Reflection on modern methods: good practices for applied statistical learning in epidemiology.

Area of Science:

• Public Health
• Epidemiology
• Data Science

Background:

• Statistical learning methods are increasingly used in public health and epidemiology for complex, high-dimensional data.
• These advanced methods, including shrinkage and kernel smoothing, often outperform traditional statistical models in challenging data settings.
• However, the use of random sampling in these novel methods can introduce variability into research findings.

Purpose of the Study:

• To assess the impact of random number generation seeds on the results of statistical learning models.
• To evaluate the robustness and interpretability of common statistical learning methods used in environmental health research.
• To provide best practice recommendations for data science in public health.

Main Methods:

• Four statistical learning models were applied to analyze the relationship between environmental mixtures and health outcomes.
• Each model was run 100 times with different random number generator seeds to assess result variability.
• Estimation and inference variability were systematically evaluated across different seeds and models.

Main Results:

• All tested statistical learning methods demonstrated some degree of variability dependent on the random seed used.
• The extent of this seed-dependent variability differed among the methods and specific environmental exposures.
• This highlights inherent seed sensitivity in statistical learning approaches that rely on random sampling.

Conclusions:

• Statistical learning methods employing random seeds are susceptible to variability.
• Researchers must conduct sensitivity analyses using multiple seeds to ensure the robustness and interpretability of their findings.
• Implementing these best practices enhances the reliability of results in public health and epidemiological studies.