Multiple imputation of zero‐inflated longitudinal count data based on a piecewise growth curve model | Dr. Kristian Kleinke

Multiple imputation of zero‐inflated longitudinal count data based on a piecewise growth curve model

Abstract

Multiple imputation based on Rubin’s (1987,1996) theory is a state-of-the- art method to address the missing data problem in both multivariate and clustered data sets. To obtain unbiased statistical inferences, the imputation method needs to be ‘proper’ in Rubin’s terms and the imputation model needs to reflect all aspects of the assumed data-generating process including non-linear relationships. The most simple solution to model an increase over time in the variable of interest, and a decrease later on (or vice versa) is to include a quadratic term into the model in addition to the linear time trend. For more complex situations, however, especially for situations, where different growth phases (developmental stages) can be clearly distinguished, a piecewise growth model is typically the better choice. We propose an imputation method for complex panel data based on a piecewise growth curve model, and evaluate this method by means of Monte Carlo simulation. We simulated data containing 12 repeated measurements of a zero-inflated count variable. The simulated data reflect ‘real’ empirical panel data of the Crime in the modern City (CrimoC) study (e.g. Boers, Reinecke, Seddig, & Mariotti, 2010) – the versatility of delinquent behaviours. 30% panel attrition was simulated following a missing at random pattern. Data were imputed and analysed by a piecewise growth model. As assumed, when both the imputation model and the subsequent analysis model are fully compatible to the data generating process (i.e. neither model is mis-specified) the proposed procedure yields unbiased statistical inferences.

Date
Sep 16, 2021 12:00 AM
Location
Mannheim