Identification Beyond Standard Criteria

Identifying causal effects, represented as $P(Y | do(X=x))$, can be achieved using various methods. For instance, backdoor and frontdoor adjustment criteria offer solutions for many common scenarios described by Directed Acyclic Graphs (DAGs). However, these criteria only address a subset of identifiable cases. Systems often present complexities like unobserved confounding or intricate causal pathways that standard adjustment criteria cannot directly handle. Structural Causal Models (SCMs) and do-calculus provide a comprehensive toolkit for addressing these more challenging identification problems.

The Completeness of Do-calculus

Recall the three rules of do-calculus, which allow us to manipulate probability distributions involving interventions ($do(\cdot)$ operators) using only observational quantities, provided certain graphical conditions are met.

1. Rule 1 (Insertion/Deletion of Observations): $P(y | do(x), z, w) = P(y | do(x), z)$ if $(Y \perp W | X, Z)_G$, where $G$ is the graph obtained by removing edges into $X$.
2. Rule 2 (Action/Observation Exchange): $P(y | do(x), do(z), w) = P(y | do(x), z, w)$ if $(Y \perp Z | X, W)_{G_{\bar{X}\underline{Z}}}$, where $G_{\bar{X}\underline{Z}}$ is the graph with incoming edges to $X$ removed and outgoing edges from $Z$ removed.
3. Rule 3 (Insertion/Deletion of Actions): $P(y | do(x), do(z), w) = P(y | do(x), w)$ if $(Y \perp Z | X, W)_{G_{\bar{X}, \overline{Z(W)}}}$, where $Z(W)$ are nodes in $Z$ not ancestors of $W$ in $G_{\bar{X}}$.

The significance of do-calculus extends past just providing sufficient conditions; it provides a complete algorithm for nonparametric identification in DAGs. If a causal effect $P(y | do(x))$ can be expressed using the observational distribution $P(V)$ derived from an SCM represented by a DAG $G$, then the do-calculus rules, applied systematically, will derive that expression. If the rules fail to eliminate the $do$-operator, the effect is generally not identifiable from observational data alone without further assumptions.

Consider a scenario slightly more complex than typical backdoor or frontdoor cases:

An unobserved confounder $U$ affects both $X$ and $Y$. $Z_1$ mediates the effect of $X$ on $Y$, while $Z_2$ confounds $X$ and $Y$ but is observed. Standard backdoor fails due to $U$. Frontdoor fails because $X$ is not blocked from $Y$ by $Z_1$ alone (due to the $Z_2 \to Y$ path).

Here, identifying $P(Y|do(X=x))$ requires a careful application of do-calculus rules, often involving conditioning and marginalizing over observed variables like $Z_1$ and $Z_2$ in sequence. For instance, one might try to first adjust for $Z_2$ and then use properties related to the mediator $Z_1$.

Identification in the Presence of Unobserved Confounders

A major challenge arises when unobserved variables confound the relationship between treatment $X$ and outcome $Y$. While do-calculus signals non-identifiability in such cases from observational data alone, specific structures or auxiliary variables can sometimes rescue identification.

Instrumental Variables (IV)

The classic approach is Instrumental Variables (IV). An instrument $Z$ is a variable that affects $X$, but affects $Y$ only through its effect on $X$, and does not share any common causes with $Y$ (other than potentially through $X$).

The canonical IV graph structure. $Z$ allows identification of the effect of $X$ on $Y$ despite the unobserved confounder $U$.

Under certain assumptions (often linearity for simple estimation), the IV estimand allows recovering the causal effect. For example, in a linear, homogeneous effect setting:

$$\beta_{IV} = \frac{Cov(Z, Y)}{Cov(Z, X)}$$

This identifies the average causal effect of $X$ on $Y$. Finding valid instruments in practice is notoriously difficult. They must satisfy the relevance condition ($Z$ causes $X$) and the exclusion restriction ($Z$ affects $Y$ only via $X$ and is independent of $U$). We explore advanced IV methods in Chapter 4.

