Accuracy and Reliability of Autoregressive Parameter Estimates: A Comparison Between Person-Specific and Multilevel Modeling Approaches

On the Use of Empirical Bayes Estimates as Measures of Individual Traits

“…Du and Wang (2018) compared the reliability of various intraindividual variability indicators based on time series data and found that EB estimates of the autoregressive parameters generally have lower reliability than other indicators, such as the intraindividual standard deviation. In terms of factors affecting reliability, they found similar results as Liu (2018). In addition, they also found substantial influences of the measurement scale reliability and a number of other factors, such as the intraindividual variance of the autoregressive process.…”

“…Previous studies on EB reliability (Du & Wang, 2018; Liu, 2018) have mostly focused on multilevel autoregressive models. We chose here to simulate data based on the stress reactivity model represented in Equation 13.…”

“…Despite the popularity of these approaches, little research has examined whether EB estimates are indeed reliable and valid measures of individual traits. Liu (2017Liu ( , 2018 compared EB estimates to regression estimates from person-specific autoregressive models with time series data and found that EB estimates generally have better accuracy, and similar reliability, when all individuals in the sample have homogeneous dynamic patterns, and when the model used to analyze the data is correctly specified. In addition, accuracy and reliability are affected by the number of within-person observations and the random effects variance, but not by sample size or distribution of the random effects.…”

Unique Contributions of Dynamic Affect Indicators – Beyond Static Variability

“…Nevertheless, the possibility of inter-individual differences in σ " ! was disregarded in studies questioning the predictive utility of IADs (Dejonckheere et al, 2019, Houben et al, 2020Wendt et al, 2020) as well as in simulation studies comparing the performance of person-specific and multilevel estimation approaches (Liu, 2017(Liu, , 2018Liu et al, 2021).…”

“…. In the N = 1 setting, a general recommendation for the estimation of 𝜑 " in univariate time-series is a minimum of 50 TP, however, in two-level AR(1) models less TP may be sufficient as N increases (Liu et al, 2017(Liu et al, , 2018Schultzberg & Muthén, 2018). However, the impact of this commonly observed shrinkage effect in multilevel models on the 𝜑 " estimates' recovery might depend on the amount of inter-individual differences in the 𝜑 " (Liu et al, 2017).…”