QuantDev Brown Bag Seminar Series (In-Person)

Google map

Detrending Longitudinal Panel Data for Modeling of Intraindividual Variability

Abstract: 

Trends typically represent systematic variations, such as developmental changes that unfold over slower time scales. These patterns of intraindividual change contrast with other more nuanced, “momentary” patterns of intraindividual variability that may be captured using popular time series models such as autoregressive models. It is well established in the time series and multilevel literature that results from fitting models of intraindividual variability might be severely biased when trends exist in the data but are not properly accounted for, or removed (i.e., “detrended). However, few guidelines exist to facilitate decisions on appropriate detrending methods for longitudinal panel data (e.g., data with T = 5). 

Using a Monte Carlo simulation study, I evaluated results from fitting a multilevel autoregressive (ML-AR) model to longitudinal panel data when nonlinear trends in the form of person-specific Gompertz curves are accounted for through: (1) two-stage detrending procedures with different degrees of misspecifications of the trend component; and (2) a single-stage Bayesian structured ML-AR model that simultaneously incorporates the trends and intraindividual variability functions. Estimation results from different trend-handling approaches will be discussed, with a step-by-step illustration of ways to diagnose and handle trends using children's reading scores from the Early Children Longitudinal Study (Kindergarten Class; ECLS-K).