Clusters that are not there: An R tutorial and a Shiny app to quantify a priori inferential risks when using...

Area of Science:

• Social Science Research
• Psychological Research
• Data Analysis

Background:

• Clustering methods are widely used in social sciences to identify distinct individual types and uncover hidden heterogeneity.
• However, the validity of clustering results depends on specific assumptions and conditions.
• Common risks include failing to detect existing clusters or identifying clusters that are not truly present.

Purpose of the Study:

• To investigate the risk of false positive cluster detection in psychological research using common clustering methods.
• To identify specific data conditions that increase the likelihood of detecting non-existent clusters.
• To provide practical tools for researchers to assess inferential risks associated with clustering methods.

Main Methods:

• The study employed simple data simulations to mimic conditions frequently encountered in applied psychological research.
• Simulated data varied in sample size, number of clusters, indicator correlation, and skewness.
• An R tutorial and a Shiny app were developed to aid researchers in quantifying a priori inferential risks.

Main Results:

• Simulations indicated that commonly used clustering methods are prone to detecting clusters that do not exist under typical psychological research conditions.
• Violations of statistical assumptions, often overlooked in psychology, contribute to this over-detection.
• Conditions that inflate the number of detected clusters are prevalent in applied psychological research.

Conclusions:

• Researchers using clustering methods in psychology face a high risk of identifying spurious clusters.
• A preliminary risk assessment using the provided tools is crucial before applying clustering techniques.
• Implementing these checks can improve the reliability and validity of cluster analysis in social science research.