Psychological Methods, Jan 16, 2025, No Pagination Specified; doi:10.1037/met0000702

Compositional data are multivariate data made up of components that sum to a fixed value. Often the data are presented as proportions of a whole, where the value of each component is constrained to be between 0 and 1 and the sum of the components is 1. There are many applications in psychology and other disciplines that yield compositional data sets including Morris water maze experiments, psychological well-being scores, analysis of daily physical activity times, and components of household expenditures. Statistical methods exist for compositional data and typically consist of two approaches. The first is to use transformation strategies, such as log ratios, which can lead to results that are challenging to interpret. The second involves using an appropriate distribution, such as the Dirichlet distribution, that captures the key characteristics of compositional data, and allows for ready interpretation of downstream analysis. Unfortunately, the Dirichlet distribution has constraints on variance and correlation that render it inappropriate for some applications. As a result, practicing researchers will often resort to standard two-sample t test or analysis of variance models for each variable in the composition to detect differences in means. We show that a recently published method using the Dirichlet distribution can drastically inflate Type I error rates, and we introduce a global two-sample test to detect differences in mean proportion of components for two independent groups where both groups are from either a Dirichlet or a more flexible nested Dirichlet distribution. We also derive confidence interval formulas for individual components for post hoc testing and further interpretation of results. We illustrate the utility of our methods using a recent Morris water maze experiment and human activity data. (PsycInfo Database Record (c) 2025 APA, all rights reserved)

Psychological Methods, Jan 09, 2025, No Pagination Specified; doi:10.1037/met0000722

Intensive longitudinal data, increasingly common in social and behavioral sciences, often consist of multivariate time series from multiple individuals. Dynamic factor analysis, combining factor analysis and time series analysis, has been used to uncover individual-specific processes from single-individual time series. However, integrating these processes across individuals is challenging due to estimation errors in individual-specific parameter estimates. We propose a method that integrates individual-specific processes while accommodating the corresponding estimation error. This method is computationally efficient and robust against model specification errors and nonnormal data. We compare our method with a Naive approach that ignores estimation error using both empirical and simulated data. The two methods produced similar estimates for fixed effect parameters, but the proposed method produced more satisfactory estimates for random effects than the Naive method. The relative advantage of the proposed method was more substantial for short to moderately long time series (T = 56–200). (PsycInfo Database Record (c) 2025 APA, all rights reserved)

Psychological Methods, Jan 06, 2025, No Pagination Specified; doi:10.1037/met0000720

Intensive longitudinal data analysis, commonly used in psychological studies, often concerns outcomes that have strong floor effects, that is, a large percentage at its lowest value. Ignoring a strong floor effect, using regular analysis with modeling assumptions suitable for a continuous-normal outcome, is likely to give misleading results. This article suggests that two-part modeling may provide a solution. It can avoid potential biasing effects due to ignoring the floor effect. It can also provide a more detailed description of the relationships between the outcome and covariates allowing different covariate effects for being at the floor or not and the value above the floor. A smoking cessation example is analyzed to demonstrate available analysis techniques. (PsycInfo Database Record (c) 2025 APA, all rights reserved)

Psychological Methods, Jan 06, 2025, No Pagination Specified; doi:10.1037/met0000673

Causality is a fundamental part of the scientific endeavor to understand the world. Unfortunately, causality is still taboo in much of psychology and social science. Motivated by a growing number of recommendations for the importance of adopting causal approaches to research, we reformulate the typical approach to research in psychology to harmonize inevitably causal theories with the rest of the research pipeline. We present a new process which begins with the incorporation of techniques from the confluence of causal discovery and machine learning for the development, validation, and transparent formal specification of theories. We then present methods for reducing the complexity of the fully specified theoretical model into the fundamental submodel relevant to a given target hypothesis. From here, we establish whether or not the quantity of interest is estimable from the data, and if so, propose the use of semi-parametric machine learning methods for the estimation of causal effects. The overall goal is the presentation of a new research pipeline which can (a) facilitate scientific inquiry compatible with the desire to test causal theories (b) encourage transparent representation of our theories as unambiguous mathematical objects, (c) tie our statistical models to specific attributes of the theory, thus reducing under-specification problems frequently resulting from the theory-to-model gap, and (d) yield results and estimates which are causally meaningful and reproducible. The process is demonstrated through didactic examples with real-world data, and we conclude with a summary and discussion of limitations. (PsycInfo Database Record (c) 2025 APA, all rights reserved)

Psychological Methods, Dec 30, 2024, No Pagination Specified; doi:10.1037/met0000723

Intervention studies in psychology often have a partially nested design (PND): after individuals are assigned to study arms, individuals in a treatment arm are subsequently assigned to clusters (e.g., therapists/therapy groups) to receive treatment, whereas individuals in a control arm are unclustered. Given the presence of clustering in the treatment arm, it can be of interest to examine the heterogeneity of treatment effects across the clusters; but this is challenging in PNDs. First, in defining a causal effect of treatment for a specific cluster, it is unclear how the treatment and control outcomes should be compared, as the clustering is absent in the control arm. Although it may be tempting to compare outcomes between a specific cluster and the entire control arm, this crude comparison may not represent a causal effect even in PNDs with randomized treatment assignments, as the cluster assignment may be nonrandomized (elaborated in this study). In this study, we develop methods to define, identify, and estimate the causal effects of treatment across specific clusters in a PND where the treatment and/or cluster assignment may be nonrandomized. Using the principal stratification approach and potential outcomes framework, we define causal estimands for the cluster-specific treatment effects in two scenarios: (a) no-interference and (b) within-cluster interference. We identify the effects under the principal ignorability assumption. For estimation, we provide a multiply-robust method that can protect against misspecification in a nuisance model and can incorporate machine learning methods in the nuisance model estimation. We evaluate the estimators’ performance through simulations and illustrate the application using an empirical PND example. (PsycInfo Database Record (c) 2024 APA, all rights reserved)

Psychological Methods, Dec 30, 2024, No Pagination Specified; doi:10.1037/met0000683

Recent reviews report that about 80% of empirical factor analyses are applied to Likert-type responses and that it is exceedingly common to treat Likert-type item responses as continuous. However, traditional model fit index cutoffs like the root-mean-square error of approximation ≤ .06 or comparative fit index ≥ .95 were derived to have 90+% sensitivity to misspecification with continuous responses. A disconnect therefore emerges whereby traditional methodological guidelines assume continuous responses whereas empirical data often contain Likert-type responses. We provide an illustrative simulation study to show that this disconnect is not innocuous—the sensitivity of traditional cutoffs to misspecification is close to 100% with continuous responses but can fall considerably if 5-point Likert responses are treated as continuous in some conditions. In other conditions, the reverse may occur, and traditional cutoffs may be too strict. Generally, applying traditional cutoffs to Likert-type responses can adversely impact conclusions about fit adequacy. This article aims to address this prevalent issue by extending the dynamic fit index (DFI) framework to accommodate Likert-type responses. DFI is a simulation-based method that was initially intended to address changes in cutoff sensitivity to misspecification because of model characteristics (e.g., number of items, strength of loadings). Here, we propose extending DFI so that it also accounts for data characteristics (e.g., number of Likert scale points, response distribution). Two simulations are included to demonstrate that—with 5-point Likert-type responses—the proposed method (a) improves upon traditional cutoffs, (b) improves upon DFI cutoffs based on multivariate normality, and (c) consistently maintains 90+% sensitivity to misspecification. (PsycInfo Database Record (c) 2024 APA, all rights reserved)

Psychological Methods, Dec 30, 2024, No Pagination Specified; doi:10.1037/met0000715

