Test statistics bottleneck

[haziqj]

The current bottleneck in the test statistics calculation is the call to lav_model_vcov from the {lavaan} package. It is more apparent when $p$ is large (e.g. $p=15$ models).

lavaan.bingof/R/04-test_statistics.R

Lines 70 to 98 in a525c2b

	get_sensitivity_inv_mat <- function(.lavobject, matrix_type = c("Sensitivity",
	"Sandwich")) {
	# Returns the matrix H^{-1}, the inverse of the sensitivity matrix
	list2env(extract_lavaan_info(.lavobject), environment())
	matrix_type <- match.arg(matrix_type, c("Sensitivity", "Sandwich"))
	# This extracts the inverse of the observed sensitivity matrix H = E(nabla^2
	# pl). It is stored as E.inv as part of the Godambe information matrix (E.inv
	# %*% B0 %*% E.inv) because lavoptions has se = "robust.huber.white"
	VCOV <- lav_model_vcov(lavmodel = lavmodel,
	lavsamplestats = lavsamplestats,
	lavoptions = lavoptions,
	lavdata = lavdata,
	lavpartable = lavpartable,
	lavcache = lavcache)
	# Note: VCOV here is (H %*% J^{-1} %*% H)^{-1} = H^{-1} %*% J %*% H^{-1} and
	# it is the "full" information matrix. E.inv is therefore the unit information
	# matrix, and J has the full information.
	if (matrix_type == "Sensitivity") {
	mat <- attr(VCOV, "E.inv") # not yet divided by sqrt n
	} else if (matrix_type == "Sandwich") {
	mat <- VCOV # already divided by sqrt n
	}
	tt <- sum(lavpartable$free != 0)
	if (is.null(mat)) mat <- diag(tt)
	mat
	}

profvis results