tests/vcov-etc.R

stopifnot(require(lme4))

(testLevel <- lme4:::testLevel())
source(system.file("testdata", "lme-tst-funs.R", package="lme4", mustWork=TRUE))# -> unn()


## "MEMSS" is just 'Suggest' -- must still work, when it's missing:
if (suppressWarnings(!require(MEMSS, quietly=TRUE)) ||
    (data(ergoStool, package="MEMSS") != "ergoStool")) {

    cat("'ergoStool' data from package 'MEMSS' is not available --> skipping test\n")
} else {

fm1   <-  lmer (effort ~ Type + (1|Subject), data = ergoStool)
##sp no longer supported since ~ 2012-3:
##sp fm1.s  <- lmer (effort ~ Type + (1|Subject), data = ergoStool, sparseX=TRUE)
## was segfaulting with sparseX (a while upto 2010-04-06)

fe1   <- fixef(fm1)
##sp fe1.s <- fixef(fm1.s)

print(s1.d <- summary(fm1))
##sp print(s1.s <- summary(fm1.s))
Tse1.d <- c(0.57601226, rep(0.51868384, 3))
stopifnot(exprs = {
    ##sp all.equal(fe1, fe1.s, tolerance= 1e-12)
    all.equal(Tse1.d, unname(se1.d <- coef(s1.d)[,"Std. Error"]),
              tolerance = 1e-6) # std.err.: no too much accuracy
    is(V.d <- vcov(fm1), "symmetricMatrix")
    ##sp all.equal(se1.d, coef(s1.s)[,"Std. Error"])#, tol = 1e-10
    ##sp all.equal(  V.d, vcov(fm1.s))#, tol = 1e-9
    all.equal(Matrix::diag(V.d), unn(se1.d)^2, tolerance= 1e-12)
})

}## if( ergoStool is available from pkg MEMSS )

### -------------------------- a "large" example -------------------------
str(InstEval)

if (FALSE) { # sparse X is not currently implemented, so forget about this:

system.time(## works with 'sparseX'; d has 1128 levels
fm7 <- lmer(y ~ d + service + studage + lectage + (1|s),
             data = InstEval, sparseX=TRUE, verbose=1L, REML=FALSE)
)
system.time(sfm7 <- summary(fm7))
fm7 # takes a while as it computes summary() again !

range(t.fm7 <- coef(sfm7)[,"t value"])## -10.94173  10.61535 for REML, -11.03438  10.70103 for ML

m.t.7 <- mean(abs(t.fm7), trim = .01)
#stopifnot(all.equal(m.t.7, 1.55326395545110, tolerance = 1.e-9)) ##REML value
stopifnot(all.equal(m.t.7, 1.56642013605506, tolerance = 1.e-6)) ## ML

hist.t <- cut(t.fm7, floor(min(t.fm7)) : ceiling(max(t.fm7)))
cbind(table(hist.t))

}# fixed effect 'd' -- with 'sparseX' only --------------------------------

if(testLevel <= 1) { cat('Time elapsed: ', proc.time(),'\n'); q("no") }

## ELSE : (testLevel > 1) :
library(lattice)
source(system.file("testdata/lme-tst-funs.R", package="lme4", mustWork=TRUE))
##--> all.equal(), isOptimized(), ...

system.time(
    fm8.N <- lmer(y ~ service * dept + studage + lectage +
                  (1|s) + (1|d), InstEval, REML=FALSE,
                  control=lmerControl("Nelder_Mead"), verbose = 1L)
    ) ## 14   sec [MM@lynne; 2022-11]
      ## 62   sec [MM@lynne; 2013-11]
      ## 59.5 sec [nb-mm3;   2013-12-31]
system.time(
    fm8.B <- lmer(y ~ service * dept + studage + lectage +
                  (1|s) + (1|d), InstEval, REML=FALSE,
                  control=lmerControl("bobyqa"), verbose = 2L)
    )
      ##  7.8 sec [MM@lynne; 2022-11]
      ## 34.1 sec [nb-mm3; 2013-12-31]

system.time(
    fm8 <- lmer(y ~ service * dept + studage + lectage +
                    (1|s) + (1|d), InstEval, REML=FALSE,
                verbose = 1L)
) ## 7.8 sec [MM@lynne; 2022-11]

stopifnot(isOptimized(fm8.N), isOptimized(fm8.B), isOptimized(fm8))
all.equal(fm8.B, fm8,   tolerance=0)# 9.78e-9 (2022-11); both versions of bobyqa
all.equal(fm8.B, fm8.N, tolerance=0)
## "Mean relative difference: 3.31 e-06"   [nb-mm3; 2013-12-31]
stopifnot(isOptimized(fm8.N), isOptimized(fm8.B), isOptimized(fm8))

str(baseOpti(fm8))
str(baseOpti(fm8.N))
str(baseOpti(fm8.B))

(sm8 <- summary(fm8.B))
str(r8 <- ranef(fm8.B))
noquote(sapply(r8, summary))
r.m8 <- cov2cor(vcov(sm8))
Matrix::image(r.m8, main="cor(<fixed eff[ lmer(*,InstEval) ]>)")

if(testLevel <= 2) { cat('Time elapsed: ', proc.time(),'\n'); q("no") }

## ELSE: testLevel > 2

## Clearly smaller X, but more RE pars
## ==> currently considerably slower than the above

system.time(
    fm9 <-
        lmer(y ~ studage + lectage +
                 (1|s) + (1|d) + (1|dept:service) + (1|dept),
             InstEval, verbose = 1L, REML=FALSE)
)
##  25.6 secs [MM@lynne; 2022-11]
## 410   secs [MM@lynne; 2013-11]

fm9
(sm9 <- summary(fm9))
rr <- ranef(fm9, condVar = TRUE) ## ~ 6 secs

noquote(sapply(rr, summary))
qqr <- qqmath(rr, strip=FALSE)
## NB: x-axis range <==> scale of RE <==> "importance" of effect
qqr$d
qqr$s
dotplot(rr,strip=FALSE)$`dept:service`

cat('Time elapsed: ', proc.time(),'\n') # for ``statistical reasons''

Try the lme4 package in your browser

Any scripts or data that you put into this service are public.

lme4 documentation built on Nov. 5, 2023, 9:06 a.m.