Previous chapters focused on identifying and estimating causal effects under the assumption that all relevant confounding variables $Z$ influencing both treatment $T$ and outcome $Y$ are observed. However, in many practical machine learning applications, this assumption is often violated. Unmeasured confounders or systematic differences between the groups being compared (selection bias) can lead to incorrect causal conclusions.

This chapter addresses this fundamental challenge. We will examine several advanced techniques specifically designed to estimate causal effects when critical variables are unobserved. You will learn about:

• Instrumental Variables (IV): How to use a third variable, an instrument, that affects the treatment but not the outcome directly (except through the treatment) to isolate the causal effect. We'll cover modern implementations like Deep IV and Kernel IV.
• Regression Discontinuity Designs (RDD): Exploiting sharp or fuzzy assignment thresholds to estimate local treatment effects.
• Difference-in-Differences (DiD): Analyzing panel data to compare changes over time between treated and control groups, controlling for time-invariant unobserved confounders.
• Proximal Causal Inference: A newer framework using proxy variables related to the unobserved confounders to achieve identification.
• Selection Bias Correction Methods: Techniques inspired by Heckman correction adapted for machine learning contexts.

By the end of this chapter, you will understand the principles behind these methods and be able to apply them to mitigate bias arising from unobserved factors in your causal analyses.