📈 economics

Dynamic causal inference with time series data

This paper generalizes the potential outcome framework to time series data by defining causal effects on entire stochastic trajectories, introducing a Dynamic Average Treatment Effect (DATE) estimator that captures how interventions alter evolutionary dynamics and persistence, thereby overcoming the systematic misestimation inherent in static causal methods.

Original authors: Tanique Schaffe-Odeleye, Kōsaku Takanashi, Vishesh Karwa, Edoardo M. Airoldi, Kenichiro McAlinn

Published 2026-02-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tanique Schaffe-Odeleye, Kōsaku Takanashi, Vishesh Karwa, Edoardo M. Airoldi, Kenichiro McAlinn

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

Imagine you are trying to figure out if a new fertilizer makes a specific type of plant grow better.

The Old Way (Static Thinking):
In traditional studies, researchers might look at a garden, pick 100 plants, give 50 fertilizer and 50 water, and measure their height once at the end of the summer. They calculate the average difference. This works well if the plants just sit there and grow at a steady rate.

The Problem with Time:
But what if the fertilizer doesn't just make the plant taller? What if it changes how the plant grows? Maybe it makes the plant shoot up quickly for a month, then stop growing, then start wilting, while the watered plants grow slowly but steadily the whole time.

If you only look at the final height, you miss the whole story. You might think the fertilizer is great because the plant is tall, or terrible because it's dead, but you wouldn't know when or how the fertilizer changed the plant's behavior.

The New Solution (This Paper):
This paper introduces a new way to think about cause and effect for things that change over time, like stock markets, disease spread, or unemployment rates. They call this "Dynamic Causal Inference."

Here is the core idea broken down into simple analogies:

1. The "Movie" vs. The "Snapshot"

Old Way: Treats data like a photograph. It asks, "What is the result at this one moment?"
New Way: Treats data like a movie. It asks, "How does the entire story of the outcome change?"
The Analogy: Imagine two runners.
- Runner A (Control) runs at a steady 5 mph.
- Runner B (Treated) takes a drug.
- Static View: You check their speed at the finish line.
- Dynamic View (DATE): You watch the whole race. You see that Runner B started fast, got tired, slowed down, and then sprinted at the end. The "Dynamic Average Treatment Effect" (DATE) is a map that shows exactly how the drug changed their speed at every single second of the race, not just the final result.

2. The Two Scenarios: The Crowd vs. The Lone Wolf

The paper explains that you need different tools depending on how much data you have.

Scenario A: The Crowd (Many Units)
Imagine you have 1,000 people. Some take a drug, some don't.

The Tool: They use a method called Dynamic Inverse Probability Weighting (DIPW).
The Analogy: Think of this like a fairness scale. In the real world, people who take a drug might be different from those who don't (e.g., sicker people take the drug). The "weighting" adjusts the data so that the "sick" people in the drug group are balanced against the "sick" people in the non-drug group. It re-arranges the crowd so it looks like a fair experiment, allowing them to see the true effect of the drug over time.

Scenario B: The Lone Wolf (One Unit)
Imagine you only have one country that went into lockdown (like the UK in the paper) and no other country to compare it to directly.

The Tool: They use a Dynamic Linear Model (DLM), which is a type of "State-Space" model.
The Analogy: Think of this as a Time Machine. Since you only have one country, you can't compare it to a twin. Instead, you build a mathematical "Time Machine" based on that country's history before the lockdown.
- The machine learns the country's natural rhythm (trends, seasons, usual ups and downs).
- When the lockdown hits, the machine projects what would have happened if the lockdown never occurred (the "counterfactual").
- The difference between the real path and the Time Machine's path is the true effect of the lockdown.

3. The "Three Parts" of the Effect

The paper shows that when you look at the whole movie, you can break the effect down into three distinct parts, which static methods often mix up:

The "Spot" Effect: The immediate shock. (e.g., The moment the lockdown started, unemployment jumped up instantly).
The "Persistent" Effect: A permanent shift in the level. (e.g., Even after things settled, unemployment stayed higher than before).
The "Trend" Effect: A change in the direction or speed. (e.g., The rate at which unemployment was growing or shrinking changed fundamentally).

The Real-World Example in the Paper:
The authors tested this on UK unemployment during the 2020 COVID-19 lockdown.

Old methods (like simple averages) saw a mess and couldn't tell if the effect was a one-time jump or a long-term change.
Their new method showed a clear story:
- There was almost no instant "spot" jump.
- There was a big "persistent" rise (unemployment went up and stayed up).
- There was a "trend" change (the rate of recovery slowed down).
- By separating these, they could see that the lockdown caused a temporary spike that eventually settled into a new, higher normal, rather than just a temporary blip.

Summary

This paper argues that to understand cause and effect in a changing world, you can't just look at the final score. You have to watch the whole game. They provide a mathematical rulebook (the DATE) and two sets of tools (one for crowds, one for lone wolves) to measure exactly how an intervention changes the story of a system over time, rather than just the ending.

