AR(1) latent class models for longitudinal count data.

Area of Science:

• Developmental psychopathology
• Biostatistics
• Longitudinal data analysis

Background:

• Identifying natural groupings in longitudinal data is crucial for understanding subject development.
• Existing methods often face computational challenges with count data and random effects models.
• Conduct problems in developmental psychopathology require robust methods for trajectory clustering.

Purpose of the Study:

• To propose a novel statistical method for clustering longitudinal count data.
• To address the specific need for recovering clusters of conduct problem trajectories.
• To offer a computationally efficient alternative to existing models.

Main Methods:

• Utilizing a first-order autoregressive process suitable for count data.
• Developing a closed-form class-specific likelihood function to avoid computational issues.
• Implementing an approximate Expectation-Maximization (EM) procedure for parameter estimation.
• Validating the method through simulations using a four-class model.

Main Results:

• The proposed method effectively recovers latent trajectories in simulated data.
• Simulations demonstrate the procedure's effectiveness, particularly in trajectory recovery.
• The method was successfully applied to analyze conduct problem trajectories in a national sample.

Conclusions:

• The developed method provides an efficient and effective way to cluster longitudinal count data.
• This approach is particularly valuable for analyzing developmental trajectories in fields like psychopathology.
• The R package 'inarmix' is available for implementing these statistical procedures.