Search for efficient complete and planned missing data designs for analysis of change

Abstract

The design of longitudinal data collection is an essential component of any study of change. A well-designed study will maximize the efficiency of statistical tests and minimize the cost of available resources (e.g., budget). Two families of designs have been used to collect longitudinal data: complete data (CD) and planned missing (PM) designs. This article proposes a systematic and flexible procedure named SEEDMC (SEarch for Efficient Designs using Monte Carlo Simulation) to search for efficient CD and PM designs for growth-curve modeling under budget constraints. This procedure allows researchers to identify efficient designs for multiple effects separately and simultaneously, and designs that are robust to MCAR attrition. SEEDMC is applied to identify efficient designs for key change parameters in linear and quadratic growth models. The identified efficient designs are summarized and the strengths and possible extensions of SEEDMC are discussed.Keywords Planned missing data designs . Growth curve modeling . Efficiency . Longitudinal data collectionThe design of longitudinal data collection is an essential component of any study of change. Factors that researchers must consider when planning such designs include but are not limited to the number and allocation of repeated measures, sample size, and minimizing unwanted factors such as carry-over effects and attrition. In addition to these considerations, researchers are typically constrained by a finite budget and limited resources which should also be taken into account at the design stage. A well-designed study will lead to statistical tests that are maximally efficient, that is, precise and powerful (Berger & Wong, 2009), given available resources.Two families of designs can be used to collect longitudinal data: complete data (CD) and planned missing (PM) designs. Each of these designs represents a unique combination of design factors (e.g., number of repeated measures per participant and allocation of the repeated measures to time points), leading to the differential efficiency of target statistical procedures. Given two designs, the more efficient one produces parameter estimates that have smaller sampling variability, leading to smaller standard errors, tighter confidence intervals, and thus higher power to detect non-zero effects. Thus, more efficient designs are preferred.Although past research has studied CD designs and specific PM designs, no study so far has searched for the most efficient design among all possible CD and PM designs. Furthermore, the effect of unplanned missingness (missingness that is not due to design, e.g., attrition) has been studied for CD designs but not for PM designs. Given that unplanned missingness is ubiquitous in longitudinal studies, this is an important issue to address. Identifying efficient designs is not a trivial matter. As shown in the examples below, using an efficient design may save up to 75 % of the research budget while achieving the same efficiency to detect target effects when compared to commonly ...

Generating missing values for simulation purposes: a multivariate amputation procedure

“…Many studies describe the use of so-called planned missing data designs (e.g. [26][27][28]), which are survey designs where 'missing data are used strategically to improve the validity of data collection.' [29,p.426].…”

Optimal number and allocation of data collection points for linear spline growth curve modeling

“…We extended the Monte Carlo method proposed by Wu et al (2015) to search for efficient data collection designs for linear-linear spline models with a fixed change point. We illustrate the use of the method, taking into account two resource constraints (sample size and budget), homogeneous and heterogeneous residual variances, as well as different degrees of uncertainty on the location of the change point.…”

Optimal planned missing data design for linear latent growth curve models

“…To formally address the question of which PM designs are optimal for modeling longitudinal change, we follow Wu et al ( 2016 ) and focus on latent growth curve models (LGCMs) as statistical models for change over time. LGCMs have become a commonly used analysis technique to capture change in longitudinal data (Duncan et al, 2006 ; Ferrer and McArdle, 2010 ; Meredith & Tisak, 1990 ).…”

“…Leveraging asymptotic results, there is no need to run time-intensive Monte Carlo simulations. Here, we revisit the insightful and inspiring simulation work presented by Wu et al ( 2016 ) to investigate optimal PM designs for the study of linear change. We complement their approach, which is available as R package SEEDMC (Jia and Wu, 2015 ), by providing an asymptotic perspective to their empirical simulations.…”