Assessing Dimensionality in Dichotomous Items When Many Subjects Have All-Zero Responses: An Example From Psychiatry...

Area of Science:

• Psychometrics
• Statistics
• Mental Health Research

Background:

• Determining latent dimensions in item sets typically uses eigenvalue analysis or factor analysis fit statistics.
• These methods can inaccurately estimate dimensionality, especially with samples exhibiting a high proportion of all-zero responses (e.g., no symptoms endorsed).

Purpose of the Study:

• To demonstrate how common dimensionality diagnostics can under- or over-estimate true latent variable dimensionality.
• To introduce and validate an all-zero inflated exploratory factor analysis (AZ-EFA) model for improved dimensionality assessment in zero-inflated data.

Main Methods:

• Simulated data experiments were conducted to evaluate common dimensionality diagnostics.
• An all-zero inflated exploratory factor analysis (AZ-EFA) model was developed and applied.
• The AZ-EFA approach was tested using simulation data and a real-world social anxiety disorder dataset.

Main Results:

• Empirical assessments of dimensionality frequently misestimate the number of latent dimensions when zero-inflated response patterns are prevalent.
• Simulation experiments confirmed that common diagnostics can lead to under- or over-estimation of true dimensionality.
• The proposed AZ-EFA model demonstrated improved accuracy in assessing dimensionality for the subgroup with measurable traits.

Conclusions:

• Standard dimensionality assessment methods are unreliable with substantial zero-inflated data, potentially leading to incorrect conclusions about underlying latent structures.
• The AZ-EFA model offers a more accurate approach for analyzing dimensionality in datasets with a high prevalence of all-zero responses.
• Findings highlight the importance of considering zero-inflation in psychometric analyses and may explain discrepancies between community and patient population studies.