Multivariate Behavioral Research最新文献

The existing literature on missing data handling in psychological network analysis using cross-sectional data is currently limited to likelihood based approaches. In addition, there is a focus on convex regularization, with the missing handling implemented using different calculations in model selection across various packages. Our work aims to contribute to the literature by implementing a missing data handling approach based on multiple imputation, specifically stacking the imputations, and evaluating it against direct and two-step EM methods. Standardized model selection across the multiple imputation and EM methods is ensured, and the comparative evaluation between the missing handling methods is performed separately for convex regularization (glasso) and nonconvex regularization (atan). Simulated conditions vary network size, number of observations, and amount of missingness. Evaluation criteria encompass edge set recovery, partial correlation bias, and correlation of network statistics. Overall, missing data handling approaches exhibit similar performance under many conditions. Using glasso with EBIC model selection, the two-step EM method performs best overall, closely followed by stacked multiple imputation. For atan regularization using BIC model selection, stacked multiple imputation proves most consistent across all conditions and evaluation criteria.

The social relations model (SRM) is a linear random-effects model applied to examine dyadic round-robin data within social networks. Such data have a unique multilevel structure in that dyads are cross-classified within individuals who may be nested within different social networks. The SRM decomposes perceptual or behavioral measures into multiple components: case-level random effects (in-coming and out-going effects) and dyad-level residuals (relationship effects), the associations among which are often of substantive interest. Multivariate SRM analyses are increasingly common, requiring more sophisticated estimation algorithms. This article evaluates Markov chain Monte Carlo (MCMC) estimation of multivariate-SRM parameters, compares MCMC to maximum-likelihood estimation, and introduces two methods to reduce bias in MCMC point estimates using empirical-Bayes priors. Four simulation studies are presented, two of which reveal dependency of small-group results on priors by manipulating location and precision hyperparameters, respectively. The third simulation study explores the impact of sampling more small groups on prior sensitivity. The fourth simulation study explores how Bayesian model averaging might compensate for underestimated variance due to empirical-Bayes priors. Finally, recommendations for future research are made and extensions of the SRM are discussed.

Intensive longitudinal data (ILD) collection methods like experience sampling methodology can place significant burdens on participants, potentially resulting in careless responding, such as random responding. Such behavior can undermine the validity of any inferences drawn from the data if not properly identified and addressed. Recently, a confirmatory mixture model (here referred to as fully constrained latent Markov factor analysis, LMFA) has been introduced as a promising solution to detect careless responding in ILD. However, this method relies on the key assumption of measurement invariance of the attentive responses, which is easily violated due to shifts in how participants interpret items. If the assumption is violated, the ability of the fully constrained LMFA to accurately identify careless responding is compromised. In this study, we evaluated two more flexible variants of LMFA-fully exploratory LMFA and partially constrained LMFA-to distinguish between careless and attentive responding in the presence of non-invariant attentive responses. Simulation results indicated that the fully exploratory LMFA model is an effective tool for reliably detecting and interpreting different types of careless responding while accounting for violations of measurement invariance. Conversely, the partially constrained model struggled to accurately detect careless responses. We end by discussing potential reasons for this.

This study introduces a new multi-process analytical framework based on the ideal point assumption for analyzing Likert scale data with three newly developed Unfolding Tree (UTree) models. Through simulations, we tested the performance of proposed models and existing Item Response Tree (IRTree) models across various conditions. Subsequently, empirical data were utilized to analyze and compare the UTree models relative to IRTree models, exploring respondents' decision-making processes and underlying latent traits. Simulation results showed that fit indices could effectively discern the correct model underlying the data. When the correct model was employed, both IRTree and UTree accurately retrieved item and individual parameters, with the recovery precision improving as the number of items and sample size increased. Conversely, when an incorrect model was utilized, the mis-specified model consistently returned biased results in estimating individual parameters, which was pronounced when the respondents followed an ideal point response process. Empirical findings highlight that respondents' decisions align with the ideal point process rather than the dominance process. The respondents' choices of extreme response options are more driven by target traits than by extreme response style. Furthermore, evidence indicates the presence of two distinct but moderately correlated target traits throughout the different decision stages.

To address the challenge of determining the number of factors to retain in exploratory factor analysis, a plethora of techniques, called stopping rules, has been developed, compared and widely used among researchers. Despite no definitive solution to this key issue, the recent Next Eigenvalue Sequence Test (NEST) showed interesting properties, such as being theoretically grounded in the factor analysis framework, robustness to cross loadings, a low false positive rate, sensitive to small but true factors, and better accuracy and unbiased compared to traditional stopping rules. Despite these strengths, there is no existing software readily available for researcher. These considerations have led to the development of the R package Rnest. This paper introduces NEST, presents the functionality of the Rnest package, and illustrates its workflow using a reproducible data example. By providing a practical and reliable approach to factor retention, this package aims to encourage its widespread adoption among practitioners, psychometricians, and methodological researchers conducting exploratory factor analyses.

