## Review of Causal-Learn Independence Tests Documentation I reviewed the **causal-learn** documentation on independence tests, specifically the page at [causal-learn.readthedocs.io](https://causal-learn.readthedocs.io/en/latest/independence_tests_index/index.html). Here is a summary of the content and an explanation of why **conditional independence** is used instead of just independence in causal inference. ### 1. Overview of Independence Tests in Causal-Learn The documentation introduces several **(conditional) independence tests** used in causal discovery algorithms. These tests are fundamental for constraint-based methods like the PC, FCI, and CD-NOD algorithms. **Available Tests:** - **Fisher-Z test**: For linear-Gaussian data (tests conditional independence via partial correlations). - **Missing-value Fisher-Z test**: Handles missing data using test-wise deletion. - **Chi-Square test**: For discrete/categorical data (tests independence based on observed vs. expected frequencies). - **Kernel-based Conditional Independence (KCI) test**: Non-parametric test for non-linear relationships. - **G-Square test**: Log-likelihood ratio test for discrete data. Each test is implemented in a unified framework (`CIT` class) and supports caching to avoid redundant computations in iterative algorithms. ### 2. Why Conditional Independence vs. Just Independence? #### **Unconditional Independence** - **Definition**: Two variables \(X\) and \(Y\) are unconditionally independent if knowing \(Y\) gives no information about \(X\), i.e., \(P(X|Y) = P(X)\) or \(P(X,Y) = P(X)P(Y)\). - **Limitation in Causal Inference**: Unconditional independence alone cannot distinguish between causal and non-causal relationships. For example, two variables may be correlated due to a common cause (confounding) or direct causation, but unconditional tests cannot separate these scenarios. #### **Conditional Independence** - **Definition**: Two variables \(X\) and \(Y\) are conditionally independent given a set of variables \(Z\) if, once \(Z\) is known, \(Y\) provides no additional information about \(X\), i.e., \(P(X|Y,Z) = P(X|Z)\). This is denoted \(X \perp\!\!\!\perp Y \mid Z\). - **Why It’s Crucial for Causal Discovery**: 1. **Distinguishing Causal Structures**: Conditional independence tests allow algorithms to infer causal relationships by checking whether variables are independent after conditioning on potential confounders or mediators. For example, in a causal graph \(X \rightarrow Y \leftarrow Z\), \(X\) and \(Z\) are unconditionally dependent but conditionally independent given \(Y\). 2. **Handling Confounding**: In the presence of confounders (variables that affect both \(X\) and \(Y\)), unconditional tests may falsely suggest a direct causal link. Conditioning on the confounder can reveal the true independence structure. 3. **Markov Property and Causal Graphs**: In causal graphical models (e.g., Bayesian networks), conditional independence corresponds to d-separation, a key concept for encoding causal assumptions. Algorithms like PC use conditional independence tests to build causal graphs from data. 4. **Interventional vs. Observational Data**: Conditional independence helps distinguish between observational associations and causal effects. For instance, in a chain \(X \rightarrow M \rightarrow Y\), \(X\) and \(Y\) are conditionally independent given \(M\), indicating mediation. #### **Example from the Documentation** The Fisher-Z test, for instance, computes **partial correlations** to test conditional independence in linear-Gaussian models. This is essential for algorithms that need to determine whether a relationship is direct or mediated by other variables. ### 3. Key Takeaways from the Documentation - **Conditional independence tests are the building blocks** of constraint-based causal discovery algorithms. - **Different tests are suited for different data types**: Fisher-Z for linear-Gaussian, KCI for non-linear, Chi-Square for discrete data. - **Conditional independence enables causal inference** by allowing algorithms to rule out spurious associations and identify true causal pathways. ### 4. Further Reading - **Wikipedia on Conditional Independence**: [Conditional independence](https://en.wikipedia.org/wiki/Conditional_independence) - **Oxford Statistics Chapter on Conditional Independence**: [Chapter 4: Conditional Independence](https://www.stats.ox.ac.uk/~evans/APTS/ci.html) – Provides a mathematical foundation and examples. In summary, causal-learn emphasizes conditional independence tests because they are essential for uncovering causal relationships from observational data, distinguishing them from mere statistical associations.