Cholesky mx status RED

The algorithm that calculates the multivariate-normal probability integral only has so much numerical accuracy, and the inaccuracy is worse out into the tails of the distribution. Further, the numerical error is compounded by the finite-differences subroutine that numerically approximates the fitfunction derivatives. As a result, sometimes models of threshold traits cannot satisfy the first or second-order conditions for a local minimum, even when they've actually reached one. In such situations, the best you can do is try to find the smallest possible -2logL. I notice you're using mxTryHard(). Some of the default argument values for mxTryHard() are downright counterproductive for optimization with threshold traits. Use mxTryHardOrdinal() instead (or equivalently use different argument values to vanilla mxTryHard()). Also, if you're not using the on-load default optimizer, CSOLNP, you should be.

The issue of "inescapable status red with threshold variables" is quite common. Two relevant forums threads that readily come to my mind are this one and this one. You could find others if you search our forums.

You could add labels.

Also consider that `umxACE` will do all of this automatically for you in one maintained and flexible function:


m1 = umxACE("multiCholACEModel", selVars = c('PAIN','CMD', 'DISABILITY'), sep = "", dzData = mzDataOrd, dzData = dzDataOrd, tryHard="yes")
umxSummary(m1)
plot(m1)


Bates, T. C., Maes, H., & Neale, M. C. (2019). umx: Twin and Path-Based Structural Equation Modeling in R. Twin Res Hum Genet, 22(1), 27-41. doi:[10.1017/thg.2019.2](https://www.cambridge.org/core/journals/twin-research-and-human-genetics/article/umx-twin-and-pathbased-structural-equation-modeling-in-r/B9658AC0CDA139E540BFAC0C9D989623)

It is pretty much impossible to help when one doesn't know what exactly the problems are. However, good that you post the code which can sometimes give the intrepid support people clues:

Note how in the example part of ?umxThresholdMatrix its result is put into an object called tmp:

tmp = umxThresholdMatrix(twinData, selDVs = selVars, sep = "", verbose = TRUE)

That object should be included in the model. Running it without a target object to put its output into simply dumps the results to the screen and does nothing else.

umxThresholdMatrix(dzDataOrd, selDVs = selVars, sep = "", method = c("auto"), threshMatName = "threshMat", droplevels = FALSE, verbose = FALSE)

should, I think, include a tmp <- at the beginning, and then be used in the mxModel. But (and it's a big one) umxACE does all this for you by detecting which variables are ordinal and which are not.

Incidentally, it is possible to generate data that have similar properties to the real data, but which are definitely not the observed data and therefore can be shared to help others to debug. See ?mxGenerateData

Finally, if code 6 is the only concern, note that the message is often a false alarm caused by the precision of numerical integration making it difficult to find the exact solution - but the solution found may be extremely close to the correct one (no differences in 2nd decimal place of estimates or -2lnL). Rerunning from different starting values can increase confidence that the approximate solution is very close indeed to the actual one.


# Example 1
obesityLevels = c('normal', 'obese')
cutPoints = quantile(twinData[, "bmi1"], probs = .2, na.rm = TRUE)
twinData$obese1 = cut(twinData$bmi1, breaks = c(-Inf, cutPoints, Inf), labels = obesityLevels)
twinData$obese2 = cut(twinData$bmi2, breaks = c(-Inf, cutPoints, Inf), labels = obesityLevels)

tmp = umxThresholdMatrix(twinData, selDVs = selVars, sep = "")
tmp[[2]]
$labels
obese1 obese2
dev_1 "obese_dev1" "obese_dev1"
$values
obese1 obese2
dev_1 -0.888 -0.888
$free
obese1 obese2
dev_1 TRUE TRUE


FYI: `umxThresholdMatrix` builds an algebra and 2 matrices: one consisting of offsets (`deviations_for_thresh`) and the other a lower-1s matrix `lowerOnes_for_thresh` which the algebra (`threshMat`) uses to add up the deviations to form thresholds). It thus handles the complex-for-humans job of building correct thresholds which are used within a model.