current status: latest tests and info about commits are in https://github.com/lme4/lme4/blob/flexSigma/misc/cs_tests.R

It looks like Lambdat as constructed by the new machinery is transposed from where it should be ... (the other difference is that the new code path has the values from theta inserted into Lambdat already, but I don't think that's a relevant difference ...)

correcting the transposition fixes a lot of problems but there are still test failures, beyond the tests that Greg introduced for his stuff (which we may want to skip temporarily) - may try to track these down tomorrow, hoping they may actually be minor.

The Dyestuff example is currently failing because the method for generic function compute_correlation_matrix with signature c(object="UnstructuredCovariance") does cov2cor(V) for V of class dsyMatrix and the method for cov2cor in package Matrix tests that V is positive definite, signaling an error if not. The old unstructured code behaved more like the cov2cor in stats by coping with standard deviations equal to zero.

Maybe it is blasphemous for me to say, but I don't see a real benefit of using Matrix classes in covariance_structures.R and covariance-helpers.R.

I commented out some of Greg's tests, fixed others, took out the dsyMatrix stuff. We're down to two errors, one of which is a mismatch in matrix dimensions (which suggests a real problem with Lambdat construction), one of which is a slight change in results from previous reference values (which I would definitely like to understand but is not necessarily problematic — I believe the example is fairly numerically unstable)

I believe the test failure in test-methods.R:43 (evaluating fmPix <- lmer(pixel ~ day + I(day^2) + (day | Dog) + (1 | Side/Dog), data = Pixel)) is serious but fixable. mkReTrms reorders the random effects terms from largest to small number of levels of the grouping variable (in this case Dog:Side=20, Dog=10, Side = 2). The s4_object_list that mkReLambdat operates on has not been correspondingly re-ordered. (I know we had conversations with Greg about this ...).

(There was an apparently unnecessary/redundant re-ordering step which I've commented out for now, which un-breaks that test and doesn't appear to break anything else in devtools::test() ... will try next with R CMD check ...)

Examples run OK now. Tests show 4 errors:

• tests/testthat/test-methods.R: lmer->lFormula->mkReLambdat constructs Lambdat with wrong dimensions (52 instead of 42)
• tests/testthat/test-summary.R: difference from reference value
• tests/boundary.R: m5B@optinfo$derivs is now NULL
• tests/priorWeights.R: difference from reference value

Oops - crossed up.

Hmm, I thought I fixed the first one just now. Pull and try again?

when running at test level 5 (export LME4_TEST_LEVEL=5), I get one example failure, in the confint.merMod examples ...

Computing profile confidence intervals ...
Warning in split.default(v, rep.int(seq_along(n), n * (n + 1)/2)) :
  data length is not a multiple of split variable
Warning in m0[lower.tri(m0, diag = TRUE)] <- x :
  number of items to replace is not a multiple of replacement length
Error in dd(opt[seq(npar1)]) : length(pars) == np is not TRUE
Calls: confint ... profile.merMod -> all.equal -> unname -> dd -> stopifnot

In addition to the test-summary reference value, I get a few problems in test-predict that are enabled at test level > 1:

══ Failed tests════════════════════════
── Failure ('test-predict.R:239:5'): only look for columns that exist in re.form ──
unname(predict(m1, re.form = ~1 | h/f, newdata = dd[1, ])) not equal to 0.14786.
1/1 mismatches
[1] 0.151 - 0.148 == 0.00264
── Failure ('test-predict.R:240:5'): only look for columns that exist in re.form ──
unname(...) not equal to 0.1533.
1/1 mismatches
[1] 0.159 - 0.153 == 0.00545
── Failure ('test-summary.R:57:3'): lmer ───────────────────────────────────────
tfun(cc2) not equal to c(...).
1/10 mismatches
x[10]: "x.2  0.135 -0.103"
y[10]: "x.2  0.136 -0.103"

Yeah, after pulling and rebuilding, I no longer see the error from tests/testthat/test-methods.R.

The confint.merMod example seems to fail because devfun2 expects that pp$theta is a concatenation of vectors of length p*(p+1)/2. That is no longer true here, where it parametrizes a diagonal covariance matrix. This issue is known: #858.

I reproduce the errors from tests/testthat/test-predict.R.

Finding the source of the differences in fitted values from reference values is going to take some more work. In theory the changes to the way unstructured models are built should result in identical problem set-up → identical solutions ... ???