Orientability of Causal Relations in Time Series using Summary Causal Graphs and Faithful Distributions

This paper establishes theoretical conditions under which micro-level causal edges in time series can be uniquely oriented by leveraging high-level summary causal graphs and assuming a faithful, causally sufficient distribution, even in the presence of macro-level cycles or bidirected edges.

The Gist

Imagine you are trying to understand a complex, high-speed dance between three friends: Alex, Blake, and Casey. You can only see them from a distance, so you can't tell exactly who nudged whom at any specific second. You just see that they are all moving together.

In the world of data science, this is the challenge of Time Series Analysis. We have data points (like stock prices, heart rates, or weather patterns) changing over time, and we want to know: Did A cause B, or did B cause A?

This paper introduces a clever way to solve this puzzle by combining expert guesses with mathematical rules. Here is the breakdown in simple terms:

1. The Two Maps: The "Zoomed-Out" vs. The "Zoomed-In"

To understand the dance, the researchers use two different maps:

• The "Summary Map" (The SCG): This is like a high-level sketch drawn by an expert. It says, "Alex and Blake influence each other," and "Casey influences Alex." It's a Summary Causal Graph. It's great because experts can draw it quickly, but it's vague. It doesn't say when the influence happens or if it's a one-way street or a two-way street. It's like saying, "The traffic in New York and London affects each other," without knowing which specific cars are involved.
• The "Full-Time Map" (The FT-DAG): This is the super-detailed, second-by-second video of the dance. It shows exactly who nudged whom at 1:00 PM, 1:01 PM, etc. This is the "truth," but it's incredibly hard to draw from scratch because there are too many details.

The Problem: We have the "Summary Map" (expert knowledge), but we want to figure out the "Full-Time Map" (the truth). We want to know: If Alex nudges Blake, does it happen instantly, or with a delay?

2. The Detective's Toolkit: The "tPC" Algorithm

Usually, to get the detailed map, we use a computer algorithm (like a detective) called tPC. It looks at the data and tries to figure out the direction of the arrows.

However, this detective has a weakness: sometimes the data is too noisy, and the detective gets stuck. It might say, "Alex and Blake are connected, but I don't know who pushed whom." The arrow remains unoriented (it looks like a line with no arrowhead: Alex — Blake).

3. The Big Discovery: Using the Summary Map to Fix the Detective

The authors of this paper asked a brilliant question: "Can we use the expert's 'Summary Map' to help the detective solve the cases where it gets stuck?"

They found that YES, we can! Even if the Summary Map looks messy (with two-way arrows or loops), it contains hidden clues that force the detailed map to have a specific direction.

Here are the three "Magic Rules" they discovered:

Rule #1: The One-Way Street

If the expert's Summary Map says "Alex influences Blake" (but Blake does not influence Alex), then in the detailed map, Alex must influence Blake.

• Analogy: If the city planner says "Water flows from the reservoir to the town," then the water pipes must flow that way. No mystery here.

Rule #2: The "No Self-Loop" Trick

Sometimes the Summary Map says "Alex and Blake influence each other" (a two-way street: Alex ⇄ Blake). Usually, this confuses the detective.

• BUT, if neither Alex nor Blake has a "Self-Loop" (meaning they don't influence themselves over time), the detective can actually figure out the direction!
• Analogy: Imagine two people pushing a swing. If neither person is also pushing themselves, and they are pushing each other, the timing of their pushes (who pushed first in the past) reveals who is the "leader" and who is the "follower," even if they are pushing back and forth.

Rule #3: The "Third Wheel"

If Alex and Blake are pushing each other, but Casey pushes Alex and doesn't push Blake, the detective can solve the puzzle.

• Analogy: If a third friend, Casey, only nudges Alex, we can use Casey's movement to figure out the direction of the push between Alex and Blake. It's like using a reference point to determine which way the wind is blowing.

4. The "Impossible" Case

The paper also tells us when we cannot solve the puzzle.
If Alex and Blake influence each other, AND they both influence themselves, AND they have the exact same set of friends pushing them, then the detective is truly stuck. The arrow will remain unoriented no matter what.

