Similarity-Based Predictive Models: Sensitivity Analysis and a Biological Application with Multi-Attributes.

Area of Science:

• Data Science
• Computational Biology
• Statistical Modeling

Background:

• Empirical similarity predictive models are crucial in biology and data science.
• Handling categorical data is a key challenge in these models.
• Two primary strategies exist: retaining original variables or converting to binary variables.

Purpose of the Study:

• To perform a sensitivity analysis comparing two strategies for handling categorical covariates in empirical similarity-based predictive models.
• To evaluate the performance of these strategies using computational simulations.
• To apply the findings to a biological data context.

Main Methods:

• Computational simulations were used to analyze the sensitivity of two strategies for handling categorical variables.
• A linear regression model served as a reference, with two parameter estimation methods.
• Similarity functions included exponential and fractional inverse types; sensitivity was measured by the coefficient of variation.

Main Results:

• The first strategy, retaining categorical variables with assigned weights, demonstrated superior performance over the second strategy (binary variable conversion).
• The first strategy showed greater parsimony, requiring fewer parameters for effective modeling.
• Relative variability of parameter estimators was lower with the first strategy.

Conclusions:

• Retaining original categorical variables is a more effective and parsimonious approach for empirical similarity-based predictive models in biological datasets.
• This method offers improved parameter estimation stability compared to binary conversion.
• The findings provide practical guidance for data scientists and biologists working with categorical data.