Evaluating shared parameter mixture models for analyzing change in the presence of non-randomly missing data Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 20, 2019
  • Gottfredson, Nisha C.
    • Affiliation: College of Arts and Sciences, Department of Psychology and Neuroscience
  • Longitudinal researchers have been slow to adopt models for assessing the sensitivity of their results to potentially non-randomly missing data, opting instead to rely exclusively on more traditional approaches to modeling growth like latent curve modeling (LCM). Implicit in this choice is the strict assumption that missing data are missing at random (MAR). Failure to meet this assumption leads to inaccurate inferences regarding growth. A number of models for assessing the impact of non-randomly missing data on growth trajectory estimates have been presented over the past quarter century. These models are briefly discussed, and a new variation on some recently developed models is introduced. The shared parameter mixture model (SPMM) described here is preferable to some other models for a few reasons. Most notably, it approximates the dependence between the missing data process and the repeated measures without requiring an explicit specification of the missingness mechanism while simultaneously allowing conditional independence between the growth model and the missing data. Performance of the SPMM is evaluated using simulation methodology across a range of plausible missingness mechanisms and across a range of longitudinal data conditions. SPMM performs well when the missing data mechanism is either latent class- or growth coefficient- dependent. Fixed effect recovery is more robust than variance component recovery. The SPMM performs best with longer observation lengths and with erratically spaced missing data than with dropout. Finally, this manuscript illustrates how the SPMM might be used in practiceby analyzing change over time in psychological symptoms of patients enrolled in psychotherapy. Results are generally encouraging for SPMM performance under a range of simulated data conditions, and for feasibility with real data. Researchers who suspect the presence of random coefficient-dependent missing data are urged to consider using the SPMM to assess sensitivity of their model results to the MAR assumption.
Date of publication
Resource type
Rights statement
  • In Copyright
  • "... in partial fulfillment of the requirements for a Doctor of Philosophy degree in the Department of Psychology (Quantitative)."
  • Bauer, Daniel
Place of publication
  • Chapel Hill, NC
  • Open access

This work has no parents.