Federated generalized additive models for location, scale and shape.

Area of Science:

• Biostatistics
• Computational Statistics
• Data Privacy

Background:

• Generalized Additive Models for Location, Scale and Shape (GAMLSS) are crucial for clinical parameter percentile curve estimation in pediatrics.
• International percentile curves necessitate large, diverse datasets, often created by pooling multi-study data.
• Ethical and legal restrictions on sharing sensitive individual-level data hinder data pooling.

Purpose of the Study:

• To develop a privacy-enhancing federated method for fitting GAMLSS models.
• To enable co-analysis of distributed data without physical transfer, addressing data privacy concerns.

Main Methods:

• Developed a federated GAMLSS algorithm enabling distributed data analysis via DataSHIELD.
• Implemented the federated algorithm in R and evaluated its theoretical properties.
• Applied the federated GAMLSS to separated epidemiological data and compared results with pooled data analysis.

Main Results:

• Theoretically and practically demonstrated that federated GAMLSS yields results virtually identical to pooled GAMLSS.
• Identified minor numerical differences between federated and pooled GAMLSS results.
• Observed a runtime increase of over 1000x for the federated approach compared to the pooled GAMLSS.

Conclusions:

• Propose a novel privacy-enhancing federated GAMLSS method.
• This approach allows for fitting GAMLSS models on distributed datasets without compromising data privacy.
• The federated GAMLSS provides a viable alternative for deriving population-level statistics when data pooling is not feasible.