Causal Inference Under Variable Aggregation

27 concepts10 papers1 proposal

In data science and machine learning, we rarely model the world at the level of its finest constituent parts. Instead, we naturally work with macro-variables: we aggregate individual product sales into "total revenue," group genes into "biological pathways," or pool graph nodes into "community embeddings." It is tempting to assume that if a causal relationship exists at the micro-level, it will cleanly manifest at the macro-level.

However, a growing body of research demonstrates that causal properties do not simply scale up. When we group variables together, the fundamental mathematical assumptions that underpin causal discovery and inference—such as the Causal Markov and Faithfulness assumptions—frequently break down. Understanding why this happens is crucial for drawing reliable conclusions from aggregated data.

Why Aggregation Destroys Faithfulness

To understand why aggregation breaks causal models, consider the concept of Causal Faithfulness. Faithfulness assumes that if two variables are statistically independent, there is no causal path between them.

When individual variables are summed or averaged into macro-variables, this assumption frequently fails. As shown by Wahl et al., aggregation can cause opposing micro-level causal effects to cancel each other out. For instance, if sub-variable $A_1$ positively influences $B_1$, but sub-variable $A_2$ negatively influences $B_2$, the aggregate macro-variable $A$ may appear completely independent of the aggregate macro-variable $B$.

Standard causal discovery algorithms interpret this statistical independence as an absence of a causal link, mistakenly deleting valid edges from the graph. Early work by Danks et al. highlighted that conditional independence relationships observed at the individual level are routinely distorted when data is summed or averaged. More recently, Gao et al. mathematically proved that aggregating graph components into single variables inherently violates both the Causal Markov and Faithfulness assumptions, invalidating standard causal modeling in graph representation learning pipelines.

Paradoxical Confounding and Policy Ambiguity

Aggregation does not just hide existing causal links; it can actively invert them. Janzing et al. uncovered a phenomenon termed paradoxical confounding, where the aggregation process itself introduces or removes confounding. A relationship that is entirely unconfounded at the micro-level can appear heavily confounded at the macro-level, and vice versa. Consequently, macro-level causal claims cannot be taken at face value without explicitly modeling the aggregation mechanism.

This distortion creates severe ambiguity when trying to estimate the effect of a policy intervention. As Tsao et al. demonstrate in the context of Instrumental Variable (IV) estimation, the causal effect of an aggregate treatment (such as "Education" or "GDP") is fundamentally ill-defined. Because an aggregate variable is composed of multiple sub-components, the macro-level outcome depends entirely on how an intervention is distributed across those components. If a policy shifts component $X_1$ instead of $X_2$, the resulting macro-effect may be completely different, rendering a single aggregate causal estimate meaningless without micro-level specification. Similarly, Tramontano et al. note that pooling data across different interventional experiments can mask causal directions if the interventions alter the sign or direction of the underlying effects.

The Topological Trap: Dual Collider and Non-Collider Properties

Why do standard graphical models fail so spectacularly when applied to groups of variables? The root of the problem lies in the internal "wiring" of the clusters.

In a standard Directed Acyclic Graph (DAG), a node acts as either a collider ($X \rightarrow Y \leftarrow Z$) or a non-collider ($X \rightarrow Y \rightarrow Z$). These properties dictate how statistical information flows when we condition on a variable (d-separation). Conditioning on a non-collider blocks the flow of information, while conditioning on a collider opens it.

However, as Anand et al. demonstrate, treating a group of variables as a single aggregate node creates a topological paradox: a single cluster can simultaneously exhibit the properties of both a collider and a non-collider. Because some internal variables within the cluster act as colliders while others act as mediators, standard d-separation rules break down. Applying traditional causal discovery logic directly to these aggregate nodes results in incorrect edge orientations and false structural conclusions.

Restoring Validity: Cluster DAGs and Groupwise Faithfulness

To resolve this granularity crisis, researchers have developed frameworks designed specifically for macro-level variables.

One approach is to relax standard faithfulness in favor of groupwise faithfulness. Parviainen et al. show that while learning dependencies between variable groups is significantly more complex due to hidden feedback loops, reliable group-level graphs can be recovered by first modeling individual interactions before aggregating them. In subsequent work, Parviainen et al. formalized that Bayesian networks are not closed under aggregation, meaning a distribution faithful at the individual level becomes unfaithful when clustered, necessitating specialized group-level structure learning algorithms.

Alternatively, researchers can use Cluster Causal Diagrams (C-DAGs). As proposed by Niu et al., grouping variables into clusters allows models to capture complex, synergistic interactions that traditional pairwise methods miss, restoring the validity of causal assumptions at the subsystem level.

However, this abstraction comes with a fundamental trade-off between model simplification and causal identifiability. Anand et al. demonstrate that if a cluster is too coarse—for instance, if a treatment variable and its confounder are grouped into the same macro-node—the causal effect becomes completely unidentifiable. To address this, the authors introduced the $\alpha$C-DAG framework (Anand et al.), providing a theoretically sound graphical language to scale causal discovery to high-dimensional systems without losing the ability to perform valid interventional and counterfactual reasoning.

Key Foundational Work

Linearity Properties of Bayes Nets with Binary Variables (2001) — contributed 3 concept elements to this direction.
Bayesian Networks for Variable Groups (2016) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_aggregation_induced_confounding_dcc49017,
  title = {aggregation-induced confounding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/aggregation-induced-confounding/rdr_dcc4901780a2420685452343a2bd6679},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_aggregation_induced_confounding_dcc49017_v1,
  title = {aggregation-induced confounding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/aggregation-induced-confounding/rdr_dcc4901780a2420685452343a2bd6679/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake