Forecasting Causal Effects of Future Interventions: Confounding and Transportability Issues

Imagine you are a chef who just perfected a new recipe for a spicy soup. You served it to a group of friends last winter, and they loved it. Now, it's summer, and you want to serve that exact same soup to a new group of friends.

The big question is: Will the soup taste just as good in the summer as it did in the winter?

This is the core problem tackled in the paper by Forastiere, Li, and Baccini. They are trying to figure out how to predict the success of a policy or event in the future based on data from the past, without just guessing.

Here is a breakdown of their ideas using simple analogies:

1. The Problem: "Time Travel" is Harder than "Space Travel"

Usually, scientists can take a result from one place (like New York) and apply it to another place (like London). This is like moving your soup recipe from your kitchen to a friend's kitchen. It's tricky, but doable if you account for the friend's different taste buds.

However, this paper is about time travel. Can you take a result from last year and apply it to next year?

The Catch: In the real world, things change over time. The weather changes, people's habits change, and the virus itself might mutate.
The Mistake: Many policymakers make a lazy guess: "If the lockdown worked in Spring 2020, it will work exactly the same way in Fall 2020."
The Reality: The "context" has changed. In the spring, people were scared and stayed home. In the fall, they might be tired of staying home, or the virus might be stronger. If you ignore these changes, your prediction will be wrong.

2. The Solution: The "Time-Traveling Recipe" Framework

The authors built a mathematical framework (a set of rules) to help us make these predictions more accurately. Think of it as a Time-Traveling Recipe Book.

To use this book, you need to check three things:

A. The "Secret Ingredients" (Effect Modifiers)

In our soup analogy, the "secret ingredients" are things like the temperature outside, how hungry your friends are, or how spicy they like their food. In the paper, these are called effect modifiers.

The Rule: You must know exactly what these ingredients were last winter, and you must be able to predict what they will be next summer.
The Challenge: If a new ingredient appears in the future that you didn't have in the past (like a new virus mutation), your recipe will fail. The authors say you must identify all the ingredients that could change the outcome.

B. The "Cooking Process" (Temporal Transportability)

This is the fancy term for "does the cooking method stay the same?"

The Assumption: The authors assume that the relationship between the ingredients and the final taste stays the same. Even if the temperature is hotter in summer, the rule "more heat = faster cooking" shouldn't change.
The Danger: If the rules of physics or human behavior change (e.g., people suddenly stop wearing masks even though it's hot), the "cooking process" breaks, and the prediction is invalid.

C. The "Simulated Future" (G-Computation)

Since we can't actually travel to the future to see what happens, the authors suggest a Simulation Game.

Step 1: Use your past data to build a model of how the world evolves. (e.g., "Usually, when it gets hotter, people go outside more.")
Step 2: Run a simulation of the future. "Okay, let's pretend it's next summer. Based on our model, the weather will be X, and people will behave Y."
Step 3: Apply your "soup recipe" (the policy) to this simulated future.
Step 4: The result is your forecast.

3. The "Time-Varying" Twist

The paper gets even more interesting when the "intervention" isn't just a one-time event (like a single protest) but something that lasts (like a lockdown that goes on for weeks).

The Analogy: Imagine you are trying to predict the effect of a 3-day rainstorm.
- Day 1: It starts raining.
- Day 2: The ground gets muddy.
- Day 3: People start slipping.
The Problem: The mud on Day 2 is caused by the rain on Day 1. This is called a time-dependent confounder. The past is affecting the future, which affects the outcome.
The Fix: The authors' framework allows you to trace this chain reaction. You don't just look at the start of the rain; you simulate how the mud builds up day by day, so you can predict exactly how many people will slip on Day 3.

4. Why This Matters (The COVID Example)

The paper uses the COVID-19 pandemic as its main example.

The Past: In Spring 2020, governments locked down cities. It worked (mostly) because people were scared and the virus was new.
The Future: In Fall 2020, governments had to decide: "Should we lock down again?"
The Prediction: If you just looked at the Spring data, you might say, "Yes, lock down!" But the authors' framework forces you to ask: "Did the virus change? Are people tired? Are hospitals better prepared?"
The Result: By simulating these changes, you might realize that a full lockdown in the Fall won't work the same way as in the Spring, or that it might need to be shorter or longer.

