A neural network analysis of Lifeways cross-generation imputed data

Area of Science:

• Biostatistics
• Genetics
• Computational Biology

Background:

• The Lifeways cross-generation study investigates factors influencing child development, including body mass index (BMI).
• Predicting child BMI from parental and grandparental data is complex due to potential nonlinear relationships and missing data.

Purpose of the Study:

• To evaluate the efficacy of neural networks in predicting child BMI using data from parents and grandparents.
• To compare the predictive performance of neural networks against traditional linear models.
• To assess the impact of various data imputation methods on prediction accuracy.

Main Methods:

• Application of neural network models for nonlinear regression analysis.
• Utilizing linear models for comparative prediction.
• Implementing and comparing different data imputation techniques for handling missing values.
• Performing analyses on both imputed and non-imputed datasets to establish a gold standard.

Main Results:

• Neural network models demonstrated superior performance compared to linear models in predicting child BMI.
• Child BMI could be predicted from family member data with an approximate accuracy of 2.84 units using neural networks.
• Different data imputation methods yielded similar mean squared errors, and imputed data performance was comparable to non-imputed data.

Conclusions:

• Neural networks are a valuable tool for detecting nonlinear and interaction effects in complex datasets.
• The predictive accuracy of child BMI from familial data is achievable using advanced statistical modeling.
• The chosen imputation methods had minimal impact on the overall prediction results, suggesting robustness in handling missing data.