Paper Summary
Share...

Direct link:

Bayesian Estimation of Nonlinear Trajectories for Multiple Baseline Design Data

Fri, April 13, 2:15 to 3:45pm, Pan Pacific, Floor: Lobby Level, Oceanview 1&2

Abstract

The multiple-baseline design (MBD) is commonly used in educational research conducted with special populations. MBDs entail repeated measures for each individual (for a small set of individuals, outcomes and/or settings) during a baseline (A) and treatment (B) phase. While visual analysis is typically used, researchers are increasingly encouraged to employ quantitative summaries of treatments’ effects. Unfortunately, MBD data is fraught with methodological complications. These complexities include autocorrelated residuals, outcomes’ scales, multilevel data structure, and the small sample sizes typically used MBDs. Models and methods have been suggested for modeling some of these complexities (e.g., Beretvas & Chung, 2008; Ferron et al., 2009; Nagler, Rindskopf, & Shadish, 2008; Van den Noortgate & Onghena, 2003b). In addition, many MBD datasets consist of nonlinear growth trajectories, however, most of the suggested models are designed for linear growth trajectories. A recent simulation study (Mulloy, 2011) assessed recovery of a three-level nonlinear model’s using a quadratic function. However, the typical MBD trajectory characterized by upper and lower asymptotes seems better described by a logistic function.
Beretvas (2010) derived two models involving multilevel logistic functions for two different types of MBD trajectories. The first trajectory involves no trend in the outcome measure during the baseline phase and a smooth transition from baseline data to the nonlinear trajectory introduced during the intervention phase:

Y_ij = (1-D_ij) *(Gamma00+u-gamma_0j) + (1)
(D_ij)*{alpha10 + (alpha20-alpha10)/[1+(pi00+u_pi0j)*exp(-pi10*Time_ij-n1)]}+e_ij

where _ij represents a subscript for measurement occasion i for person j, n1 represents the number of measurement occasions in baseline, D is the dummy-coded phase indicator and Time identifies the measurement occasion (coded with a zero for the first occasion).
The second trajectory also entails a smooth transition to a nonlinear trajectory during the treatment phase, however, the preceding baseline phase involves a linear developmental trajectory:

Y_ij = (1-D_ij) *[(Gamma00+u-gamma_0j) + (Gamma10+u_gamma_1j)*Time_ij] + (2)
(D_ij)*{alpha10 + (alpha20-alpha10)/[1+(pi00+u_pi0j)*exp(-pi10*Time_ij-n1)]}+e_ij.

Although Beretvas (2010) recommended use of SAS PROC NLMIXED for estimating these nonlinear models, she did not evaluate parameter recovery. The current study involved a simulation study that was designed to assess recovery of the two multilevel logistic model’s parameters [including the lower and upper asymptote parameters (alpha10 and alpha20, respectively) as well as the pseudo-intercept and slope parameters (pi00 and pi10)] comparing estimates using SAS PROC NLMIXED versus Bayesian estimation (using BUGS code). The number of measures and individuals were manipulated. Generating values for the pseudo-slope and intercept parameters as well as variance component values (for the level one residual and for pseudo-intercept and pseudo-slope level two residuals) were varied across conditions. For each combination of conditions, 1,000 replications were generated. Parameter bias was assessed and compared across estimation procedures for each parameter. Last, convergence rates were also tallied.
The R code has been written for generating the datasets. The BUGS code has also been written for the Bayesian estimation of the nonlinear model’s parameters (see Equations 1 and 2). Data generation and model estimation will be complete by October, 2011 and the full paper will be finished by January, 2012.

Authors