Refining the Notion of No Anticipation in Difference-in-Differences Studies
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Refining the Notion of No Anticipation in Difference-in-Differences Studies".
Jane: Here's the gist of the paper. We address an ambiguity in identification strategies using difference-in-differences, which are widely applied in empirical research, particularly in economics The authors argue that confusion surrounding the no-anticipation assumption arises from ambiguity in the intervention considered and that current formulations of the assumption are ambiguous They propose new definitions and identification results to clarify this widespread ambiguity in difference-in-differences (DiD) and related causal inference methods Pruning redundant reasoning tokens The paper aims to clarify a widespread ambiguity in identification arguments used in difference-in-differences (DiD) and related causal inference methods Some scholars have claimed that DiD methods require an assumption of “no anticipation”
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11: and some authors have even called the no-anticipation assumption the “arrow of time”
14: The authors claim that confusion around anticipation effects stems from an inadequate definition of the intervention considered They propose a solution relying on an expanded causal model and a new formalization of the no-anticipation assumption How the training signal is built The paper first considers a canonical setting with treatment and outcome observed at two time points, defining the average treatment effect among the treated as AT TA2 Under Assumptions 1–3, this estimand is identified as AT TA2 = E(Yk A2 = 1) − E(Yk A1 = 0) Many authors explicitly state an additional assumption: the no-anticipation, or no anticipatory effects, assumption The standard no-anticipation assumption is sometimes called “NEPT”
26, 27: Where the gains come from The authors claim that the standard definition fails to encode the conditions that researchers intend to formalize: that is, to preclude that prior knowledge about the plan of implementing the treatment, and not the treatment itself They introduce a variable P representing the decision (or plan) of implementing the policy Assumption 4 states that A p=p∗ 2 = p∗ for every p∗ ∈
0, 1: They introduce a new formalization of the no-anticipation assumption as Assumption 5: E(Y p=1 P = 1) = E(Y p=0 P = 1) This formulation rules out that the decisions to implement the policies affect outcomes even before the policy is enacted Extension to multiple time points The paper extends results to settings with multiple time points, such as staggered policy adoption and multiple outcome measurements They propose a new estimand for g ≤ τ − s and k ≤ τ: AT TP¯(g, k) This estimand represents the effect, at time k, of first deciding at time g to implement the policy versus deciding during the study period to not implement the policy Under Assumptions 9–13, AT TP¯(g, k) is identified by a functional similar to the one proposed by Callaway and Sant’Anna Proposition 3 shows that this result is a direct generalization of the scenario discussed in Section 5.
Paper summary: Tom: So, to wrap up on "Refining the Notion of No Anticipation in Difference-in-Differences Studies," the authors are essentially saying that we need a better way to define what it means for an intervention not to anticipate its own effect.
Jane: They focus on how the standard definition fails to capture researchers' concerns about prior knowledge regarding policy plans, which they model using a variable P.
Lu: The main implication is that by adopting their new formalization, Assumption five we can derive identification results that better reflect the actual scenario of planning and implementing treatments over time.
Meng: This means for anyone building production systems that test sequential changes—say, updating a recommendation engine based on a plan to change its algorithm—we need to be careful about how we set up our comparison points.
Lalam: I think the core idea is that the standard assumption implies you must explicitly rule out backward causation because of how it's phrased, whereas their refinement allows for a more realistic interpretation of the decision process.
Tom: So, what does this title really suggest? It’s not just about fixing a math problem; it’s about refining the very notion of no anticipation itself in these studies.
Jane: It sounds like they are pushing for a deeper conceptual understanding before we just keep using these identification strategies as they are.
Lu: The paper is valuable because it shows how to move from an ambiguous statement to a formal, testable condition that better captures the intended causal structure of policy adoption over time.
Conclusion: Tom: So we’re wrapping up on "Refining the Notion of No Anticipation in Difference-in-Differences Studies." This paper by some researchers is trying to fix a big confusion around what we mean when we assume no anticipation in these kinds of studies.
Jane: It sounds like they are taking that old assumption, the no anticipation idea, and really unpacking what it actually means for how we set up our comparisons between two different time points.
Lu: They introduce a new way to look at the decision to implement a policy, calling it P, and then formalize the no anticipation idea as this specific condition about what happens when that plan P is made.
Meng: From my side, I’m wondering if this new formalization actually helps us predict when our models will break when we try to deploy them in a real system.
Lalam: The paper suggests that by making this change, the identification results we get are much more accurate for situations where decisions are made sequentially over time.
Tom: So the main point is that the way they define no anticipation matters because it directly impacts whether our results hold up when we look at policy changes happening in stages.
Jane: Exactly. It’s not just about a math trick; it’s about making sure the theoretical assumptions we use actually match what happens in the real world of planning and implementing new systems.
Lu: If this works, it gives us a much clearer tool to analyze things like staggered policy adoption where treatments happen at different times.
Meng: I still have to see if this new setup is practical for running large-scale simulations, but conceptually, defining the decision plan P sounds more robust than the old version.
Lalam: I’m quite confident in this direction; it moves the discussion from a vague statement to a concrete condition we can test against our data.
Tom: It seems like the authors are giving us a better way to measure causality when things happen over time, and that’s pretty important for any serious AI or economics work.
Jane: It really is about sharpening the tools we use to understand how decisions influence outcomes across time.
