📊 statistics

Causal Data Fusion for Panel Data without a Pre-Intervention Period

This paper introduces two data-fusion methods that enable causal inference in panel data settings where pre-intervention data are unavailable by leveraging auxiliary reference domains to estimate counterfactuals.

Original authors: Zou Yang, Seung Hee Lee, Julia R. Köhler, AmirEmad Ghassami

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zou Yang, Seung Hee Lee, Julia R. Köhler, AmirEmad Ghassami

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 Problem: The "Sudden Storm" Dilemma

Imagine you are a scientist trying to study whether a new type of umbrella actually works. To prove it, you’d normally want to watch a person walking in the sun before the rain starts, and then watch them again during the rain to see if they stay dry. This "before and after" comparison is the gold standard in science.

But what if a sudden, massive storm hits out of nowhere? You didn't have time to set up your cameras or watch the person walking in the sun. The rain is already pouring, and you have no "before" data.

In the real world, this happens all the time. During the COVID-19 pandemic, certain community groups in cities like Chelsea, Massachusetts, launched massive vaccination campaigns almost immediately after vaccines became available. Because the intervention happened so fast, researchers had no "pre-intervention" data to show what the vaccination rates in that specific community would have looked like if the group hadn't stepped in. Without that "before" picture, it’s hard to prove if the community group actually made a difference or if people were just getting vaccinated on their own.

The Solution: The "Mirror World" Strategy (Data Fusion)

The authors of this paper propose a clever workaround called Data Fusion.

Since they can't look at the "before" in the target city (Chelsea), they look for a "Mirror World" (a Reference Domain). They look at a different group of people—in this case, the Black sub-population in those same cities—who were also being vaccinated during the same time.

Think of it like this: Imagine you want to know if a new fertilizer helped a specific patch of red roses grow faster, but you didn't record how fast they were growing before you added the fertilizer. You can't go back in time, but you can look at a nearby patch of blue hydrangeas. If you know that blue hydrangeas and red roses usually grow at a similar relative rate, you can use the "growth story" of the hydrangeas to predict what the roses would have done without the fertilizer.

Two Ways to Solve the Puzzle

The paper offers two main mathematical "tools" to do this:

The "Equal Balance" Method (Equi-Confounding):
This is like saying, "If the gap between the target group and the average group is 10 points in the 'Mirror World,' it was probably also 10 points in the 'Target World' before the intervention." It assumes the "vibe" or the underlying trends are balanced between the two groups. It’s simple and direct, like using a ruler to match two lines.
The "Custom Matchmaker" Method (Synthetic Control Data Fusion):
This is much more sophisticated. Instead of just comparing Chelsea to an "average" city, the researchers act like a high-end matchmaker. They look at all the other cities and say, "City A is a bit too rich, City B is too old, but if I combine 30% of City A and 70% of City B, I can create a 'Synthetic Chelsea'—a digital twin that looks and acts almost exactly like the real Chelsea did."

By creating this "Digital Twin" using data from the "Mirror World," they can compare the real Chelsea to the digital Chelsea. The difference between the two tells them exactly how much the community organization helped.

Why It Matters: The Real-World Result

The researchers applied this to the Chelsea vaccination case. Even without "before" data, their math showed a clear, powerful result: the community organization's efforts led to a massive boost in vaccination rates (estimated between 10% and 13%).

The Big Picture: This paper gives scientists a "time machine" of sorts. It allows them to study the impact of emergency responses—like disaster relief, sudden medical interventions, or rapid policy changes—even when they arrive too late to record the "before" state. It turns "we missed the chance to study this" into "we can use what we know elsewhere to find the truth."

Technical Summary: Causal Data Fusion for Panel Data without a Pre-Intervention Period

1. Problem Statement

Traditional causal inference frameworks for panel data—most notably Difference-in-Differences (DiD) and Synthetic Control Methods (SCM)—rely heavily on "pre-intervention" data. These methods use the target unit's behavior prior to a treatment to establish a baseline or to match it with control units.

However, in real-world scenarios involving sudden shocks or emergencies (e.g., public health crises, epidemiological outbreaks, or sudden policy shifts), interventions are often implemented immediately. In such cases, no data exists for the target unit between the onset of the event and the implementation of the intervention. This creates a "missing pre-intervention period" problem, rendering standard DiD and SCM inapplicable.

2. Methodology

The authors propose two novel data-fusion methods that leverage an auxiliary reference domain to estimate the counterfactual (what would have happened to the target unit without treatment). The reference domain contains panel data from related units (e.g., a different demographic sub-population in the same geographic area) that does have observable data during the period of interest.

A. Equi-Confounding Data Fusion
This approach extends the logic of "negative controls" and DiD. It relies on the assumption that the confounding effect (unobserved variables) in the target domain mirrors the confounding effect in the reference domain.

Linear Equi-Confounding: Assumes the difference in mean potential outcomes between the target unit and the control group in the target domain is equal to the same difference in the reference domain.
Logarithmic Equi-Confounding: A variation for data measured on a multiplicative scale, where the relationship is defined by the difference in the logarithms of the means. This is shown to be more robust to certain data-generating processes.

B. Synthetic Control Data Fusion (SCDF)
This is a more flexible extension of the SCM. Instead of weighting all control units equally, SCDF seeks an optimal weight vector $w$ to create a "transferred synthetic control."

The Challenge: A naive SCM might match the target unit perfectly in the reference domain (e.g., matching Black vaccination rates) but fail to match the observed covariates in the target domain (e.g., Hispanic median income).
The Solution: The authors introduce a data-driven constrained optimization algorithm. It minimizes the Normalized Squared Error (NSE) of the reference outcomes while enforcing constraints that ensure the weights also provide a "good enough" match for the observed covariates in the target domain. This prevents the model from overfitting to the reference domain at the expense of target domain relevance.

3. Key Contributions

Theoretical Framework: The paper formalizes the conditions (Equi-confounding and Factor Models) under which causal effects can be identified via data fusion.
Bias Bounds: The authors derive mathematical proofs showing that the absolute bias of their estimators converges to zero as the number of reference-domain time periods ( $T$ ) or control units ( $J$ ) increases, provided certain conditions are met.
Novel Optimization Algorithm: The introduction of a constrained optimization method for SCDF that balances matching across two different domains (target and reference).
Practical Application: A real-world application to COVID-19 vaccination rates where no pre-intervention data was available.

4. Results

Simulation Studies: Simulations demonstrated that the SCDF method generally outperforms the equi-confounding methods. The bias of the SCDF estimator decreases as the number of reference time periods increases, confirming the theoretical bounds.
Real-World Application (Chelsea, MA): The researchers evaluated the impact of a community organization's vaccination campaign for the Hispanic population in Chelsea, MA.
- Target Domain: Hispanic vaccination rates.
- Reference Domain: Black vaccination rates in the same cities.
- Findings: All three proposed methods (Linear, Logarithmic, and SCDF) yielded a large, positive causal effect. The estimated increase in the Hispanic vaccination rate was 13.6% (Linear), 13.2% (Logarithmic), and 10.1% (SCDF). The results were statistically significant according to placebo tests.

5. Significance

This paper provides a vital toolkit for researchers working in urgent, data-constrained environments. By allowing the "fusion" of disparate datasets (e.g., different demographics or different time periods), it enables robust causal inference in the wake of sudden humanitarian or medical crises where traditional longitudinal data is non-existent. It bridges the gap between the need for immediate policy evaluation and the mathematical rigor of panel-data econometrics.