The effects of measurement non-invariance on parameter estimation in latent growth models Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 21, 2019
  • Wirth, R.J.
    • Affiliation: College of Arts and Sciences, Department of Psychology and Neuroscience
  • Researchers are increasingly taking advantage of the latent growth modeling framework to evaluate complex behaviors over time. However, many constructs in the social sciences change in definition over the course of development. These changes may include an item's relationship to a construct over time (i.e., measurement non-invariance). Unfortunately, the effect of changing measurement structures on latent growth model (LGM) parameter estimates is not well understood. This paper begins with a brief introduction to LGMs followed by the restrictions placed on the traditional measurement model used within the LGM framework. Following this, an introduction to measurement invariance is provided within the context of longitudinal confirmatory factor analysis (CFA). Methods available to incorporate flexible measurement structures within longitudinal models of growth are introduced. The various roles measurement invariance may play in the study of stability and growth are presented. The results of a Monte-Carlo simulation are provided and generally show that when complete measurement invariance was maintained, both mean and factor score-based methods recovered generating parameter values well. If at least partial lambda invariance was maintained, factor scores based on a single invariant item also provided unbiased estimates of the random effects. This result was not found using mean or factor scores based on constraining all factor loadings (within item) to equality over time. A measurement structure that changes systematically over time led to biased estimates of almost all parameters regardless of which scoring method was used as well as observed non-linear trends over time. This study did find that the use of factor scores as indicators in LGMs led to consistently positively biased fit statistics. Possible sources of this misfit are discussed. This paper concludes with a discussion of future research and the implications of these findings in applied research.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Curran, Patrick
Degree granting institution
  • University of North Carolina at Chapel Hill
  • Open access

This work has no parents.