1 The Hidden Assumption in Every DiD Regression   

The difference-in-differences (DiD) estimator is the workhorse of applied causal inference. Its logic is so familiar it feels like arithmetic: subtract the treated group's pre-period mean from its post-period mean, subtract the same difference for the control group, and the difference of differences is the average treatment effect on the treated (ATT). The identifying assumption, parallel trends, says that absent treatment the two groups' mean outcomes would have moved in lockstep.

Yet parallel trends carries a stowaway most users never inspect: it is not invariant to how the outcome is scaled. Parallel trends in levels of earnings is a different assumption from parallel trends in log earnings, and they cannot both hold except in knife-edge cases. A DiD that is "valid" in dollars can be "invalid" in log dollars, and the data cannot tell you which scale is right. For an outcome with no natural cardinalisation- a test score, a happiness index, a price this is deeply uncomfortable. We are letting an arbitrary choice of units drive a causal conclusion.

Athey and Imbens [2006] confronted this head-on with the changes-in-changes (CIC) model, one of the most elegant and underused ideas in the DiD canon. CIC delivers an estimator of the entire counterfactual distribution not just the mean that is invariant to monotone rescalings of the outcome, and it nests standard DiD as a special case. A recent revival, recasting CIC through the lens of optimal transport, has extended it to multivariate outcomes and reconnected it to the modern distributional-effects literature [Torous et al., 2024].  

2 A Model of Outcomes, Not of Means   

The conceptual move in CIC is to model the outcome-generating process rather than the trend in means. Let G ∈ {0,1} index the control and treated groups and T ∈ {0,1} index the pre- and post-periods. Athey and Imbens [2006] posit that the untreated (no-treatment) potential outcome is produced by   

where U is an unobserved scalar capturing all individual heterogeneity ("ability", "health", "latent demand") and h(·,t) is strictly increasing in U for each t. The production function h may change arbitrarily between periods this is how the world evolves over time but it is common to both groups. The groups differ only in the distribution of U, which is allowed to be arbitrary, and crucially that distribution is held fixed within a group over time:   

Assumption (2) replaces parallel trends. It says that the population of each group does not change in unobservable composition between periods, and that both groups experience the same time transformation h(·,0) → h(·,1), whatever its shape. Because h need not be additive or even linear, CIC is invariant to any monotone transformation of Y: if the model holds for Y, it holds for log Y, √Y, or any strictly increasing relabelling.  

3 Identification via the Quantile-Quantile Transform   

The payoff is a closed-form expression for the counterfactual distribution of the treated group's untreated outcome in the post-period. The construction is intuitive. The control group reveals how the time transformation h(·,0) → h(·,1) moves an individual from one point in the distribution to another. We learn this map by lining up the control group's period-0 and period-1 distributions quantile by quantile:   

where F_Y,gt is the observed CDF of Y in group g, period t. The function k(·) takes a control unit's period-0 outcome and returns the period-1 outcome of the control unit at the same rank. We then apply this transform to the treated group's period-0 outcomes to obtain their counterfactual period-1 outcomes. The implied counterfactual CDF is   

and the average treatment effect on the treated is   

Because (4) delivers the whole counterfactual distribution, CIC identifies not only the mean ATT but quantile treatment effects on the treated: compare the observed treated quantiles to the counterfactual quantiles, and read off how the treatment reshaped the distribution did it lift the bottom, stretch the top, or shift everything uniformly? Standard DiD as a special case. Suppose the production function is additive in time, h(u,t) = u + δ_t Then k(y) = y + (δ₁ - δ₀) is a pure location shift, equation (5) collapses to the textbook double difference of means, and the quantile effects are constant across the distribution. Standard DiD is therefore the CIC model under the extra, often-implausible restriction that time affects everyone by the same additive amount regardless of their position in the distribution. CIC exposes this restriction and lets the data relax it.  

4 The Optimal Transport Revival   

Equation (3) is a one-dimensional optimal transport map: the increasing rearrangement that pushes the control's period-0 distribution onto its period-1 distribution is exactly the L²-optimal coupling on the real line. This observation, made precise by Torous et al. [2024], unlocks a powerful generalisation. When the outcome is multivariate- vector of test scores, a bundle of health markers, a consumption basket the scalar quantile transform no longer exists, but the optimal transport (Monge) map between the control distributions does. Replacing k(·) with the multivariate optimal transport map yields a nonlinear, distribution- respecting DiD for vector outcomes, with the univariate CIC of Athey and Imbens [2006] as the one-dimensional special case.

This reframing connects CIC to a fast-moving toolkit. Optimal transport now underpins distributional synthetic controls, counterfactual distribution estimation, and fairness- constrained prediction; importing its computational machinery (entropic regularisation, Sinkhorn iterations) makes multivariate CIC tractable at scale. It also clarifies the assumption: the key requirement becomes that the optimal transport map describing untreated evolution is common across groups -a clean geometric statement of "common dynamics" that generalises parallel trends.  

5 Costs, Limits, and Why It Is Underused

If CIC is so principled, why does it not appear in every empirical paper? Several frictions explain its niche status.  

• Support requirements. Identification of the counterfactual at a point y requires that the corresponding quantile be present in the control group's period-0 distribution. Where supports do not overlap, the counterfactual is only partially identified, and Athey and Imbens [2006] provide bounds rather than point estimates.  
• Scalar unobserved heterogeneity. The baseline model compresses all heterogeneity into a single index U with a monotone production function. Rich, multidimensional selection on unobservables can break this though the optimal transport extension relaxes the scalar restriction.  
• Staggered adoption. The recent revolution in DiD has been about staggered treatment timing and heterogeneous effects [Callaway and Sant'Anna, 2021, de Chaisemartin and D'Haultfœuille, 2020]. Extending CIC to many groups and many periods is an active but incomplete project, and practitioners reach for the better-tooled group-time ATT estimators instead.  
• Inference. Distributional estimands require functional inference uniform confidence bands over quantiles which is more delicate than a single clustered standard error.  

None of these is fatal, and the central message survives. The comfortable habit of running a levels-or-logs DiD and trusting parallel trends conceals a scale-dependence that should worry anyone reporting effects on an arbitrarily measured outcome. Changes-in-changes shows that we can do better: model the data-generating process, respect the ordinal content of the outcome, and recover the full counterfactual distribution. Two decades after Athey and Imbens [2006], the optimal transport revival has given the idea fresh computational legs and a fresh claim on the attention of applied researchers who care about who a policy affects, not just the average.  

References

1. Athey, S., and Imbens, G. W. (2006). Identification and inference in nonlinear difference-in-differences models. Econometrica, 74(2), 431-497.  
2. Callaway, B., and Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200-230.  
3. de Chaisemartin, C., and D'Haultfœuille, X. (2020). Two-way fixed effects estimators with heterogeneous treatment effects. American Economic Review, 110(9), 2964-2996.  
4. Roth, J., and Sant'Anna, P. H. C. (2023). When is parallel trends sensitive to functional form? Econometrica, 91(2), 737-747.  
5. Torous, W., Gunsilius, F., and Rigollet, P. (2024). An optimal transport approach to estimating causal effects via nonlinear difference-in-differences. Journal of Causal Inference, 12(1), 20230004.