Models where some part of a mediation is moderated (conditional process models) are commonly used in psychology research, allowing for better understanding of when the process by which a focal predictor affects an outcome through a mediator depends on moderating variables. Methodological developments in conditional process analysis have focused on between-subject designs. However, two-instance repeated-measures designs, where each subject is measured twice: once in each of two instances, are also very common. Research on how to statistically test mediation, moderation, and conditional process models in these designs has lagged behind. Judd et al. (2001) introduced a piecewise method for testing for mediation, that Montoya and Hayes (2017) then translated to a path-analytic approach, quantifying the indirect effect. Moderation analysis in these designs has been described by Judd et al. (2001, 1996), and Montoya (2018). The generalization to conditional process analysis remains incomplete. I propose a general conditional process model for two-instance repeated-measures designs with one moderator and one mediator. Simplifications of this general model correspond to more commonly used moderated mediation models, such as first-stage and second-stage conditional process analysis. An applied example shows both how to conduct the analysis using MEMORE, a free and easy-to-use macro for SPSS and SAS, and how to interpret the results of such an analysis. Alternative methods for evaluating moderated mediation in two-instance repeated-measures designs using multilevel approaches are also discussed. (PsycInfo Database Record (c) 2024 APA, all rights reserved)

Psychological Methods, Dec 30, 2024, No Pagination Specified; doi:10.1037/met0000682

Using instruments comprising ordered responses to items is ubiquitous for studying many constructs of interest. However, using such an item response format may lead to items with response categories infrequently endorsed or unendorsed completely. In maximum likelihood estimation, this results in nonexisting estimates for thresholds. This work focuses on a Bayesian estimation approach to counter this issue. The issue changes from the existence of an estimate to how to effectively construct threshold priors. The proposed prior specification reconceptualizes the threshold prior as prior to the probability of each response category, which is an easier metric to manipulate while maintaining the necessary ordering constraints on the thresholds. The resulting induced-prior is more communicable, and we demonstrate comparable statistical efficiency with existing threshold priors. Evidence is provided using a simulated data set, a Monte Carlo simulation study, and an example multigroup item-factor model analysis. All analyses demonstrate how at least a relatively informative threshold prior is necessary to avoid inefficient posterior sampling and increase confidence in the coverage rates of posterior credible intervals. (PsycInfo Database Record (c) 2024 APA, all rights reserved)

Psychological Methods, Dec 30, 2024, No Pagination Specified; doi:10.1037/met0000721

Often in educational and psychological studies, researchers are interested in understanding the mediation mechanism of longitudinal (repeated measures) variables. Almost all longitudinal mediation models in the literature stem from structural equation modeling framework and hence, cannot directly estimate intrinsically nonlinear functions (e.g., exponential, linear–linear piecewise function with an unknown changepoint) without using reparameterizations. The current study aims to develop a framework of Bayesian (non)linear random effects mediation models, B(N)REMM, to directly model intrinsically linear and nonlinear functions. Specifically, we developed two distinct longitudinal mediation models where all variables under consideration were longitudinal and followed either a linear trend (L-BREMM) or a segmented trend captured by linear–linear piecewise functions with unknown random changepoints (P-BREMM). Additionally, no research has assessed the impact of omitting confounder(s) when modeling mediation effects for intrinsically nonlinear functions. We used an empirical data example from the Early Childhood Longitudinal Study—Kindergarten Cohort to contrast the fit of two models where one included the confounder and the other omitted it. The empirical example illustrated the need to study the impacts of model misspecification with respect to omitting confounder(s). We further explored this issue and its effect on model estimation for both L-BREMM and P-BREMM via Monte Carlo simulation studies under a variety of data conditions. The simulation study results showed that omitting confounder(s) negatively impact parameter recovery for both L-BREMM and P-BREMM but only had an impact on model convergence of P-BREMM. We provide R scripts to estimate both L-BREMM and P-BREMM to aid the dissemination of these models. (PsycInfo Database Record (c) 2024 APA, all rights reserved)