It's difficult to measure the causal effect of actions or interventions. Here, I present three different causal machine learning frameworks that we can use to estimate the effect of a treatment--in order of the simplest to the most complex.

An introduction to SHAP values and how they can be used to understand machine learning model predictions. Using a small climbing competition example, I show how SHAP differs from feature importance, how to read common SHAP plots, and a few things to watch out for when interpreting the results.

The one in which the author writes a surprisingly long introduction to multilevel modelling (with math and all) and then proceeds to show how to fit one in both R and Julia with frequentist and Bayesian methods, respectively, despite not being very proficient in either language (because every company she has ever worked at has been python-houses).

If data is noisy, or we want to capture an underlying trend without assuming a fixed global shape, we can use smoothing methods. Here, I describe one such method and briefly discuss its usage in more complex cases than single-variable smoothing.

An introduction to the Metropolis-Hastings algorithm, one of the core methods behind Bayesian inference via MCMC. I explore how it works, why it works, and how to implement it — with references for further learning.

I derive the maximum a posteriori (MAP) estimator and show how, under different prior choices, MAP estimation leads naturally to L2-regularization (Ridge regression) with a normal prior and L1-regularization (LASSO regression) with a Laplace prior.

I provide a complete Bayesian analysis for a standard linear regression model with a Normal-inverse-gamma prior, demonstrating how the conjugate structure yields closed-form expressions for both the posterior distribution and the posterior predictive.

I show that the Normal-inverse-gamma distribution is a conjugate prior to the normal distribution with unknown mean and variance.

I give an introduction to Bayesian inference and Bayes' Theorem, and derive the posterior of the binomial model under a Beta prior, showing that the Beta prior is conjugate in the process.

I explain the difference between ROC and precision-recall curves, highlighting their use in evaluating classifiers with imbalanced data.

I show how to extend PySpark’s built-in cross-validation to support time-series data, with custom folds and windowing strategies.