Aggregate-Constrained Component Intervention (ACID)

• ACID is a formal framework that defines the joint interventional distribution for aggregate treatments, decomposing them into component-level interventions under strict aggregation constraints.
• It provides methodological guidance in instrumental variable analysis and causal inference by specifying how interventions at the component level impact outcomes.
• The framework aids in identifying causal effects and supports algorithms for mixture recovery in Bayesian networks, highlighting its practical and theoretical significance.

The Aggregate-Constrained Component Intervention Distribution (ACID) formalizes how interventions on an aggregate-level treatment—composed of underlying, finer-grained components—are instantiated at the component level within a structural causal model. In contexts such as instrumental variable (IV) analysis or causal inference in Bayesian networks, ACID provides the precise interventional joint law that bridges the aggregate variable with its constituent parts under constraints enforced by the aggregation rule. Proper interpretation or identification of causal effects for aggregate treatments (e.g., GDP, total caloric intake) critically depends on specifying the ACID, especially when component-level heterogeneity is present (Tsao et al., 17 Jan 2026).

1. Formal Definition and Factorization

Given an aggregate treatment variable $A$ formed by

$A = \sum_{j=1}^k \alpha_j\,A_j$

for unobserved components $\{A_j\}$, the ACID is the joint interventional distribution $P(\mathbf v)_{do(A=a)}$ under the "surgery" $do(A = a)$ (Tsao et al., 17 Jan 2026). Here, $\mathbf{V} = \{I, U, A_1, ..., A_k, Y\}$, where $I$ denotes an instrument, $U$ optional confounders, and $Y$ the outcome.

Three canonical desiderata underlie the ACID construction:

• Surgicality: The intervention severs all incoming edges into each $A_j$.
• Aggregation Constraint: The values $A = \sum_{j=1}^k \alpha_j\,A_j$0 must lie on the hyperplane defined by $A = \sum_{j=1}^k \alpha_j\,A_j$1.
• Value-Independence: The effect of a one-unit change in $A = \sum_{j=1}^k \alpha_j\,A_j$2 does not depend on the specific level $A = \sum_{j=1}^k \alpha_j\,A_j$3.

Under these, ACID factorizes as: $A = \sum_{j=1}^k \alpha_j\,A_j$4 where the ACID kernel $A = \sum_{j=1}^k \alpha_j\,A_j$5 encodes the interventional mechanism for the components subject to the aggregation constraint.

2. Role in Interpreting Aggregate Interventions

Various choices for the ACID kernel $A = \sum_{j=1}^k \alpha_j\,A_j$6 correspond to distinct "versions" of an intervention on $A = \sum_{j=1}^k \alpha_j\,A_j$7, e.g., increasing just one $A = \sum_{j=1}^k \alpha_j\,A_j$8, shifting all $A = \sum_{j=1}^k \alpha_j\,A_j$9 proportionally, or imposing some Gaussian law over $\{A_j\}$0 conditioned on their sum. Calculation of the aggregate causal effect proceeds only once $\{A_j\}$1 is made explicit: $\{A_j\}$2

$\{A_j\}$3

When value-independence holds, the aggregate causal effect is invariant in $\{A_j\}$4. If not, the effect can vary as a function of the aggregate level itself (Tsao et al., 17 Jan 2026).

3. Conditions for Alignment with Instrumental Variable Estimation

The standard IV estimand in a linear-Gaussian structural causal model (SCM) with components $\{A_j\}$5 and

$\{A_j\}$6

is

$\{A_j\}$7

Two main sufficient conditions ensure $\{A_j\}$8:

• Proportional Aggregation: If $\{A_j\}$9 is the same for all $P(\mathbf v)_{do(A=a)}$0, any ACID yields $P(\mathbf v)_{do(A=a)}$1. This collapses all component pathways to one aggregate treatment effect.
• Instrument-Tuned Interventions: Choosing a Gaussian ACID kernel where the mean shift vector $P(\mathbf v)_{do(A=a)}$2 matches the instrument's structure ($P(\mathbf v)_{do(A=a)}$3) ensures that the IV estimand matches the causal effect. This construction is highly contrived and relies on exact knowledge of both instrument and outcome coefficients.

