Does Acquiescence Bias Distort Measurement Invariance Across Multiple Groups? A Monte Carlo Comparison of Tree-Based Factor Analytic Methods
Year: 2026
Type: preprint
Abstract: Acquiescence (ACQ), the tendency to agree regardless of item content, can differ across groups and thereby mimic or mask measurement non-invariance. Model-based recursive partitioning (MOB) combined with factor-analytic models (e.g. EFAtree, CFAtree) is increasingly used to detect non-invariance across many covariates, but whether the detected heterogeneity can be attributed to its correct source is unknown. In a Monte Carlo simulation (432 conditions, 500 replications each), we compared EFAtree and CFAtree with their random-intercept (RI) versions under four data-generating conditions: groups differed in neither source, in ACQ only, in the content loadings only, or in both. We also propose a block-wise efficient score test that separates instability in the RI variance from instability in the content loadings by projecting out all other parameters. All trees held the nominal Type I error rate. When groups differed only in ACQ, trees without an RI factor mistook the difference for metric non-invariance under unbalanced keying or missed it under balanced keying, whereas the RI trees detected it regardless of keying. Applied to the true grouping variable, the standard block test, which selects a block from the fully decorrelated scores, attributed purely substantive non-invariance to ACQ in 79% of replications, more often in larger samples. The efficient test kept this misattribution near the nominal level (.06) and attributed heterogeneity correctly at the attainable maximum. With one content factor, EFA- and CFA-based trees produced identical results. In an empirical example with the Rosenberg Self-Esteem Scale (N = 26236), the RI variance differed across countries but not across age groups or genders, and even across countries it was only one of several unstable parameter blocks, alongside the intercepts and the factor-mean and factor-variance directions. Before a split is interpreted as measurement non-invariance, its source should be examined with an RI factor and efficient block-wise tests.
Source: PsyArXiv (OSF Preprints)
Authors: Tugay Kaçak, David Navarro González
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225319035