When should I use ADE instead of ACE?

ISG Workshop 2026 Day 3
1

When should I use ADE instead of ACE?

2

A number of scenarios. But the big one is when the MZ correlation is more than twice the DZ correlation.

3

The ACE model will fit the data better when your rDZ is more that 1/2 of the rMZ. Otherwise, the ADE model will fit better.

4

We usually use a heuristic - if DZr >= .5 MZr then use ACE otherwise ADE

5

OR add an extra relative type (eg. half-sibs) and model ACDE :)

6

You can also use generalized method-of-moments to fit ACDE in the single-trait case: Using Non-Normal SEM to Resolve the ACDE Model in the Classical Twin Design - PMC

7

Many of the below points have been made by others in the thread.

In a univariate analysis of monozygotic (MZ) and dizygotic (DZ) twins, parameter estimates are based upon three observed statistics: the variance of the phenotype, the covariance between MZ twins, and the covariance between DZ twins. Note that there is only one variance statistic, as all variances are parameterized to have the same expectation across MZ and DZ twins, and for first and second born twins. Each statistic may be expressed as a function of the model parameters:

VAR(P) = V_A + V_D + V_C + V_E

COV(MZ) = V_A + V_D + V_C

COV(DZ) = 0.5*V_A + 0.25*V_D + V_C

where VAR(P) is the phenotypic variance of the variable under study, COV(MZ) is the covariance between MZ twins, COV(DZ) the covariance between DZ twins, and V_A, V_D, V_C and VE
are the latent additive genetic, dominance genetic, common environmental and unique environmental variances that are being estimated (i.e. the parameters).

With three observed statistics, you can only estimate a maximum of three parameters, less the model be unidentified. Typically investigators estimate either an ACE model (if the DZ correlation > 0.5*MZ correlation) and assume D is zero, or an ADE model (if the MZ correlation < 0.5*MZ correlation) and assume that C is zero. Indeed estimates of C and D negatively confounded (see answer to a question above in the thread).

The addition of other sorts of relatives can enable simultaneous estimation of C and D.

I would not advocate looking at higher order moments in the classical twin design to try to resolve C and D simultaneously.

Dave

8

Note that this method is suitable when the data are on an interval scale, where the assumptions of multivariate normality of the phenotypes are met. Neuroimaging traits may be good for this, but beware of the need for large N. The use of method-of-moments with sum scores, factor scores etc may prove misleading - behavior geneticists beware.