lav_test_print() skips $stat.group when NULL

[TDJorgensen]

When I added lav_update_test_custom_h1() in PR #328, I set $stat.group <- NULL because when there is a more restricted h1.model=, the updated test is only available for the whole model. This causes an error when lav_test_printing multigroup models that have a hidden @external$h1.model.

HS.model <- '
    visual  =~ x1 + x2 + x3
    textual =~ x4 + x5 + x6
    speed   =~ x7 + x8 + x9
'
fit1 <- cfa(HS.model, data = HolzingerSwineford1939, group = "school")
fit0 <- cfa(HS.model, data = HolzingerSwineford1939, group = "school", 
            orthogonal = TRUE)
fit0@external$h1.model <- fit1
fit1 # prints group stats
fit0 # no group stats found, error

So I just added an additional condition in lav_test_print() to only print group stats if (ngroups > 1L && !is.null(TEST[[block]]$stat.group)).

[TDJorgensen]

The second commit avoids an error due to the @test$satorra.bentler$scaling.factor set to NA when the user's h1.model fits perfectly. The commit message ends "df==0", but it should have said "$stat==0"

[TDJorgensen]

Something strange is going on that I need time to track down. Please don't merge this PR yet.

[TDJorgensen]

OK, this PR is ready now. Weirdly, the last issue was hard to track down, because it had nothing to do with the lav_test_update_custom_h1() function I made in April. There was an issue with running lavTestLRT() on a single model when multiple test= options were specified, in which case the hidden single_model function still did the old behavior of selecting the second @test element whenever there was one. So I added arguments or lavTestLRT() to pass, which retain the ability to select specific tests.

HS.model <- '
    visual  =~ x1 + x2 + x3
    textual =~ x4 + x5 + x6
    speed   =~ x7 + x8 + x9
'
TESTS <- c("satorra.bentler","scaled.shifted","browne.residual.adf")
fit1 <- cfa(HS.model, data = HolzingerSwineford1939, group = "school",
            test = TESTS)
fit0 <- cfa(HS.model, data = HolzingerSwineford1939, group = "school", 
            test = TESTS, orthogonal = TRUE)
fit0@external$h1.model <- fit1
fit0 # no more error!

I wrote a function to compare the @test slot (which does not adjust for a custom H1 model) to the short summary output that does adjust, followed by calling lavTestLRT() for each test to verify that summary() shows all the right stats.

checkTests  <- function(obj) {
  cat("UNCORRECTED:\n\n")
  lavaan:::lav_test_print(obj@test, nd = 5)

  cat("\n\nCORRECTED:\n\n")
  print(summary(obj, nd = 5, estimates = FALSE))

  cat("\n\nSTANDARD\n")
  print(lavTestLRT(obj, method = "standard"))

  cat("\n\nSCALED\n")
  sb <- lavTestLRT(obj, test = "satorra.bentler")
  print(sb)
  print(c(scale = attr(sb, "scale")[-1]))

  cat("\n\nSCALED and SHIFTED\n")
  sh <- lavTestLRT(obj, test = "scaled.shifted")
  print(sh)
  print(c(scale = attr(sh, "scale")[-1], shift = attr(sh, "shift")[-1]))

  cat("\n")
  print(lavTestLRT(obj, type = "browne.residual.adf"))
  return(invisible(obj))
}

These don't change (no hidden @external$h1.model):

checkStats(fit1)

These do get updated correctly:

checkStats(fit0)

[yrosseel]

Thanks! Merged now.

[TDJorgensen]

Thanks!
FYI, the changes I made do lead to a different test statistic returned by default for single- vs. multiple-model tests. In the absence of other arguments, the single-model test returns the second element, whatever that is, which will be Browne's (1984) when in @test, even if other scaled tests are in @test. In contrast, a multimodel test (in the absence of other arguments) will default to the first scaled @test element.

I'm not sure if you find that problematic, since most users will never call lavaan(test=) with a vector. If they do, they should be careful to know what they are doing.

Should I add a Note to the lavTestLRT.Rd file about this?
Or adjust the lav_test_lrt_single_model() checks to prefer scaled over Browne tests?

[yrosseel]

I will take a look at this soon (next week), when the semester is finally over.