My question concerns choosing the best fitting model. Given the twin correlations in a given phenotype (rMZ = 0.50, rDZ = 0.34), it seems that the model to be tested should be the ACEmodel (because rMZ < 2rDZ).
Runing the ACE model has given the following output.
> mxCompare( fitACE, fitAE ) base comparison ep minus2LL df AIC diffLL diffdf p 1 oneACEc <NA> 4 4824.870 532 3760.870 NA NA NA 2 oneACEc oneAEc 3 4827.013 533 3761.013 2.143052 1 0.1432167 lbound estimate ubound oneAEc.h2[1,1] 0.3833 0.4961 0.5925 oneAEc.c2[1,1] 0.0000 0.0000 0.0000 oneAEc.e2[1,1] 0.4075 0.5039 0.6167 > mxCompare( fitACE, fitCE ) base comparison ep minus2LL df AIC diffLL diffdf p 1 oneACEc <NA> 4 4824.870 532 3760.870 NA NA NA 2 oneACEc oneCEc 3 4826.179 533 3760.179 1.308865 1 0.2526002 lbound estimate ubound oneCEc.h2[1,1] NA 0.0000 NA oneCEc.c2[1,1] 0.3308 0.4331 0.5253 oneCEc.e2[1,1] 0.4747 0.5669 0.6692 > mxCompare( fitAE, fitE ) base comparison ep minus2LL df AIC diffLL diffdf p 1 oneAEc <NA> 3 4827.013 533 3761.013 NA NA NA 2 oneAEc oneEc 2 4881.842 534 3813.842 54.82935 1 1.314631e-13 lbound estimate ubound oneEc.h2[1,1] 0 0 0 oneEc.c2[1,1] 0 0 0 oneEc.e2[1,1] 1 1 1
Based on minimizing the AIC and parsimony, I would choose the CE as the best fitting model.
However, I have also run the ADE model and the results confused me.
> mxCompare( fitADE, fitAE ) base comparison ep minus2LL df AIC diffLL diffdf p 1 oneADEc <NA> 4 4827.013 532 3763.013 NA NA NA 2 oneADEc oneAEc 3 4827.013 533 3761.013 -2.692104e-10 1 1 lbound estimate ubound oneAEc.h2[1,1] 0.3833 0.4961 0.5925 oneAEc.d2[1,1] NA 0.0000 NA oneAEc.e2[1,1] 0.4075 0.5039 0.6167 > mxCompare( fitADE, fitDE ) base comparison ep minus2LL df AIC diffLL diffdf p 1 oneADEc <NA> 4 4827.013 532 3763.013 NA NA NA 2 oneADEc oneDEc 3 4881.842 533 3815.842 54.82935 1 1.314631e-13 lbound estimate ubound oneDEc.h2[1,1] NA 0 NA oneDEc.d2[1,1] NA 0 NA oneDEc.e2[1,1] NA 1 NA > mxCompare( fitAE, fitE ) base comparison ep minus2LL df AIC diffLL diffdf p 1 oneAEc <NA> 3 4827.013 533 3761.013 NA NA NA 2 oneAEc oneEc 2 4881.842 534 3813.842 54.82935 1 1.314631e-13 lbound estimate ubound oneEc.h2[1,1] NA 0 NA oneEc.d2[1,1] NA 0 NA oneEc.e2[1,1] NA 1 NA
Here it would seem that the AE model provides the best fit, and that removing A significantly worsens the model fit, therefore, A should not be dropped.
By choosing the CE model from the ACEmodel, I would be dropping exactly A.
Should I take into account the results from the ADEmodel, or just go ahead with applying the rule rMZ < 2rDZ and test ACEmodel?
I worry about not reporting misleading results.
Thank you,
Mirela
Printer-friendly version
The CE model has the lowest AIC, right? Assuming you're only going to report the results from one model, then select the CE model. Also, it's not biologically plausible for there to be D but no A, so your DE model should be excluded from consideration.