Spurious Results in the Analysis of Longitudinal Data in Organizational Research

Abstract

Organizational scientists increasingly focus on the dynamics of human behavior through longitudinal and event sampling methodologies. Random coefficient modeling such as hierarchical linear modeling and latent growth modeling is the dominant analytical method for longitudinal data. These models require that the covariance matrix of the errors is time invariant. Unfortunately, if unmeasured or unpredictable influences (i.e., unmeasured variables) consistently impact the dynamic process under investigation, the error term can become time-dependent. If random coefficient modeling is used to analyze data with time-dependent errors, then a serious inflation of Type I error rates, known as spurious regression, is observed. Monte Carlo simulation results from several common random coefficient models are presented to highlight the scope and severity of the problem, focusing on the potential mistaken inferences researchers can make. An analytic strategy is proposed to aid researchers in determining the underlying structure of the error covariance matrix, and alternative statistical techniques are given to analyze data that may contain a time-dependent error term.