📊 statistics

Modifying causal models to distinguish between transient and lasting causal effects

This paper proposes a system and state-based approach to causal modeling that overcomes the limitations of traditional potential outcomes and DAGs in time-varying contexts by introducing a novel null effect definition to distinguish between transient and lasting causal impacts on equilibrium behavior.

Original authors: Russell Steele, Naftali Weinberger, Tess Baker, Ian Shrier

Published 2026-05-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Russell Steele, Naftali Weinberger, Tess Baker, Ian Shrier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Fixing the "Snapshot" Problem

Imagine you are trying to understand how a car engine works. Most traditional studies take a "snapshot" of the car at specific moments: "At 10:00 AM, the car was moving at 50 mph. At 10:05 AM, it was moving at 55 mph."

The authors of this paper argue that this "snapshot" approach is often misleading when studying systems that change over time, like human diseases, economies, or ecosystems. They say that just because a car speeds up for a moment after you press the gas, it doesn't mean you've changed how the engine works. You might have just pushed the car while it was already rolling.

The paper proposes a new way to think about cause and effect that distinguishes between temporary bumps and real changes to the system's rules.

The Problem: The "Bump" vs. The "Rule Change"

To explain this, the authors use an analogy of cancer cells growing in the body.

1. The Old Way (The Snapshot):
Imagine a doctor gives a drug that shrinks a tumor.

The Measurement: The doctor measures the tumor size at Day 10, Day 20, and Day 30.
The Result: The tumor shrinks at Day 10, but by Day 30, it has grown back to its original size.
The Flaw: In traditional models, if you only look at Day 30, you might say, "The drug had no effect." If you look at Day 10, you say, "The drug worked!" But both answers are incomplete. The drug didn't actually change the rules of how the tumor grows; it just temporarily pushed the tumor back. The tumor's "engine" (its biology) is still running the same way, so it eventually returns to its normal size.

2. The New Way (The System Dynamics):
The authors suggest we shouldn't just look at the size of the tumor (the state); we should look at the rules governing its growth (the dynamics).

Transient Effect: Like pushing a swing. You give it a big push (the drug), it goes higher for a moment, but gravity (the system rules) eventually pulls it back to its normal rhythm. The swing's engine hasn't changed.
Lasting Effect: Like changing the length of the swing's chain. Now, the swing moves at a completely different speed and height forever. The rules of the swing have changed.

The paper argues that current scientific methods are great at measuring the "push" (transient effects) but terrible at identifying if we actually changed the "chain length" (lasting effects on the system).

The Solution: Changing the "Knobs," Not the "Dials"

The authors propose a new framework to tell these two scenarios apart. They use a metaphor of a machine with Dials and Knobs.

The Dials (System States): These are the things you can measure right now, like the tumor size, the temperature, or the stock price.
- Intervention: If you manually turn a dial to a new number, the machine might react for a while, but it will eventually settle back into its old pattern. This is a Transient Effect.
- Paper's Claim: Traditional studies often mistake turning a dial for changing the machine.
The Knobs (System Parameters): These are the hidden settings that determine how the machine behaves. In the cancer example, these are the biological rates of cell growth or how cells fight each other.
- Intervention: If you turn a knob (e.g., you change the biological rule so cells grow slower), the machine behaves differently forever. It finds a new "equilibrium." This is a Lasting Effect.
- Paper's Claim: To truly change a system, you must intervene on the "knobs" (the parameters), not just the "dials" (the current state).

The "Null Effect" Redefinition

The paper introduces a clever new definition of "No Effect" (Null Effect).

Old Definition: "If the result is the same as before, there was no effect."
New Definition: "If the result is the same as before, but only because we shifted the timeline, there was no real effect on the system's rules."

The Analogy:
Imagine a runner on a track.

Scenario A: You push the runner. They speed up for 10 seconds, then slow down and run at their normal pace.
Scenario B: You give the runner better shoes. They run faster the whole time.

In the old view, if you check the runner at the 1-hour mark, both scenarios might look the same (they are running at their normal pace). The old view says, "The push didn't work."
The new view says, "The push didn't change the runner's potential or rules. It just shifted their position in time. The runner is still running the same race, just starting a few seconds later."

Why This Matters for Research

The authors conclude that to get better answers from studies (especially in health and medicine), researchers need to:

Stop relying only on snapshots: Measuring a system at just one or two points in time can hide whether a treatment actually changed the underlying biology or just gave a temporary boost.
Measure more often: To see if a system has truly changed its "rules," you need to watch it long enough to see if it settles into a new pattern or returns to the old one.
Target the right things: If you want a permanent cure, you need to find a way to change the "knobs" (the biological mechanisms/parameters), not just push the "dials" (the current symptoms).

Summary

This paper is a guide for scientists to stop confusing temporary fixes with permanent changes. It argues that if you only look at the "what" (the current state), you miss the "how" (the system's rules). To truly understand cause and effect in complex, changing systems, we need to design our studies to see if we have changed the engine, or just pushed the car.

Technical Summary: Modifying Causal Models to Distinguish Between Transient and Lasting Causal Effects

Problem Statement
Current causal inference frameworks, primarily based on potential outcomes and causal Directed Acyclic Graphs (DAGs), face significant limitations when applied to systems with complex temporal dynamics. While these methods successfully analyze cross-sectional and longitudinal data, they struggle to distinguish between interventions that cause transient changes (where the system eventually returns to its baseline equilibrium) and those that induce lasting changes to the system's dynamics.

In standard causal models, variables represent measurements at specific time points. Consequently, causal effects are defined relative to the timing of these measurements. If an intervention temporarily alters a state variable (e.g., tumor volume) but the system's underlying differential equations remain unchanged, the system will eventually return to its original equilibrium trajectory, merely shifted in time. Standard definitions of the null effect ( $E[Y(1) - Y(0)] \neq 0$ ) fail to differentiate between a temporary perturbation of a state and a fundamental modification of the system's causal parameters. This limitation is critical in fields like health research, where distinguishing between a treatment that temporarily reduces tumor size versus one that alters the biological mechanisms of tumor growth is essential.

Methodology
The authors propose a shift from a "measurement-based" approach to a "system-and state-based" approach. The methodology involves:

Dynamic Causal Modeling: Utilizing systems of ordinary differential equations (ODEs) to represent the dynamics of the system, rather than discrete time-point variables alone. The paper uses a Lotka-Volterra competition model for tumor growth as a motivating example, where cell volumes ( $V_{target}$ and $V_{proxy}$ ) evolve according to growth rates ( $r$ ), carrying capacities ( $K$ ), and interaction parameters ( $\gamma$ ).
Differentiating Interventions: The framework distinguishes between two types of interventions:
- Interventions on States: Interventions that alter the value of an endogenous variable (a node on the causal DAG) at a specific time. These shift the system's trajectory but do not change the underlying differential equations. The system may reach the same equilibrium at a different time or a different equilibrium reachable under the original dynamics.
- Interventions on Dynamics: Interventions that alter the exogenous parameters of the system (e.g., $K$ , $\gamma$ , or $r$ ). These changes modify the causal relationships and the equations governing the system, leading to trajectories that were impossible under the original dynamics.
Redefining the Null Effect: The paper proposes a novel definition of the null causal effect for dynamic systems. A causal effect is considered "null on dynamics" if the post-intervention trajectory can be mapped to a pre-intervention trajectory by shifting the initial time or initial state, without altering the system's governing equations. Conversely, a "non-null dynamic effect" exists only if the intervention changes the system's parameters, creating trajectories that could not have been generated by the original dynamics regardless of initial conditions.
Causal Graph Representation: The authors utilize dynamic causal graphs (incorporating derivatives and parameters) to visualize these distinctions. Interventions on state nodes (red nodes) affect the trajectory but not the graph structure, while interventions on parameter nodes (blue nodes) alter the dynamics of the system.

Key Contributions and Results

Limitation of Standard DAGs: The paper demonstrates that standard causal DAGs, when applied to time-varying systems, can violate the Causal Faithfulness condition. If an intervention causes a transient effect that eventually subsides (returning the system to equilibrium), the variable becomes probabilistically independent of the intervention after a certain time, leading standard discovery algorithms to incorrectly infer no causal relationship.
Formal Distinction of Effects: The authors provide a formal mechanism to distinguish between:
1. Transient Effects: Interventions on state values that result in the system reaching the same equilibrium later or earlier, or a different equilibrium achievable under the original dynamics.
2. Lasting (Dynamic) Effects: Interventions on system parameters that fundamentally alter the system's behavior and equilibrium possibilities.
Identification of Causal Parameters: The paper argues that identifying true causal effects on system dynamics requires targeting exogenous variables (parameters) rather than endogenous states. It synthesizes ideas from mainstream causal DAGs (Pearl, 2009), dynamic causal models (Blom and Mooij, 2023; Weinberger, 2023), and causal ordering at equilibrium (Iwasaki and Simon, 1994).
Implications for Study Design: The selection of timepoints for measurement is shown to be critical. If measurements are taken only after a system has returned to equilibrium following a state-based intervention, the causal effect may be misidentified as null. The proposed framework clarifies that to detect dynamic effects, one must measure the system's trajectory or target the system's parameters.

Significance and Claims
The paper claims to offer a new structure for improving the specificity of interventions, study design, and data collection in dynamic systems. By redefining the null effect to account for system dynamics, the authors aim to allow researchers to formally differentiate between interventions that merely shift a system's state and those that change the system's underlying causal mechanisms.

The authors explicitly state that the paper focuses on the causal identification problem rather than the estimation of these effects from observed data. They acknowledge that while their approach clarifies what constitutes a causal effect on dynamics, estimating these effects (e.g., via functional registration or G-computation) presents significant challenges, particularly with irregular or infrequent data. The work is presented as a theoretical framework to guide future research in distinguishing transient from lasting causal effects in observational and experimental studies involving time-varying systems.