Multivariate modelling of responses to conditional items: New possibilities for latent class analysis.

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Questionnaire data frequently exhibit missing values, particularly for conditional items not applicable to all respondents.
• Traditional analysis often restricts conditional item analysis to subpopulations, potentially limiting insights.
• Missingness in conditional items arises from non-existent features, not survey design flaws.

Purpose of the Study:

• To develop and demonstrate a joint multivariate modeling approach for questionnaire data incorporating both unconditional and conditional items.
• To enable analysis of conditional item data without restricting the study population.
• To explore the application of this approach within latent class modeling.

Main Methods:

• Joint multivariate modeling integrating conditional item structures into statistical models.
• Application to latent class analysis.
• Utilizing common parameters to model distributions across the entire population.

Main Results:

• The proposed method allows inference from both unconditional data (entire population) and conditional data (relevant subjects).
• This approach enhances multivariate analysis possibilities for complex questionnaire data.
• Demonstrated utility using respiratory symptom data (wheeze, cough) in children.

Conclusions:

• Joint modeling of unconditional and conditional questionnaire items offers a more comprehensive analytical framework.
• This methodology is particularly valuable in medical research where conditional data structures are prevalent.
• The approach is adaptable to various multivariate models beyond latent class analysis.