Planned Missing Data Designs in Longitudinal Research: A Practical Guide

• In Event: Measurement Issues in Developmental Research

Abstract

Multiform designs are gaining traction with researchers, as an increasing number of empirical studies are using this type of design. Yet, there remain several barriers to adopting planned missing designs in developmental research (Baraldi & Enders, 2010), including the complexity of design considerations as well as the technical nature of handling missing data in statistical analyses. Beyond the complexities of designing a longitudinal planned missing study, including item assignment and participant assignment, an additional barrier to the implementation of three-form planned missing designs in longitudinal studies is a lack of concrete recommendations for how best to account for missing data in statistical analyses. For datasets with missing data, planned or unplanned, statisticians recommend using auxiliary variables in either the imputation process or the analysis model as they make it more likely for MAR assumptions to be satisfied (Collins et al., 2001; Enders & Peugh, 2004; Graham, 2003).
Thus, the aims of this study were two-fold: 1) provide a detailed example of the use and considerations of planned-missing designs in the context of longitudinal accelerated cohort designs and 2) to illustrate the benefits of using PC-Aux to inform missing data estimation using example analyses of empirical data from an accelerated cohort longitudinal three-form planned missing study. In sum, the current paper aims to encourage the use of planned missing designs and PC-Aux by providing an intuitive guide on how to adopt this approach.
Data come from the Roots of Engaged Citizenship Study, an accelerated cohort study designed using a 3-form planned missing to examine the developmental roots of civic engagement. A total of 5,573 youth ranging in age from 9 to 21 (MageW1 = 12.07; 53.9% female) were surveyed annually from 2013 to 2018 in schools across three U.S. regions. Social responsibility values and civic discussions with parents were used as example data to illustrate the benefits of using PC-Aux comparing traditional FIML estimation to stochastic regression imputation, FIML, and MI estimation using principal components as auxiliary variables.
Chi-square and RMSEA values were lowest in models using FIML with 45 auxiliary variables whereas CFI and TLI values were highest in the MI model. Fraction of missing information values (𝝀j ) for factor loadings, intercepts, and residual variances were nearly always lowest in the single stochastic regression imputation model. Interestingly, however, for autoregressive and cross-lagged paths, they were lowest in the MI model (see Table 1). In growth models, however, the 𝝀j for the means and variances of the growth factors as well as the regression coefficients of the TVCs were lowest in the MI models, suggesting the efficiency of these parameters were less impacted by missing values than in other approaches.
Results from the current study suggest MI is preferable to FIML because auxiliary variables are already included in the dataset resulting in reduced model complexity, estimation times, and fewer convergence issues. Moreover, fraction of missing information values showed less of a loss in efficiency for structural paths and for all parameters in growth models (key model parameters). Implications for application in developmental research will be discussed.

Authors

• Jennifer C. Shubert, Utah Valley University
  Presenting Author

• Laura Wray-Lake, University of California - Los Angeles
  Non-Presenting Author

• Amy K. Syvertsen, Search Institute
  Non-Presenting Author