If neither proportional aggregation nor instrument-tuned ACID is posited, the exclusion restriction is generally violated in the mis-specified single-treatment IV model: the instrument will induce a direct effect on the outcome not solely mediated via the aggregate variable (Tsao et al., 17 Jan 2026).

4. Illustrative Examples and Counterexamples

• Gaussian ACID: The joint component distribution is multivariate normal conditioned to satisfy $P(\mathbf v)_{do(A=a)}$4, with appropriate mean and covariance constraints. Here, $P(\mathbf v)_{do(A=a)}$5 depends on the vector $P(\mathbf v)_{do(A=a)}$6.
• Counterexample to Value-Independence: Assigning a uniform distribution to $P(\mathbf v)_{do(A=a)}$7 over a rectangle under the aggregation constraint demonstrates the failure of value-independence: $P(\mathbf v)_{do(A=a)}$8 is not constant in $P(\mathbf v)_{do(A=a)}$9 (Tsao et al., 17 Jan 2026).

5. Practical Algorithms for Mixture Recovery in Causal Bayesian Networks

A related aggregate-constrained mixture problem arises in causal Bayesian networks, where observed data may derive from a mixture of interventions. With a known DAG $do(A = a)$0 and access to all single-node interventional marginals $do(A = a)$1, one seeks the mixture weights $do(A = a)$2 in the expansion

$do(A = a)$3

Identifiability is achieved via the "exclusion per node" assumption: for each node, at least one possible intervention does not participate in the mixture.

An explicit recovery algorithm, based on solving a sequence of linear systems, computes the unique mixture weights in polynomial time provided exact marginals are known. In practical scenarios, when only empirical estimates are available, a constrained quadratic program—solved via SLSQP—provides robust estimation of the mixture weights (Sinha et al., 2019). Performance is validated empirically on benchmark networks (e.g., ALARM) and proprietary e-commerce datasets.

6. Implications, Limitations, and Extensions

The ACID framework exposes fundamental ambiguities in defining and estimating the causal effect of aggregate treatments:

• The choice of ACID kernel $do(A = a)$4 is often underdetermined unless justified by substantive knowledge or additional assumptions. Without such specification, the meaning of $do(A = a)$5 is ambiguous.
• Standard IV conditions—relevance, exchangeability, exclusion—are insufficient when aggregation-induced heterogeneity exists. Either proportional aggregation or an explicit defense of the chosen ACID kernel is necessary for policy-relevant inference (Tsao et al., 17 Jan 2026).
• In the Bayesian network mixture context, identifiability requires full knowledge of the DAG and all component-wise interventional marginals, along with the exclusion assumption. Extensions to continuous, latent, or multi-node interventions remain open problems (Sinha et al., 2019).

Summary Table: ACID Key Elements and Consequences

Principle	Description	Consequence
Surgicality	Severs all arrows into each $do(A = a)$6 under intervention	Ensures valid "surgery" semantics
Aggregation constraint	Constrains $do(A = a)$7 so $do(A = a)$8	Restricts support of intervention law
Value-independence	Effect per unit of $do(A = a)$9 does not depend on $\mathbf{V} = \{I, U, A_1, ..., A_k, Y\}$0	Ensures constancy of causal effect
Proportional aggregation	$\mathbf{V} = \{I, U, A_1, ..., A_k, Y\}$1 equal for all $\mathbf{V} = \{I, U, A_1, ..., A_k, Y\}$2	IV equals causal effect for any ACID
Instrument-tuned ACID	Kernel mean vector matches effect of instrument on components	Forces IV = ACE by construction

The specification of the ACID is central to any rigorous causal analysis involving aggregate treatments and crucial in justifying the interpretation of instrumental variable estimands in such settings (Tsao et al., 17 Jan 2026).