Proximal Causal Inference

A more recent framework, Proximal Causal Inference, offers an alternative when direct instruments are unavailable but proxies for the unobserved confounders exist. Suppose $U$ is the unobserved confounder between $X$ and $Y$. Proximal inference seeks two observed variables:

• $W$: A "confounding proxy" influenced by $U$ and influencing $X$, but not directly influencing $Y$ except through $U$ (and possibly $X$).
• $Z$: An "outcome proxy" influenced by $U$ and influencing $Y$, but not directly influencing $X$ except through $U$.

Crucially, $W$ and $Z$ must satisfy certain conditional independence assumptions, essentially acting as noisy measurements of the underlying confounding $U$.

Proximal Causal Inference setup. $W$ and $Z$ are observed proxies for the unobserved confounder $U$.

Under specific conditions, particularly solving integral equations (often requiring regularization or kernel methods in practice), proximal methods can identify the causal effect $P(Y|do(X=x))$. For example, one identification strategy involves solving for bridge functions $h(x, z) = E[Y | X=x, Z=z]$ and $g(w, x) = P(X=x | W=w)$ from the observational data, and then relating them via the latent confounder $U$. This is particularly relevant in high-dimensional settings where $U$ might be complex, but high-dimensional proxies $W$ and $Z$ are available (e.g., text embeddings, past user behavior). We explore this in Chapter 4.

Generalized Adjustment Criteria

Identification can sometimes be achieved by adjusting for sets of variables that do not satisfy the standard backdoor criterion, but fulfill alternative graphical conditions. For example, the conditional backdoor criterion might apply if adjustment is valid only within specific strata of other variables. Other criteria might involve adjusting for descendants of the treatment variable under specific graph structures, in addition to the simpler frontdoor setup. These often arise from careful application of do-calculus or analysis of the graph structure to find alternative ways to block confounding paths.

The General Identification Algorithm (ID Algorithm)

For DAGs possibly including unobserved confounders (often represented using bidirected edges in related graphs like Maximal Ancestral Graphs, or MAGs), Tian and Pearl developed a provably complete algorithm, often called the ID algorithm. Given a causal query $P(y | do(x))$ and a causal graph (potentially with hidden variables), the ID algorithm determines if the query is identifiable from the observed joint distribution. If it is, the algorithm outputs the corresponding estimand as a function of the observed distribution; otherwise, it correctly reports non-identifiability.

While the algorithm itself is intricate, involving graph decompositions and recursive application of do-calculus logic, its existence is theoretically significant. It guarantees that any identifiable effect can be found systematically. Implementations exist in libraries like DoWhy, allowing practitioners to check identifiability for complex custom graphs.

Parametric vs. Nonparametric Identification

It's important to distinguish between nonparametric and parametric identification.

• Nonparametric Identification: The effect can be determined from the observational distribution without making assumptions about the functional forms of the relationships between variables (e.g., linearity, additivity). Backdoor, frontdoor, do-calculus, and the ID algorithm primarily target nonparametric identification.
• Parametric Identification: The effect is identifiable only if specific functional form assumptions are made (e.g., linear SCMs, specific distributional forms). Standard IV estimation often relies on linearity assumptions. Identification in cyclic models frequently requires parametric assumptions.

While nonparametric identification is generally preferred for its robustness, parametric assumptions are sometimes necessary, especially when dealing with feedback loops or specific types of unobserved confounding where nonparametric methods fail.

Practical Implications

Moving past standard criteria opens up possibilities for causal inference in more challenging, realistic settings. However, these advanced strategies come with caveats:

• Stronger Assumptions: Methods like IV and Proximal Inference rely on assumptions (e.g., exclusion restrictions, proxy validity) that can be difficult to verify and may be strong.
• Data Requirements: They often require specific auxiliary variables (instruments, proxies) that may not be available.
• Estimation Complexity: Estimators derived from these strategies can be complex, involving multi-stage regression, integral equations, or specialized optimization, often requiring larger sample sizes for reliable estimation.

Therefore, while these methods expand our capabilities, they demand careful consideration of the underlying assumptions and increased attention to sensitivity analysis (covered in the next section) to assess the robustness of conclusions to potential violations of these assumptions. Applying do-calculus systematically or using algorithmic identification tools can provide formal justification, but domain knowledge remains indispensable for validating the plausibility of the graph structure and the required identifying assumptions.