Institute of Mathematics, Ecole Polytechnique Férale de Lausanne · Department of Statistics and Data Science, The Wharton School, University of Pennsylvania · Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
stat.ME, econ.EM
Submitted: 2025-07-17
Updated: 2026-11-02
Project page: https://bcallaway11.github.io/did
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 75/100
The gist: We address an ambiguity in identification strategies using difference-in-differences, which are widely applied in empirical research…
Key concepts
- No Anticipation Assumption
- This is an assumption in causal inference suggesting that decisions made before a treatment is actually implemented do not affect the outcome. The authors argue the standard definition is too vague, leading to confusion about what exactly needs to be excluded.
- Intervention vs. Decision
- The paper distinguishes between the actual treatment being applied and a researcher's 'decision' or 'plan' to implement that treatment later. They propose separating these concepts to clarify the no-anticipation assumption, focusing on whether prior knowledge of the plan matters.
- New Formalization (Assumption 5)
- The authors introduce a specific mathematical formulation for the no-anticipation assumption: E(Y p=1 | P = 1) = E(Y p=0 | P = 1). This new rule explicitly states that the decision to implement the policy (P=1) does not influence outcomes before the policy is enacted, offering a clearer way to test this condition.
Terminology
Summary
We address an ambiguity in identification strategies using difference-in-differences, which are widely applied in empirical research, particularly in economics The authors argue that confusion surrounding the no-anticipation assumption arises from ambiguity in the intervention considered and that current formulations of the assumption are ambiguous They propose new definitions and identification results to clarify this widespread ambiguity in difference-in-differences (DiD) and related causal inference methods
Pruning redundant reasoning tokens
The paper aims to clarify a widespread ambiguity in identification arguments used in difference-in-differences (DiD) and related causal inference methods Some scholars have claimed that DiD methods require an assumption of “no anticipation” [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11] and some authors have even called the no-anticipation assumption the “arrow of time” [14] The authors claim that confusion around anticipation effects stems from an inadequate definition of the intervention considered They propose a solution relying on an expanded causal model and a new formalization of the no-anticipation assumption
How the training signal is built
The paper first considers a canonical setting with treatment and outcome observed at two time points, defining the average treatment effect among the treated as AT TA2 Under Assumptions 1–3, this estimand is identified as AT TA2 = E(Yk A2 = 1) − E(Yk A1 = 0) Many authors explicitly state an additional assumption: the no-anticipation, or no anticipatory effects, assumption The standard no-anticipation assumption is sometimes called “NEPT” [26, 27]
Where the gains come from
The authors claim that the standard definition fails to encode the conditions that researchers intend to formalize: that is, to preclude that prior knowledge about the plan of implementing the treatment, and not the treatment itself They introduce a variable P representing the decision (or plan) of implementing the policy Assumption 4 states that A p=p∗ 2 = p∗ for every p∗ ∈ [0, 1] They introduce a new formalization of the no-anticipation assumption as Assumption 5: E(Y p=1 P = 1) = E(Y p=0 P = 1) This formulation rules out that the decisions to implement the policies affect outcomes even before the policy is enacted
Extension to multiple time points
The paper extends results to settings with multiple time points, such as staggered policy adoption and multiple outcome measurements They propose a new estimand for g ≤ τ − s and k ≤ τ: AT TP¯(g, k) This estimand represents the effect, at time k, of first deciding at time g to implement the policy versus deciding during the study period to not implement the policy Under Assumptions 9–13, AT TP¯(g, k) is identified by a functional similar to the one proposed by Callaway and Sant’Anna Proposition 3 shows that this result is a direct generalization of the scenario discussed in Section 5.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved identification of treatment effects by explicitly modeling decision variables P instead of just implementation A2. This allows researchers to distinguish between
the effect of implementing the policy
(AT TA2) andthe effect of a policy decision
(AT TP). -
Development of an expanded causal model that includes the variable P, representing the plan to implement the treatment. This formalization clarifies that confusion arises from ambiguity in whether prior knowledge about a plan affects outcomes before implementation.
-
Creation of a new no-anticipation assumption, Assumption 5, which rules out
the average effect of the decision P on the first outcome measurement Y1, for units that decided to implement the policy (P = 1).
This addresses concerns thatprior knowledge about the plan of implementing the treatment
affects pre-treatment outcomes. -
Identification of a functional for AT TP using Assumptions 1, 4, 5, 6, and 7. This functional allows for identifying effects based on the decision to implement versus not implement, even when there is anticipation present in Y1.
Abstract
We clarify the interpretation of identification strategies using difference-in-differences (DiD), which are widely applied in empirical research. In particular, DiD researchers often invoke an assumption commonly referred to as the "no-anticipation assumption". Taken literally, this assumption states that treatments in the future have no effect on outcomes in the past. Because standard causal models rely on a temporal structure in which causes precede effects, such an assumption seems to be inherently satisfied. However, most researchers attribute a different meaning to the assumption: they interpret the assumption as ensuring no effect of policy decision, e.g., immediate behavioral changes due to knowledge about the plan to implement the policy. We discuss the consequences of this mismatch, and give an alternative formulation of no anticipation that captures DiD researchers' concerns and clarifies identification arguments.
Sources
- Transporting treatment effects from difference-in-differences studies
- Bounds for Treatment Effects in the Presence of Anticipatory Behavior
- Applied Causal Inference Powered by ML and AI
- A Nonparametric Framework for Universal Difference-in-Differences
- Proximal Causal Inference for Conditional Separable Effects
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes