Multivariate Behavioral Research最新文献

A regression discontinuity (RD) design is often employed to provide causal evidence when the randomization of the treatment assignment is infeasible. When variables of interest are latent constructs measured by observed indicators, the conventional RD analysis using observed variable scores does not allow researchers to examine heterogeneity in the estimated local average treatment effect (ATE) and to generalize the ATE to participants away from the cutoff. We propose a novel methodological augmentation to the conventional RD analysis, which assumes the availability of multiple indicator variables (i.e., raw item responses) that measure the latent construct underlying the running variable. By specifying an explicit measurement model based on those indicator variables, our latent RD framework allows 1) defining the local ATE conditional on the latent construct, 2) disentangling the heterogeneity of the local ATE, and 3) generalizing the local ATE to running variable scores away from the cutoff. In a proof-of-concept simulation we illustrate the proposed augmentation recovers parameters of interest well under practical test length and sample size conditions.

In a linear latent growth model parameterized by intercept (α) and slope (β) factors, those factors' relation is often of interest. The model typically captures this through their covariance parameter, which inherently assumes linearity in their relation. However, this assumption may not always hold. For instance, α and β might be unrelated below a certain threshold along the α-axis but show a meaningful relation above it. That is, even though individual growth trajectories may follow a linear pattern over time, the relation between α and β can be nonlinear, potentially featuring distinct segments separated by a transition point. To address such relations, we propose a semiparametric approach that combines Bayesian P-splines for flexible nonlinear modeling of the α-β relation along with a segmented regression-based transition point detection method. This two-stage analytic approach provides for a more nuanced understanding of the α-β relation, including estimation of a potential transition point where the α-β relation structure fundamentally changes. Simulation results and an empirical data illustration support this approach's effectiveness with single transition point scenarios, offering deeper insights into aspects of the growth process.

State-of-the-art causal inference methods for observational data promise to relax assumptions threatening valid causal inference. Targeted maximum likelihood estimation (TMLE), for example, is a template for constructing doubly robust, semiparametric, efficient substitution estimators, providing consistent estimates if the outcome or treatment model is correctly specified. Compared to standard approaches, it reduces the risk of misspecification bias by allowing (nonparametric) machine-learning techniques, including super learning, to estimate the relevant components of the data distribution. We briefly introduce TMLE and demonstrate its use by estimating the effects of private tutoring in mathematics during Year 7 on mathematics proficiency and grades using observational data from starting cohort 3 of the National Education Panel Study ($N=$ 4,167). We contrast TMLE estimates to those from ordinary least squares, the parametric G-formula, and the augmented inverse-probability weighted estimator. Our findings reveal close agreement between methods for end-of-year grades. However, variations emerge when examining mathematics proficiency as the outcome, highlighting that substantive conclusions may depend on the analytical approach. The results underscore the significance of employing advanced causal inference methods, such as TMLE, when navigating the complexities of observational data and highlight the nuanced impact of methodological choices on the interpretation of study outcomes.

Norms play a critical role in high-stakes individual assessments (e.g., diagnosing intellectual disabilities), where precision and stability are essential. To reduce fluctuations in norms due to sampling, normative studies must be based on sufficiently large and well-designed samples. This paper provides formulas, applicable to any sample composition, for determining the required sample size for normative studies under the simplified parametric norming framework. In addition to a sufficiently large sample size, precision can be further improved by sampling according to an optimal design, that is, a sample composition that minimizes sampling error in the norms. Optimal designs are, here, derived for 45 (multivariate) multiple linear regression models, assuming normality and homoscedasticity. These models vary in the degree of interaction among three norm-predictors: a continuous variable (e.g., age), a categorical variable (e.g., sex), and a variable (e.g., education) that may be treated as either continuous or categorical. To support practical implementation, three interactive Shiny apps are introduced, enabling users to determine the sample size for their normative studies. Their use is demonstrated through the hypothetical planning of a normative study for the Trail Making Test, accompanied by a review of the most common models for this neuropsychological test in current practice.

Metamodels, or the regression analysis of Monte Carlo simulation results, provide a powerful tool to summarize simulation findings. However, an underutilized approach is the multilevel metamodel (MLMM) that accounts for the dependent data structure that arises from fitting multiple models to the same simulated data set. In this study, we articulate the theoretical rationale for the MLMM and illustrate how it can improve the interpretability of simulation results, better account for complex simulation designs, and provide new insights into the generalizability of simulation findings.

Multilevel compositional data, such as data sampled over time that are non-negative and sum to a constant value, are common in various fields. However, there is currently no software specifically built to model compositional data in a multilevel framework. The R package multilevelcoda implements a collection of tools for modeling compositional data in a Bayesian multivariate, multilevel pipeline. The user-friendly setup only requires the data, model formula, and minimal specification of the analysis. This article outlines the statistical theory underlying the Bayesian compositional multilevel modeling approach and details the implementation of the functions available in multilevelcoda, using an example dataset of compositional daily sleep-wake behaviors. This innovative method can be used to robustly answer scientific questions from the increasingly available multilevel compositional data from intensive, longitudinal studies.