Summary: The "Crystal Ball" with Rules

This paper doesn't give us a magic crystal ball. Instead, it gives us a rulebook for building a crystal ball.

It tells us:

Don't just guess. You can't assume the future is identical to the past.
List your variables. You must know every factor that could change the result (weather, behavior, virus mutations).
Simulate the path. You have to mathematically "walk" from the past to the future, step-by-step, accounting for how the world changes along the way.
Check your assumptions. If you miss a key ingredient (like a new virus variant), your prediction is garbage.

In short: The paper provides the mathematical tools to stop policymakers from saying, "It worked last time, so it will work next time," and instead helps them say, "It worked last time under these specific conditions. If we change the conditions, here is exactly how the result will change."

Here is a detailed technical summary of the paper "Forecasting Causal Effects of Future Interventions: Confounding and Transportability Issues" by Forastiere, Li, and Baccini.

1. Problem Statement

The paper addresses a critical gap in causal inference: forecasting the causal effects of interventions or events in a future time window based on data observed in a past time window.

Context: While existing literature extensively covers transportability (generalizing causal effects from a study sample to a different target population in the same time frame) and generalizability (extending inferences from a trial sample to a target population), these methods typically assume time-fixed treatments and baseline effect modifiers.
The Challenge: Forecasting into the future introduces unique complexities:
1. Time-Varying Confounders and Effect Modifiers: The characteristics of the population (e.g., epidemic status, behavioral compliance, environmental factors) evolve over time.
2. Unobserved Future States: The covariates and effect modifiers in the future time window are not yet observed.
3. Dynamic Shifts: The relationship between treatment and outcome may change due to the evolution of the context (e.g., a lockdown in Spring 2020 vs. Fall 2020 during a pandemic).
Motivating Example: The authors use the impact of public policies (e.g., lockdowns, school closures) and social events (e.g., protests, elections) on COVID-19 mortality. The core question is: Will the effect of a policy implemented in Spring 2020 hold true if replicated in Fall 2020, given that population behaviors, viral variants, and healthcare capacity have changed?

2. Methodology and Framework

The authors develop a potential outcomes framework specifically designed for temporal transportability. They distinguish between evaluation (assessing effects in the observed window) and prediction (forecasting effects in a future window).

A. Notation and Setup

Units ( $i$ ): Geographical areas (e.g., states, cities).
Time ( $t$ ): Discrete time points within an observed window $T$ and a future window $T+F$ .
Treatment ( $Z_{it}$ ): Binary indicator of an intervention or event.
Exposure ( $S_{it}$ ): Continuous variable affected by the treatment (e.g., number of contagious interactions).
Outcome ( $Y_{it}$ ): Aggregate health outcome (e.g., daily deaths).
Covariates ( $X_{it}$ ): Time-varying confounders and effect modifiers.

B. Identification of Observed Effects

Before forecasting, the authors establish how to identify effects in the observed sample:

Point Treatments: Use standard Treatment Ignorability (Assumption 1) conditional on baseline and current covariates.
Time-Varying Treatments: Use Sequential Ignorability (Assumption 5). This accounts for:
- Latency ( $B$ ): Delay between treatment and outcome.
- Carry-over ( $K$ ): Duration of treatment effect.
- Time-Dependent Confounders: Variables affected by past treatments that influence future treatments and outcomes. The authors utilize g-formulas (g-computation) to adjust for these, distinguishing between scenarios with and without time-dependent confounding.

C. Identification of Future Effects (Temporal Transportability)

The core contribution is the derivation of nonparametric identification formulas for future causal effects ( $ATTF$ ). This requires two layers of assumptions:

Temporal Transportability of Potential Outcomes (Assumptions 2 & 6):
- Assumes that the functional relationship between covariates/history and the potential outcome is invariant across time.
- Formally: $P(Y_{it}(d) | \text{History}) = P(Y_{i, T+F}(d) | \text{History})$ .
- This implies that all effect modifiers are measured and included in the history vector.
Temporal Transportability of Time-Varying Modifiers (Assumptions 3 & 7):
- Since future covariates ( $X_{T+F}$ ) are unobserved, their distribution must be predicted.
- Assumes the transition mechanism of covariates (and outcomes) is stable over time.
- The evolution of modifiers depends only on their own history and past treatments, not on unobserved time-specific shocks.

