Causal invariance treats each conditional distribution in a causal model as a distinct real-world mechanism. When an environment changes one mechanism, only its corresponding causal kernel should change; kernels for other mechanisms should remain stable. section 3.2.3
This differs from relying on correlations that may shift with the data distribution. A model that captures an invariant causal relationship can remain useful under deployment drift, because it represents how variables are generated rather than merely how they co-occur in training data. section 1.3.3
Causal invariance supports several practical reasoning moves:
- Localize change: determine which causal mechanism an environmental change affects instead of refitting every model component. section 3.2.3
- Estimate parameters from domain knowledge and population statistics, using assumptions about the mechanism rather than statistical learning alone. section 3.2.3
- Prefer causal structure when prediction must transfer across environments, such as predicting temperature from elevation in cities with different elevation profiles. section 1.3.3
Transportation mechanisms
Suppose residence R depends on education E, while temperature T depends on observed variables O and residence R. A change to the mechanism producing R should alter P(R|E) without changing the parameters of P(T|O,R). section 3.2.3