The appeal of lagged-effects models, like the first-order vector autoregressive (VAR(1)) model, is the interpretation of the lagged coefficients in terms of predictive-and possibly causal-relationships between variables over time. While the focus in VAR(1) applications has traditionally been on the strength and sign of the lagged relationships, there has been a growing interest in the residual relationships (i.e., the correlations between the innovations) as well. In this article, we will investigate what residual correlations can and cannot signal, for both the discrete-time (DT) and continuous-time (CT) VAR(1) model, when inspecting a CT process. We will show that one should not take on a DT perspective when investigating a CT process: Correlated (i.e., non-zero) DT residuals can flag omitted common causes and effects at shorter intervals (which is well-known), but-when having a CT process-also effects at longer intervals. Furthermore, when inspecting a CT process, uncorrelated (i.e., zero) DT residuals do not imply that the variables have no effect on each other at other intervals, nor does it preclude the risk of having omitted common causes. Additionally, we will show that residual correlations in a CT model signal omitted causes for one or more of the observed variables. This may bias the estimation of lagged relationships, implying that the found predictive lagged relationships do not equal the underlying causal lagged relationships. Unfortunately, the CT residual correlations do not reflect the magnitude of the distortion.

Seemingly routine data-preprocessing choices can exert outsized influence on the conclusions drawn from randomized controlled trials (RCTs), particularly in behavioral science where data are noisy, skewed and replete with outliers. We demonstrate this influence with two fully specified multiverse analyses on simulated RCT data. Each analysis spans 180 analytical pathways, produced by crossing 36 preprocessing pipelines that vary outlier handling, missing-data imputation and scale transformation, with five common model specifications. In Simulation A, which uses linear regression families, preprocessing decisions explain 76.9% of the total variance in estimated treatment effects, whereas model choice explains only 7.5%. In Simulation B, which replaces the linear models with advanced algorithms (generalized additive models, random forests, gradient boosting), the dominance of preprocessing is even clearer: 99.8% of the variance is attributable to data handling and just 0.1% to model specification. The ranges of mean effects show the same pattern (4.34 vs. 1.43 in Simulation A; 15.30 vs. 0.56 in Simulation B). Particular pipelines-most notably those that standardize or log-transform variables-shrink effect estimates by more than 90% relative to the raw-data baseline, while pipelines that leave the original scale intact can inflate effects by an order of magnitude. Because preprocessing choices can overshadow even large shifts in statistical methodology, we call for meticulous reporting of these steps and for routine sensitivity or multiverse analyses that make their impact transparent. Such practices are essential for improving the robustness and replicability of behavioral-science RCTs.

Despite the popularity of structural equation modeling in psychological research, accurately evaluating the fit of these models to data is still challenging. Using fixed fit index cutoffs is error-prone due to the fit indices' dependence on various features of the model and data ("nuisance parameters"). Nonetheless, applied researchers mostly rely on fixed fit index cutoffs, neglecting the risk of falsely accepting (or rejecting) their model. With the goal of developing a broadly applicable method that is almost independent of nuisance parameters, we introduce a machine learning (ML)-based approach to evaluate the fit of multi-factorial measurement models. We trained an ML model based on 173 model and data features that we extracted from 1,323,866 simulated data sets and models fitted by means of confirmatory factor analysis. We evaluated the performance of the ML model based on 1,659,386 independent test observations. The ML model performed very well in detecting model (mis-)fit in most conditions, hereby outperforming commonly used fixed fit index cutoffs across the board. Only minor misspecifications, such as a single neglected residual correlation, proved to be challenging to detect. This proof-of-concept study shows that ML is very promising in the context of model fit evaluation.

Indicators of affect dynamics (IADs) capture temporal dependencies and instability in affective trajectories over time. However, the relevance of IADs for the prediction of time-invariant outcomes (e.g., depressive symptoms) was recently challenged due to results suggesting low predictive utility beyond intraindividual means and variances. We argue that these results may in part be explained by mathematical redundancies between IADs and static variability as well as the chosen modeling strategy. In three extensive simulation studies we investigate the accuracy and power for detecting non-null relations between IADs and an outcome variable in different relevant settings, illustrating the effect of the length of a time series, the presence of missing values or measurement error, as well as of erroneously fixing innovation variances to be equal across persons. We show that, if uncertainty in individual IAD estimates is not accounted for, relations between IADs (i.e., autoregressive effects) and a time-invariant outcome are underestimated even in large samples and propose the use of a latent multilevel one-step approach. In an empirical application we illustrate that the different modeling approaches can lead to different substantive conclusions regarding the role of negative affect inertia in the prediction of depressive symptoms.