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

SIMPLE SUMMARY: In this study, we perform a sensitivity analysis in similarity-based predictive models using computational simulations and two distinct methodologies, while focusing on a biological application. We utilize a linear regression model as a reference point. We gauge sensitivity by calculating the coefficient of variation of the parameter estimators from three different models. Our findings show that the first approach outperforms the second one when dealing with categorical variables. Moreover, this first approach offers the advantage of being more parsimonious due to a smaller number of parameters. ABSTRACT: Predictive models based on empirical similarity are instrumental in biology and data science, where the premise is to measure the likeness of one observation with others in the same dataset. Biological datasets often encompass data that can be categorized. When using empirical similarity-based predictive models, two strategies for handling categorical covariates exist. The first strategy retains categorical covariates in their original form, applying distance measures and allocating weights to each covariate. In contrast, the second strategy creates binary variables, representing each variable level independently, and computes similarity measures solely through the Euclidean distance. This study performs a sensitivity analysis of these two strategies using computational simulations, and applies the results to a biological context. We use a linear regression model as a reference point, and consider two methods for estimating the model parameters, alongside exponential and fractional inverse similarity functions. The sensitivity is evaluated by determining the coefficient of variation of the parameter estimators across the three models as a measure of relative variability. Our results suggest that the first strategy excels over the second one in effectively dealing with categorical variables, and offers greater parsimony due to the use of fewer parameters.