There has been considerable interest in estimating causal cross-lagged effects in two-variable, multi-wave designs. However, there does not currently exist a strategy for ruling out unmeasured time-varying covariates that may act as confounders. In this paper, we propose a new strategy for testing whether an unmeasured time-varying covariate explains all covariation between the two "causal" variables in the data. That model, called the Latent Time-Varying Covariate (LTVC) model, can be tested with observations for two variables assessed across three or more measurement waves. If the LTVC model fits well, then a time-varying covariate can explain the covariance structure, which undermines the plausibility of causal cross-lagged effects. Although the LTVC model tends to be underpowered when causal cross-lagged effects are small, if testable stationarity constraints on the LTVC model are imposed, adequate power can be achieved. We illustrate the LTVC approach with three examples from the literature. Additionally, we introduce the LTVC-CLPM model, which is identified given strong stationarity constraints. Also considered are multivariate and multi-factor models, the inclusion of measured time-invariant covariates in model, measurement of the stability of the LTVC, and the lag-lead model. These methods allow researchers to probe the assumption that an unmeasured time-varying confounder is the source of all the X-Y covariation. Our methods help researchers to rule out certain forms of confounding in two-variable, multi-wave designs.

Causal conclusions about non-randomized treatments rest on the absence of unobserved confounding. While often made in practice, this fundamental yet empirically untestable assumption can rarely be definitively justified. In most realistic settings, the threat of unobserved confounding lurks. Can causal effects be unbiasedly estimated when unobserved confounding is present? In this tutorial, we introduce an approach from the causal inference and epidemiological literature that permits doing so: negative control outcomes. We explain what a negative control outcome is and how to leverage it to counteract bias due to unobserved confounding. Estimation using a negative control outcome is carried out using the Control Outcome Calibration Approach (COCA). To encourage the adoption of COCA in practice, we implement COCA using lavaan, a popular and free statistical modeling software in R. We illustrate COCA using two publicly available real-world datasets. COCA is practically elegant, straightforward to implement, and under certain assumptions about the potential outcomes, able to unbiasedly estimate causal effects even when unobserved confounding is present.

This article describes a simple and fast algorithm for generating uniformly sampled correlation matrices (R) with a fixed mean absolute discrepancy (MAD) relative to a target (population) $R_{\text{pop}}\text{.}$ The algorithm can be profitably used in many settings including model robustness studies and stress testing of investment portfolios, or in dynamic model-fit analyses to generate R matrices with a known degree of model-approximation error (as operationalized by the MAD). Using results from higher-dimensional geometry, I show that $R_{n\times n}$ matrices with a fixed MAD lie in the intersection of two sets that represent: (a) an elliptope and (b) the surface of a cross-polytope. When n = 3, these sets can be visualized as an elliptical tetrahedron and the surface of an octahedron. An online supplement includes $R$ code for implementing the algorithm and for reproducing all of the results in the article.

Suppression effects are important for theoretical and applied research because these effects occur when there is an unexpected increase in an effect when it is adjusted for a third variable. This paper investigates three approaches to testing for statistical suppression. The first test was proposed in 1978 and is based on the relationship between the zero-order and semi-partial correlations. The second test comes from a condition that is necessary for suppression proposed in 1997. The third test is an extension of the test for the inconsistent mediated effect. We derive standard errors for the Velicer, and Sharpe and Roberts tests, conduct a statistical simulation study, and apply all three tests to two real data sets and several published correlation matrices. In the simulation study, the test based on inconsistent mediation had the best properties overall. For the data examples, when raw data were available, we constructed bootstrap confidence intervals to assess significance, and for correlations, we compared each test statistic to the normal distribution to assess statistical significance. Each test gave consistent results when applied to the example data sets. Analytical work demonstrated conditions where each test gave conflicting results. The mediation test of suppression based on the sign of the product of the mediated effect and the direct effect had the best overall performance.

LOO (Leave-One-Out cross-validation) and WAIC (Widely Applicable Information Criterion) are widely used for model selection in Bayesian statistics. Most studies select the model with the smallest value based on point estimates, often without considering the differences in fit indices or the uncertainty of the estimates. To address this gap, we propose a sequential method for comparing models based on confidence intervals for $\Delta \text{LOO}$ or $\Delta \text{WAIC.}$ A simulation study was conducted to evaluate this method in selecting mixed-effects location-scale models (MELSMs). Our study revealed that the sequential methods, especially when using a 90% confidence interval, achieved a higher accuracy rate of model selection compared to the point method when the true model was simple, had a large magnitude of random intercept in the scale model, or had a large sample size. Models selected by the sequential methods demonstrated higher power, narrower credible interval width, smaller standard errors for the fixed effect in the location model, and lower bias for the random effect of the intercept in the location model. Differences between LOO and WAIC were significant only when the level-1 sample size was small, with LOO performing better when the true model had homogeneous or severe heterogeneity in residual variances.