Lenzner
2013-Sep-19 14:03 UTC
[R] SAR: nonlinear and linear mixed model comparison using mmSAR - package
Hello, I have a problem trying to compare model-fit between linear and non-linear mixed models. I am using the mmSAR-package from Guilhaumon (Version 1.0) to fit different models (power, exponential, lomolino, weibull etc.) to my data. Besides that I am as well fitting a linear model to the data. My dataset looks like this: a s 1 21.28038 2.944439 2 21.22179 3.091042 3 22.09917 3.526361 4 21.66947 2.197225 5 21.66702 3.713572 6 21.58499 3.044522 7 22.03311 3.465736 8 22.03289 2.397895 9 22.16306 3.610918 . . . with a = area estimate and s = species richness. Both variabels are log - transformed to reach normality and variance homogeneity. Now I want to compare the model fit of the linear model with the fit of the non-linear models by using information theory. I use corrected AIC values to estimate model fit but when I use the "multiSAR" function from the mmSAR package to genereate non-linear model fit I obtain negative AICc values: >mod.selection <- multiSAR(MODELS,DATA) >mod.selection$filtOptimRes p1 p2 p3 AICc power 1.349178e-01 9.999853e-01 0.00000000 -55.64722 expo -1.201421e+01 4.856363e+00 0.00000000 -56.93415 negexpo 1.063781e+08 1.268287e-09 0.00000000 -55.64728 logist 3.762325e+00 3.364975e-01 6.02467719 -54.79978 ratio -2.427444e-02 8.295920e-02 -0.01806225 -54.56568 lomolino 3.925678e+00 5.436657e+02 18.23351785 -54.79876 weibull 3.550748e+00 2.352926e-06 4.39606412 -54.81074 on the other hand I get a positive AICc value for the linear model >lin <- lm(log(s) ~ log(a)) >AICc(lin) > 181.50 How do I interpret these results? Visually inspecting the regressions shows that the exponential regression line is almost identical to the linear one so I would expect that the AICc value should be somewhere in the same range. Thanks for the help or suggestions.