fix: adjust tests for compatibility with lavaan 0.7-1

Author: felixdidi

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: mlstats
 Title: Multilevel Descriptive Statistics and Data Preparation
-Version: 0.1.0.9000
+Version: 0.1.1
 Authors@R: 
     person("Felix", "Dietrich", , "mail@felix-dietrich.de", role = c("aut", "cre", "cph"))
 Description: Provides tools for multilevel descriptive statistics and data preparation.

NEWS.md

@@ -1,4 +1,4 @@
-# mlstats 0.1.0.9000 (development version)
+# mlstats 0.1.1
 
 * `decompose_within_between()` now defaults to `components = c("between",
   "within")`, so grand-mean-centered scores are no longer returned by
@@ -17,7 +17,7 @@
   markdown italics (`*N*`) already used on the same label and with how the
   "a"/"b" correlation-note superscripts are marked elsewhere in the table.
 
-  * `mldesc()` now reports the observed minimum and maximum in the `range`
+* `mldesc()` now reports the observed minimum and maximum in the `range`
   column with two decimals instead of rounding them to whole numbers
   (decimals are dropped when both the minimum and the maximum are whole
   numbers, e.g., for integer scales). Previously, a variable observed
@@ -42,6 +42,27 @@
   the ICC models); `decompose_within_between()` keeps the rows but sets
   their between- and within-group components to `NA`, with a warning.
 
+* Test-only fix for compatibility with lavaan 0.7-1 (no user-visible changes
+  to `mlstats` itself).
+
+  Four `method = "sem"` tests depended on `lavaan` converging on an
+  inadmissible (negative between-level variance) solution for particular
+  degenerate/small-sample models, either expecting `.wb_cor_sem()`'s
+  "out-of-range" warning to fire or a specific between-group correlation to
+  come back as `NA`. lavaan 0.7-1 converges on an admissible solution
+  instead for those same models, so the warning no longer fires and a valid
+  correlation is returned. Three of the affected tests now simply suppress
+  any warning instead of requiring the "out-of-range" one, since that
+  warning was incidental to what they actually check (flip= symmetry, print
+  methods, between-only variable handling). The fourth test, which checks
+  that within-only variables are excluded from the between-group model, now
+  constructs a variable with exactly zero between-group variance so it
+  exercises that exclusion path deterministically rather than relying on
+  lavaan returning an improper estimate.
+
+  The package's test suite now passes with both lavaan 0.6-21 and lavaan
+  0.7-1.
+
 # mlstats 0.1.0
 
 Initial CRAN release.

cran-comments.md

@@ -1,21 +1,28 @@
-## R CMD check results
-
-0 errors | 0 warnings | 1 note
+## Reason for this submission
 
-* This is a new release.
+This is an urgent maintenance release requested by Yves Rosseel (lavaan
+maintainer). A new lavaan feature (`optim.fix.saturated`, default `TRUE` in
+the upcoming lavaan 0.7-1) causes 4 of mlstats' `method = "sem"` tests to
+fail, because those tests depended on lavaan converging on an inadmissible
+(out-of-range) standardized solution for certain degenerate/small-sample
+models. lavaan 0.7-1 now converges on an admissible solution instead for
+those same models. This is currently blocking lavaan 0.7-1's own submission
+to CRAN. The relevant tests have been adjusted accordingly (see NEWS.md);
+no user-facing behavior of mlstats itself has changed. The package's test
+suite passes with both lavaan 0.6-21 (current CRAN release) and lavaan
+0.7-1 (verified locally against the pre-release version from the
+maintainer's r-universe).
 
-## Notes
+This release also bundles a handful of unrelated bug fixes and
+documentation/output improvements accumulated since 0.1.0 (see NEWS.md).
 
-`devtools::check(cran = TRUE)` reports "unable to verify current time" in
-some local runs; this is a network-reachability artifact of the local check
-environment (it cannot reach the time-verification service) and is unrelated
-to package content.
-
-## win-builder / R-hub
+## R CMD check results
 
-Checked on win-builder (R-devel) and R-hub (Linux, macOS x86_64/arm64,
-R-devel). All clean except:
+0 errors | 0 warnings | 2 notes
 
-* win-builder flags "APA", "ICCs", "REWB", and "intraclass" as possibly
-  misspelled in DESCRIPTION. These are established abbreviations/terms in
-  the multilevel-modeling literature, not typos.
+* `checking CRAN incoming feasibility ... NOTE`: "Days since last update: 4".
+  Expected, and the reason for the rush is explained above.
+* `checking for future file timestamps ... NOTE`: "unable to verify current
+  time". This is a network-reachability artifact of the local check
+  environment (it cannot reach the time-verification service) and is
+  unrelated to package content; it also appeared on the 0.1.0 submission.

tests/testthat/test-mldesc.R

@@ -1067,26 +1067,29 @@ test_that("mldesc method='sem' works with flip=TRUE", {
     z = rnorm(200)
   )
 
-  result_normal <- expect_warning_value(
+  # Whether this particular model yields an out-of-range standardized
+  # solution (and hence the "out-of-range" warning) depends on the lavaan
+  # version: older versions converge on an inadmissible between-level
+  # solution here, newer ones do not. That is incidental to what this test
+  # checks -- the flip= symmetry -- so any warning is simply suppressed.
+  result_normal <- suppressWarnings(
     mldesc(
       data = data,
       group = "group",
       vars = c("x", "y", "z"),
       method = "sem",
       flip = FALSE
-    ),
-    "out-of-range"
+    )
   )
 
-  result_flipped <- expect_warning_value(
+  result_flipped <- suppressWarnings(
     mldesc(
       data = data,
       group = "group",
       vars = c("x", "y", "z"),
       method = "sem",
       flip = TRUE
-    ),
-    "out-of-range"
+    )
   )
 
   # Upper triangle of normal should equal lower triangle of flipped
@@ -1134,14 +1137,15 @@ test_that("mldesc method='sem' handles between-only variables correctly", {
   data$x <- rnorm(500, 5, 2)
   data$y <- rnorm(500, 5, 2)
 
-  result <- expect_warning_value(
+  # As above, an "out-of-range" warning here is lavaan-version dependent and
+  # incidental: this test checks the handling of the between-only variable.
+  result <- suppressWarnings(
     mldesc(
       data = data,
       group = "group",
       vars = c("trait", "x", "y"),
       method = "sem"
-    ),
-    "out-of-range"
+    )
   )
 
   # Check structure

tests/testthat/test-within_between_correlations.R

@@ -1102,6 +1102,13 @@ test_that("method='sem' excludes within-only variables from the between-group mo
   data$wi <- rnorm(200)
   data$x <- rnorm(200)
 
+  # Group-mean-center `wi` so that it has *exactly* zero between-group
+  # variance, i.e. it is genuinely within-only and must be dropped from the
+  # between (level 2) model. Plain `rnorm()` would leave a small but nonzero
+  # between-group variance, which does not meet the ICC ~ 0 criterion used by
+  # .wb_cor_sem(), so the exclusion path being tested here would not be taken.
+  data$wi <- data$wi - stats::ave(data$wi, data$group)
+
   result <- suppressWarnings(
     within_between_correlations(data, "group", c("wi", "x"), method = "sem")
   )