• Analogy: It's like two identical twins dancing in a mirror room with no outside reference. You literally cannot tell who is leading.

5. Why Does This Matter?

Knowing the direction of the arrow isn't just a game; it changes how we make decisions.

• Total Effect: If we want to know "If I change Alex's behavior, how much will Blake change?" we need to know the direction.
• Direct Effect: If we want to know "If I change Alex, how much does Blake change directly (ignoring Casey)?" we need the arrows to be clear.

The paper proves that by using the expert's "Summary Map" as a guide, we can guarantee that we will know the direction of the arrows in many more situations than we thought possible. This means we can make better predictions and better policies (like in medicine or economics) without needing perfect data.

Summary

Think of this paper as a guidebook for detectives. It says: "Don't just look at the crime scene (the data). Look at the suspect's sketch (the expert's Summary Map). Even if the sketch looks messy, it contains hidden clues that will force the suspect's identity (the arrow direction) to reveal itself."

This allows scientists to be more confident in their conclusions about cause and effect, even when the data is tricky.

Technical Summary

1. Problem Statement

In time series analysis, determining the causal structure (specifically the direction of edges) between variables is a fundamental challenge. While Full-Time Directed Acyclic Graphs (FT-DAGs) represent the complete causal structure including all temporal lags and instantaneous relations, they are often too complex to be fully specified by domain experts.

Instead, experts often provide Summary Causal Graphs (SCGs), which abstract time series into single nodes, capturing high-level causal relations (e.g., $X$ causes $Y$) but ignoring micro-level details like specific time lags or instantaneous effects. Conversely, causal discovery algorithms (like the PC algorithm) can recover partially oriented structures (FT-CPDAGs) from data but require strong assumptions and are computationally expensive.

The Core Question: Given an SCG (background knowledge) and a faithful, causally sufficient distribution from observational data, can we a priori guarantee that specific micro-level edges (especially instantaneous ones, $X_t \leftrightarrow Y_t$) will be oriented in the resulting Full-Time Maximally Oriented Partially Directed Acyclic Graph (FT-MPDAG)? If so, under what conditions?

2. Methodology and Framework

2.1. Definitions and Assumptions

The paper operates under three key assumptions:

1. Causal Sufficiency: No unobserved confounders; all noise terms are independent.
2. Faithfulness: All conditional independencies in the data are entailed by the causal Markov condition of the true graph.
3. Stationarity: The causal mechanism is time-invariant (the function generating $Y_t$ from its parents is the same for all $t$).

Key Concepts:

• FT-DAG: The ground truth graph with nodes representing variables at specific time steps ($X_t, Y_t$).
• SCG ($G_s$): A macro-graph where nodes represent entire time series ($S_X, S_Y$). An edge $S_X \to S_Y$ exists if any $X_{t'} \to Y_t$ exists in the FT-DAG.
• FT-MPDAG ($C$): The output of a causal discovery algorithm (e.g., tPC) augmented with background knowledge from the SCG. It contains the maximum number of oriented edges possible given the data and the SCG.
• s-orientability: A property defined by the authors. An edge $X_t \to Y_t$ is s-orientable if, for every FT-DAG compatible with the SCG and every faithful distribution, the resulting FT-MPDAG has the edge oriented (i.e., it is never left as an unoriented undirected edge $X_t - Y_t$).

2.2. Theoretical Approach

The authors analyze the conditions under which the combination of the SCG constraints and the orientation rules of causal discovery (specifically Meek's rules and the Unshielded Collider rule) forces the orientation of instantaneous edges. They distinguish between:

• Lagged edges: Always oriented by time ($X_{t-1} \to Y_t$).
• Instantaneous edges: The focus of the paper, as these are the only edges that may remain unoriented in the FT-MPDAG.

3. Key Contributions and Results

The paper provides a complete characterization of when instantaneous edges are guaranteed to be oriented. The results are summarized in Theorem 1, which relies on three lemmas:

3.1. Lemma 1: Directed Macro-Edges

If the SCG contains a directed edge $S_X \to S_Y$ (and not $S_Y \to S_X$), then any corresponding instantaneous edge $X_t \to Y_t$ in the FT-MPDAG is guaranteed to be oriented in that direction.

• Reasoning: The definition of the SCG implies the existence of a causal link; without a reverse link in the SCG, the reverse link cannot exist in the FT-DAG.

3.2. Lemma 2: Bidirected Edges without Self-Loops

If the SCG contains a bidirected edge $S_X \leftrightarrow S_Y$ (indicating mutual influence), the instantaneous edge $X_t - Y_t$ is still s-orientable provided that neither $S_X$ nor $S_Y$ has a self-loop in the SCG.

• Mechanism: The absence of self-loops ensures that specific unshielded triples (e.g., $Y_{t-1} \to X_t \leftarrow Y_t$) or chains exist, which force orientation via the Unshielded Collider (UC) rule or Meek's Rule 1.

3.3. Lemma 3: Asymmetric Parentage

If $S_X \leftrightarrow S_Y$ in the SCG, and both have self-loops, the edge is still s-orientable if there exists a node $S_Z$ that is a parent of $S_X$ but not a parent of $S_Y$ (i.e., $S_Z \in Pa(S_X) \setminus Pa(S_Y)$).

• Mechanism: The asymmetry in the parent sets creates unshielded triples or directed paths that force the orientation of the $X_t - Y_t$ edge to avoid cycles or satisfy collider constraints.

3.4. Theorem 1: The Necessary and Sufficient Condition

An instantaneous edge $X_t - Y_t$ is NOT s-orientable (i.e., it may remain unoriented in the FT-MPDAG) if and only if all three of the following conditions are met simultaneously in the SCG:

1. There is a bidirected edge: $S_X \leftrightarrow S_Y$.
2. Both nodes have self-loops: $S_X \to S_X$ and $S_Y \to S_Y$.
3. They share the exact same set of parents: $Pa(S_X) = Pa(S_Y)$.

Completeness: The theorem proves that if these three conditions hold, there exists at least one compatible FT-DAG and distribution where the edge remains unoriented. Conversely, if any condition is violated, the edge is guaranteed to be oriented.

3.5. Implications for Causal Effect Identification

The paper extends these orientability results to the identification of causal effects:

• Total Effect: If all edges connected to the treatment variable in the FT-MPDAG are s-orientable, the total effect is identifiable.
• Controlled Direct Effect: If all edges connected to the outcome variable are s-orientable, the controlled direct effect is identifiable.
• Significance: This allows researchers to verify identifiability conditions directly on the SCG (a high-level graph) without needing to run complex discovery algorithms first.

4. Significance and Practical Implications

1. Bridging Expert Knowledge and Data: The work demonstrates that high-level expert knowledge (SCGs) is not just a static constraint but actively enables the resolution of micro-level ambiguities that data alone might leave unresolved.
2. Efficiency in Causal Discovery: Researchers can determine before running computationally expensive algorithms whether specific edges are guaranteed to be oriented. This helps in focusing discovery efforts on the truly ambiguous parts of the graph.
3. Robustness to Cycles: Unlike standard DAG-based methods, this framework explicitly handles the cycles and bidirected edges common in time series summary graphs, providing theoretical guarantees even in these complex scenarios.
4. Quantitative Guarantees: The results provide a pathway to quantify causal effects (Total and Direct) in time series settings where full FT-DAGs are unknown, provided the SCG satisfies the derived conditions.

5. Limitations and Future Work

• Assumptions: The results rely heavily on Causal Sufficiency (no latent confounders) and Stationarity. If these assumptions are violated (e.g., non-stationary data or unobserved confounders), the guarantees may not hold.
• SCG Accuracy: The method assumes the SCG is a faithful representation of the underlying truth. If the SCG contains incorrect edges (e.g., a false bidirected edge), the orientation guarantees could lead to incorrect conclusions.
• Future Directions: The authors suggest extending these results to the FCI algorithm (which handles latent confounders) and investigating the completeness of the identifiability conditions for causal effects.

In summary, this paper provides a rigorous theoretical framework for leveraging Summary Causal Graphs to guarantee the orientation of causal relations in time series, significantly enhancing the reliability and efficiency of causal discovery and effect estimation in complex temporal systems.