Technical Summary: Dynamic Causal Inference with Time Series Data

Problem Statement
Traditional causal inference frameworks, such as the Rubin potential outcomes model, treat outcomes as random variables realized at a single point in time. This static approach is ill-suited for dynamic systems where interventions do not merely shift outcome levels but fundamentally alter the evolutionary dynamics of the process itself (e.g., changing persistence, volatility, and transition laws). Existing methods like Difference-in-Differences (DiD), Synthetic Control Methods (SCM), and Interrupted Time Series (ITS) often assume time-invariant dynamics or parallel trends. Consequently, when interventions modify the underlying stochastic evolution, these static methods conflate causal effects with temporal dependence, leading to biased estimates. The paper addresses the need for a causal framework that treats time as intrinsic to the causal mechanism, defining estimands directly on the space of stochastic trajectories rather than single-period observations.

Methodology
The authors generalize the potential outcome framework to time series by defining potential outcomes as entire stochastic trajectories $\{Y_t(Z)\}_{t \in [0,T]}$ .

The Dynamic Average Treatment Effect (DATE):
The core estimand is the DATE, defined as the expected difference between the treated and control stochastic processes at each time point $t$ , averaged over the distribution of initial states. Formally:
$\text{DATE}(t) \triangleq \int \left( E_{y_0}[Y_t(Z=1)] - E_{y_0}[Y_t(Z=0)] \right) p(y_0) dy_0$
This reduces to the classical Average Treatment Effect (ATE) when $T=1$ or under i.i.d. assumptions.
Identification and Estimation Regimes:
The paper distinguishes two empirical regimes based on data availability:
- Multi-unit/Design-based Regime: When multiple comparable units are available, the authors propose the Dynamic Inverse-Probability Weighting (DIPW) estimator. This extends standard IPW to dynamic settings by conditioning on the filtration $\mathcal{F}^X_{t_c}$ (the history of confounders up to the intervention time). Under Dynamic Ignorability (Assumption 4.1), which posits that potential outcomes are independent of treatment assignment conditional on pre-intervention history, the DIPW estimator is shown to be unbiased and consistent.
- Scarce-treated/Model-based Regime: When only one or few treated units are observed, cross-sectional reweighting is insufficient. The authors demonstrate that the conditional mean trajectories underlying the DATE admit a state-space representation. Specifically, they prove (Theorem 5.1) that under mild regularity conditions (square-integrable semimartingales), the conditional mean dynamics can be expressed as a Dynamic Linear Model (DLM). This allows for the estimation of the DATE by fitting a state-space model to the full series, decomposing the intervention effect into "spot" (immediate), "persistent" (level shift), and "trend" (slope change) components.
Theoretical Foundations:
The framework introduces Intertemporal Strong Ignorability (Assumption 5.1), a time-series extension of strong ignorability that requires treatment assignment to be independent of future confounders (no feedback). This distinguishes the approach from sequential potential outcome frameworks that treat temporal dependence as network interference; here, dependence is treated as intrinsic stochastic evolution.

Key Results

Simulation Study: The authors evaluate the framework against existing methods (LM, AR(1)-LM, ARIMAX, SCM, DiD, C-ARIMA) across four scenarios: Many-Many, One-Many, One-One, and One-None (single treated unit, no controls).
- The DLM consistently achieves the lowest Mean Squared Error (MSE) and accurate 95% coverage probabilities across all scenarios, particularly in the "One-None" case where other methods fail to capture time-varying effects.
- Static methods (LM, ARIMAX) and methods relying on observed series directly (SCM, DiD) exhibit high bias and poor coverage, often estimating flat averages or failing to adapt to evolving dynamics.
Empirical Application: The framework is applied to the effect of the UK's 2020 COVID-19 lockdown on unemployment (a One-None scenario).
- The DLM successfully captures a non-linear, arch-shaped trajectory of the treatment effect, revealing an initial sharp increase in unemployment followed by a slow recovery over nearly a year.
- Decomposition analysis reveals that the effect is driven by a persistent level shift (temporary increase) and a negative trend component (deterioration in growth rate), while the "spot" effect was negligible. This level of granularity was obscured in competing models.

Significance and Claims
The paper claims to provide a unified probabilistic foundation for causal inference in dynamic systems. Its primary contributions are:

Theoretical Extension: Extending the Rubin causal model from random variables to stochastic processes, thereby formally defining causal effects as properties of entire trajectories.
Identification Strategy: Deriving conditions for dynamic identification (Dynamic Ignorability) and providing a valid estimator (DIPW) for settings with time-dependent confounding.
Practical Implementation: Showing that when data is limited, the theoretical DATE admits a state-space representation, justifying the use of Dynamic Linear Models for causal estimation in single-unit or scarce-treated scenarios.

The authors emphasize that their framework clarifies when model-based approaches (like Structural Time Series or CausalImpact) target well-defined causal estimands. It distinguishes between forward-looking counterfactual prediction (based only on pre-treatment data) and retrospective causal attribution (using the full series to decompose baseline dynamics from intervention effects), advocating for the latter to sharpen the decomposition of treatment effects. The paper concludes that treating time as a confounder and part of the causal mechanism is essential to avoid the systematic misestimation of dynamic effects inherent in static methods.

1. The "Movie" vs. The "Snapshot"

2. The Two Scenarios: The Crowd vs. The Lone Wolf

3. The "Three Parts" of the Effect

Summary

Technical Summary: Dynamic Causal Inference with Time Series Data

More like this