关键词: 反应风格, 项目反应树, 李克特量表, 项目反应理论

Abstract: Item response tree models decompose Likert-type item responses into a sequence of binary decision nodes. This decomposition provides a refined framework for separating response styles from the trait of interest in self-report measures. Recent years have seen substantial extensions to the IRTree framework, yet a systematic survey of these developments, especially regarding their theoretical assumptions and comparative utility, is still lacking. This review fills that gap by tracing the framework’s evolution along two interconnected threads: node-level structural refinements and response-process-level heterogeneity. Based on this systematic examination, we distill a stepwise analytical workflow and provide supporting computational tools to facilitate empirical application.
The node-level thread begins with the standard IRTree, which rests on two implicit assumptions. The first is within-node unidimensionality, meaning each decision node is driven by a single latent variable. The second is the dominance process, meaning the probability of a positive decision increases monotonically with the corresponding latent trait. Subsequent developments have relaxed both assumptions. The first major extension introduces within-node multidimensionality. In this specification, the trait of interest exerts a global influence across all decision nodes, while response styles operate locally on specific nodes, such as extreme or midpoint choices. This configuration addresses the mimicry effect, improving the separation of the trait of interest and response styles, and reduces bias in the parameter estimation. The second major extension incorporates ideal point processes into the IRTree framework. Models that assign ideal point functions to agreement nodes while retaining dominance functions for intensity nodes offer a more nuanced representation of the underlying response mechanism. A further synthesis allows dominance and ideal point processes to co-occur within the same node, accommodating mixed decision mechanisms in a unified structure. These node-level advancements collectively enrich the theoretical specification of what drives each binary choice in the response sequence.
The process-level thread moves beyond the assumption that all respondents follow identical decision paths. Mixture IRTree models identify latent subgroups characterized by distinct response strategy configurations. Recent extensions capture graded differences in the relative weight of the trait of interest versus response styles within subgroups, rather than merely indicating the presence or absence of a style. Multi-path IRTree models account for within-response pathway uncertainty by allowing the same observed category to arise from multiple cognitive routes, such as the acquiescence response style bypassing the regular decision tree. Dynamic IRTree models extend the scope to intra-individual changes. They capture systematic shifts in response style slopes across item positions, reflecting transitions from thoughtful evaluation to satisficing strategies. They also accommodate longitudinal changes through mean drift and autoregressive processes in response style levels. These developments challenge the traditional view of response styles as static person parameters and open new avenues for studying the temporal dynamics of response behavior. Taken together, the process-level thread reveals a gradual expansion from homogeneous and static models to frameworks that accommodate heterogeneity across individuals, across individual responses, and across time.
The article makes three primary contributions. First, it provides a comprehensive review of the fragmented IRTree literature. It clarifies the theoretical assumptions and boundary conditions of each major model variant, offering researchers a coherent map for navigating the expanding toolkit. Second, it distills the reviewed advancements into a stepwise analytical workflow that guides empirical users through three phases: determining the necessity of IRTree modeling and the appropriate response function for the trait of interest, evaluating the added value of within-node multidimensionality, and exploring process-level heterogeneity when theoretically motivated. Third, to bridge the gap between methodological sophistication and practical application, the article provides custom R functions that facilitate rapid implementation of IRTree models for common Likert-scale formats. These functions include flexible constraints for directional invariance and response function choices. An empirical demonstration using a real personality questionnaire illustrates the applicability of this workflow and shows how it assists researchers in model selection and trait estimation.
In conclusion, the advanced IRTree framework reflects two complementary perspectives. It controls for response styles as measurement noise, and it treats response styles as psychologically meaningful information about cognitive strategies. Future directions include refining the framework for scales with more than five categories, establishing clear guidelines for parameter constraints and sample size requirements, developing methods for detecting differential item functioning within IRTree nodes, and extending IRTree models toward a unified framework for modeling complex response behaviors in self-report instruments.

Key words: response styles, item response tree, Likert scale, item response theory