Forecasting Causal Effects of Future Interventions: Confounding and Transportability Issues

Recent developments in causal inference allow us to transport a causal effect of a time-fixed treatment from a randomized trial to a target population across space but within the same time frame. In contrast to transportability across space, transporting causal effects across time or forecasting causal effects of future interventions is more challenging due to time-varying confounders and time-varying effect modifiers. In this article, we seek to formally clarify the causal estimands for forecasting causal effects over time and the structural assumptions required to identify these estimands. Specifically, we develop a set of novel nonparametric identification formulas–g-computation formulas–for these causal estimands, and lay out the conditions required to accurately forecast causal effects from a past observed sample to a future population in a future time window. Our overarching objective is to leverage the modern causal inference theory to provide a theoretical framework for investigating whether the effects seen in a past sample would carry over to a new future population. Throughout the article, a working example addressing the effect of public policies or social events on COVID-related deaths is considered to contextualize the developments of analytical results.

💡 Research Summary

The paper addresses a gap in causal inference literature: while existing work on generalizability and transportability focuses on spatial transfer of effects for time‑fixed treatments, real‑world policy evaluation often requires forecasting the impact of interventions into future time periods where both covariates and effect modifiers may evolve. The authors formalize this “temporal transportability” problem, define appropriate causal estimands, and derive non‑parametric identification results using g‑computation.

First, the authors introduce a longitudinal setting with units (e.g., geographic regions) observed over T time points. For each unit i and time t they denote exposure S_it, treatment indicator Z_it (policy or event), time‑varying covariates X_it, and outcome Y_it (e.g., COVID‑19 deaths). They distinguish between two estimands: (1) an evaluation estimand that quantifies the average potential outcome under a specified treatment trajectory within the observed window, and (2) a prediction estimand that projects the same trajectory into a future window t* > T.

Identification relies on three core assumptions. The first is a standard ignorability condition: conditional on observed past covariates, exposures, and treatment history, the treatment at each time point is independent of future potential outcomes. The second, novel “temporal transportability” assumption, requires that the structural causal mechanisms governing the relationships among S, Z, X, and Y remain invariant across time, or that any change can be accurately modeled using external information. The third assumption is the stability of the exposure‑response function μ_t(s, x) across periods. Under these conditions, the authors derive a recursive g‑computation formula that integrates over the joint distribution of future covariates and exposures, which are estimated from past data and then “transported” to the future.

The paper extends the basic framework to handle time‑varying treatments with histories, distinguishing between historical effects (past treatments influencing current covariates) and duration effects (persistent influence of a treatment over time). For such settings, the identification conditions become more intricate, requiring additional assumptions about the absence of unmeasured time‑varying confounders that are themselves affected by earlier treatments.

A concrete motivating example involves COVID‑19 policy evaluation. The authors illustrate how to estimate the effect of spring‑2020 social‑distancing measures on daily death counts, then forecast the effect of applying the same measures during a later wave when population behavior and vaccination coverage have changed. They compare their g‑computation based forecasts with traditional system‑dynamics and agent‑based simulations, showing that the causal approach yields comparable predictions while providing a clear, assumption‑driven interpretation.

The authors also discuss practical threats to identification: (i) omitted time‑varying confounders, (ii) shifts in effect modifiers that are not captured in the observed data, and (iii) exogenous shocks such as new viral variants. To address these, they propose sensitivity‑analysis techniques, partial identification strategies, and the incorporation of auxiliary data sources (e.g., mobility data, census updates) within a Bayesian hierarchical framework to quantify uncertainty.

In conclusion, the paper contributes a rigorous theoretical framework for forecasting causal effects of future interventions, bridging the gap between static transportability theory and dynamic policy forecasting. It highlights the importance of explicit structural assumptions, offers practical g‑computation formulas for implementation, and points to future work on multi‑region, multi‑time‑point extensions, real‑time updating with streaming data, and robust methods for handling non‑stationary external shocks.