D. Identification Formulas (G-Computation)

The authors derive g-computation formulas to identify future effects:

Step 1 (Projection): Recursively predict the distribution of future effect modifiers ( $X_{T+F}$ ) using historical transition models fitted on the observed data.
Step 2 (Imputation): Apply the stable outcome mechanism (estimated from the past) to these projected future covariates.
Formula Structure:
$E[Y_{T+F}(d)] = \sum_{x} E[Y_{obs} | \text{History}=x] \times f(X_{T+F} | \text{History})$
Where $f(X_{T+F})$ is obtained via recursive integration (Monte Carlo simulation) over the transition kernels estimated from the past.

3. Key Contributions

Formal Definition of Causal Estimands for Forecasting: The paper rigorously defines the Average Treatment Effect on the Treated ( $ATTF$ ) for future time windows, distinguishing it from standard transportability to different populations.
Novel Nonparametric Identification Formulas: The authors provide the first set of g-computation formulas that allow for the nonparametric identification of future causal effects under time-varying treatments and confounders.
Decomposition of Assumptions: The framework explicitly separates the assumptions required for:
- The stability of the outcome mechanism (transportability of potential outcomes).
- The stability of the covariate evolution (transportability of time-varying modifiers).
Handling Time-Dependent Confounding: The methodology extends to complex settings where past treatments influence future confounders (e.g., a lockdown affecting hospital capacity, which then influences the decision to extend the lockdown).
Practical Implementation Guide: The paper outlines a Monte Carlo imputation procedure:
- Fit a generative model for covariate evolution.
- Simulate future covariate trajectories.
- Impute potential outcomes using the historical outcome model.

4. Results and Theoretical Findings

Proposition 1 & 2: Establish that if effect modifiers are known and their evolution is stationary (or follows a stable transition model), the future causal effect is identified by marginalizing the observed outcome model over the predicted future distribution of modifiers.
Proposition 3: Extends this to time-varying treatments with carry-over effects. It proves that one only needs to predict modifiers before the treatment starts in the future window; the evolution of modifiers during the treatment can be averaged out conditional on pre-treatment history.
Stationarity vs. Trends: If effect modifiers are stationary, the future effect equals the past effect. If they follow trends or complex dependencies, the identification relies on the recursive projection of these trends (Equation 9 and 19).
History-Conditioned Estimands: The paper introduces an alternative estimand conditioning on the full observed history up to the prediction point, which is more relevant for real-time forecasting than conditioning only on baseline covariates.

5. Significance and Implications

Theoretical Advancement: Moves causal inference beyond static "generalizability" to dynamic "forecasting," providing a rigorous mathematical foundation for policy prediction.
Policy Relevance: Offers a structured way to answer "What if" questions for future interventions (e.g., "What will be the impact of a new lockdown in Winter if we implement it now?"). It moves beyond informal qualitative assumptions that "everything stays the same."
Limitations and Validity: The paper critically highlights that these predictions are only valid if:
- All relevant time-varying effect modifiers are measured.
- No unobserved structural breaks (e.g., new virus variants, sudden behavioral shifts) occur between the observed and future windows.
- The "no-interference" assumption holds (e.g., interventions in one region don't spill over to others).
Broader Application: While motivated by COVID-19, the framework applies to environmental epidemiology (e.g., forecasting health effects of climate policies), economics, and any field requiring the prediction of intervention effects in evolving systems.

Conclusion

Forastiere, Li, and Baccini provide a comprehensive theoretical framework for forecasting causal effects. By formalizing the assumptions of temporal transportability and deriving g-computation formulas, they enable researchers to rigorously project past intervention effects into future contexts, provided that the dynamic mechanisms governing effect modifiers are stable and measurable. This bridges the gap between retrospective policy evaluation and prospective policy planning.