Dynamic Factor Models: A Tool for Analyzing Longitudinal Public Health and Medical Databases

Public health researchers have access to a wide variety of longitudinal data bases with cohorts for analysis of trends in health behaviors, risk factors and disease outcomes. However, when using the same instrument with the same subjects on different occasions, there is a tendency for autocorrelation of errors over time, also referred to as lagged measurement error. This presentation utilizes data from a random sample of participants in the National Longitudinal Survey of Youth's 1979 cohort to illustrate a technique for assessing stability of latent constructs over time while allowing for and controlling for lagged measurement error. A seven item subscale of the CES-D was administered to the same participants in 1992, 1994 and 2004. The analysis utilized Lisrel models with polychoric correlations and asymptotic covariance matrices, the appropriate analytic technique with ordinal scales. Measurement errors for identically worded items for each time were allowed to correlate across time periods. Results indicated a good model fit (Chi Square = 39.6, p = 0.77, NFI = 0.98, CFI = 0.99, RMSEA = 0.0), with factor loadings ranging from 0.45 to 0.83 and significant relationships between the latent variables across time. Furthermore, there was a significant improvement in Chi Square for the lagged measurement error model compared with a model without it. Discussion of the conclusions includes implications for modeling longitudinal data in public health research.