umxACE, standardized estimates not standardized

[jpritikin]

umx 4.3

library(umx)

data(twinData) # ?twinData from Australian twins.
twinData[, c("ht1", "ht2")] = twinData[, c("ht1", "ht2")] * 10
mzData = twinData[twinData$zygosity %in% "MZFF", ]
dzData = twinData[twinData$zygosity %in% "DZFF", ]
m1 = umxACE(selDVs = "ht", selCovs = "age", sep = "", dzData = dzData, mzData = mzData)

sum(m1$top$a_std$result,
    m1$top$c_std$result,
    m1$top$e_std$result)  # not 1.0

m2 = umxACEv(selDVs = "ht", selCovs = "age", sep = "", dzData = dzData, mzData = mzData)
sum(m2$top$A_std$result,
    m2$top$C_std$result,
    m2$top$E_std$result) # 1.0

• Looks like the umxACE formulas use SD %*% a instead of SD %&% a.
• Why the naming difference? Why use lowercase a_std and uppercase A_std?

[tbates]

a is for paths, A is for Variance (case change is a pretty standard nomenclature for paths and variances)

Variance is what is being standardized, hence summing to 1 (a_std^2 + c_std^2 + e_std^2 sums to 1).

0.929^2+ 0.082^2+ 0.36^2 = 1

[jpritikin]

OK, maybe you can add the missing lower & uppercase algebras so it is easy to switch between umxACE and umxACEv?

[tbates]

For umxACE, adding algebras to compute A_std etc. would mean they get computed every time, so would slow execution (unless there's a way to tell an algebra to compute only once at model completion?)

But also, Is there a use case for wanting, e.g. A_std from an ACE model? The model paths are in the lower-triangle a, c, e matrices, and don't survive the a %*% t(a) algebras except in the univariate 1*1 case.

For umxACEv it doesn't contain any paths/lower-case variables.

[jpritikin]

adding algebras to compute A_std etc. would mean they get computed every time, so would slow execution (unless there's a way to tell an algebra to compute only once at model completion?)

Nope, the algebras are only computed twice: once at the beginning and once at the end,

library(umx)

data(twinData) # ?twinData from Australian twins.
twinData[, c("ht1", "ht2")] = twinData[, c("ht1", "ht2")] * 10
mzData = twinData[twinData$zygosity %in% "MZFF", ]
dzData = twinData[twinData$zygosity %in% "DZFF", ]
m1 = umxACE(selDVs = "ht", selCovs = "age", sep = "", dzData = dzData, mzData = mzData,
              autoRun=FALSE)

m1 <- mxModel(m1, mxModel(m1$top, mxAlgebra(SD %&% A, name="A_std", verbose=2L)))
m1 <- mxRun(m1)

For umxACEv it doesn't contain any paths/lower-case variables.

Yeah, that's true. I didn't really think it through.

[mcneale]

Is that because certain matrices and algebras do not affect the fit function? Either because they are in an mxModel that has no data and there are no label-based equality constraints, nor any uses of the matrices and algebras by mxModels that have data? Getting something like this wrong would be bad… getting it right is definitely more efficient.

[jpritikin]

Is that because certain matrices and algebras do not affect the fit function?

Yeah, exactly

[tbates]

Good to know, re algebras! (slightly amazed that can be automated) umxACE (on GitHub) now has:

mxAlgebra(name = "A_std", SD %&% A), # standardized A
# ...

in addition to the existing.

mxAlgebra(name = "a_std", SD %*% a), # standardized a

[jpritikin]

Awesome! Thank you.