Presentations

Robust regression for noisy biological data

Nonparametric

Regression

Biological variables are often non-normally distributed, contain outliers, may have lower limits of detection, excessive zeros or multiple modes. The proportional odds (PO) regression model can easily handle these issues and is a robust alternative to linear regression and other parametric models. Indeed, the PO model can be thought of as an extension of the robust Wilcoxon rank-sum test that can adjust for covariates.

Linear regression diagnostics

Regression

This presentation introduces linear regression, covering its core assumptions, diagnostic techniques for assessing them, and strategies for addressing violations.

Automatic feature selection and engineering with MARS

Variable selection

Machine Learning

Nonparametric

Multivariate adaptive regression splines (MARS) is a nonparametric machine learning method that automatically identifies which features are important for predicting the target variable (feature selection) as well as what forms the features should take (feature engineering), i.e. whether a feature should have a linear, non-linear, additive, or interaction effect. It is fairly fast method for working with big data and produces an interpretable model (not a black box).

Causal inference with observational data using propensity score matching

Causal inference

This presentation introduces propensity score matching as a method for attempting causal inference with observational data.

Introduction to Bayesian statistics

Bayesian statistics

An introduction to the basics of Bayesian statistics, including Bayes’ Theorem, the 3 key aspects of Bayesian analysis (prior, model and posterior), and a comparison with traditional Frequentist statistics.

Triplet matching: propensity score matching with 3 groups

Causal inference

Propensity score matching is a popular method for attempting causal inference with observational data. However, the vast majority of implementations focus on matching only two groups for a binary exposure variable. This presentation introduces triplet matching, a generalization of propensity score matching to three groups.

Multi-omics data integration

Multi-omics

Multi-omics data integration is the process of combining data from different omics platforms (e.g. genomics, transcriptomics, epigenomics) to gain a more comprehensive understanding of biological systems. This presentation explains how to download multi-omics data from the TCGA database. Then dimension reduction methods are considered (PCA, PLS and CCA), and are shown to identify groups of correlated features. Lastly, two machine learning methods are compared for using multi-omic features to predict survival in kidney cancer patients.

Interpretable machine learning

Machine Learning

Many machine learning models are considered “black boxes”: data is input and predictions are output, but we have little insight into how the model makes its decisions. This presentation covers methods for interpreting models, including visualizing the relationships between predictors and outcome, variable importance scores, interaction scores to identify predictors that interact, and methods to explain how a model makes a prediction for a given subject (LIME and Shapley values).