Non-Additive Difference-in-Differences

• The paper introduces NA-DiD as a framework that departs from additive linear models to capture heterogeneous and nonlinear treatment effects via regime-specific potential outcomes.
• It employs alternative identification strategies such as odds ratio equi-confounding, geodesic means, and exposure mappings to address issues like interference, continuous treatments, and non-Euclidean outcomes.
• NA-DiD enhances causal inference robustness by tailoring estimands to varying treatment regimes, addressing missing data challenges and validating sensitivity analysis under non-additive conditions.

Non-Additive Difference-in-Differences (NA-DiD) denotes a family of difference-in-differences frameworks that depart from the idea that causal identification and aggregation must be anchored to an additive linear model or a constant treatment effect. In this literature, treatment effects may vary with treatment level, group, period, exposure, latent response type, or outcome geometry, and identification is formulated directly with potential outcomes rather than through specifications such as $Y_{it}=\alpha_i+\lambda_t+\beta D_{it}+u_{it}$ (Chaisemartin et al., 2024). Some contributions replace additive parallel trends with odds ratio equi-confounding, some show that the familiar two-by-two DID algebra identifies different causal objects in non-canonical treatment regimes, some extend switcher logic to continuous treatments with no stayers, and some reinterpret DID when interference, missingness, or non-Euclidean outcomes destroy the usual additive subtraction of means (Park et al., 2022, Shahn et al., 2024, Zhou et al., 29 Jan 2025). A separate paper uses the label “NA-DiD” literally, replacing additive aggregation by the Choquet integral under a capacity (Halkiewicz, 16 Jul 2025). This suggests not a single formal model, but a research program in which the meaning of a DID contrast is allowed to depend on treatment regime, outcome scale, exposure environment, and aggregation rule.

1. Classical template and the non-additive departure

The canonical two-group, two-period DID contrast is

$E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$

Under the canonical treatment pattern $A_0=0,\ A_1=G$ and group parallel trends

$E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$

this identifies

$E[Y_1(1)-Y_1(0)\mid G=1].$

That is the standard ATT interpretation of two-by-two DID (Shahn et al., 2024).

NA-DiD begins where that algebra is no longer read through a single additive bias model or a constant coefficient. One sensitivity-analysis paper makes the classical structure explicit by writing untreated control potential outcomes as additive in a baseline term $\mu_i$, a constant group difference $\beta_i$, and a uniform time trend $\alpha_i$, and then stresses that DID’s bias-reducing properties are “functional form dependent” and that “the additive pattern of distorting effects comes and goes with strictly monotone transformations of the response” (Keele et al., 2019). The continuous-treatment DID literature makes the same point from a different direction: two-way fixed effects with a continuous regressor is said to rely on linearity and homogeneous effects, whereas the non-additive alternative works directly with $Y_t(d)$ and allows treatment effects to be heterogeneous and nonlinear (Chaisemartin et al., 2024).

The central conceptual shift is therefore from additive outcome decompositions to regime-specific potential-outcome restrictions. In NA-DiD, the same observed difference in changes may represent an ATT, a contrast of treatment effects across groups, a contrast of treatment effects across periods, a weighted average of dose-response slopes, or a net difference between direct and spillover effects. The algebra of differencing is stable; the identified causal object is not.

2. Scale-free and regime-specific identification

One major strand replaces additive parallel trends with a scale-invariant restriction on the association between treatment and untreated potential outcomes. In the “universal” DID framework, the key assumption is odds ratio equi-confounding,

$\alpha_0^*(y,x)=\alpha_1^*(y,x),$

which states that the generalized odds ratio relating treatment and the untreated potential outcome is stable across periods. Under consistency, no anticipation, overlap, and OREC, the post-treatment untreated distribution for treated units is identified, not just its mean, so the framework can target the ATT, transformed-outcome means, medians, quantiles, the QTT, and more generally any estimand defined by a moment equation. The papers emphasize that OREC is scale-invariant under monotone transformations and that PT and OREC are generally non-nested (Park et al., 2022, Tchetgen et al., 2023).

A second strand shows that the same two-by-two algebra can identify different causal objects when the treatment regime is non-canonical. With

$E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$0

the canonical design $E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$1 identifies the period-1 ATT. But under the pre-post design $E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$2, GDiD identifies

$E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$3

a difference in second-period ATTs across groups. Under the always-treated versus never-treated design $E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$4, it identifies

$E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$5

an intertemporal effect contrast among treated units (Shahn et al., 2024). A recurrent misconception in DID practice is therefore that a DID-style coefficient always estimates an ATT level. In NA-DiD, the treatment regime itself determines which causal remainder survives the differencing algebra.

A third strand keeps parallel trends but abandons additive-linear estimation. DiD-BCF models

$E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$6

where $E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$7, so untreated outcomes and treatment effects are represented by flexible nonlinear functions rather than fixed effects plus a constant coefficient. Its PTA-based reparameterization forces the pre-treatment effect to zero structurally through $E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$8, and the framework is presented as a unified estimator of average, group-average, and conditional treatment effects under staggered adoption and heterogeneous treatment effects (Souto et al., 14 May 2025). This does not replace parallel trends; it replaces additive outcome modeling.

3. Continuous treatments and no-stayer settings

In continuous-treatment DID, NA-DiD moves from discrete treated-versus-untreated states to path-specific dose responses. In a two-period setup with scalar doses $E[Y_1-Y_0\mid G=1]-E[Y_1-Y_0\mid G=0].$9, outcomes $A_0=0,\ A_1=G$0, and $A_0=0,\ A_1=G$1, one paper targets the WAOSS,

$A_0=0,\ A_1=G$2

equivalently

$A_0=0,\ A_1=G$3

This is a weighted average of observed secant slopes, with weights proportional to $A_0=0,\ A_1=G$4, rather than a constant effect or a global derivative. The distinctive case is $A_0=0,\ A_1=G$5: there are no exact stayers, but there are quasi-stayers because $A_0=0,\ A_1=G$6 for all $A_0=0,\ A_1=G$7. Under a static model, generalized parallel trends,

$A_0=0,\ A_1=G$8

and Lipschitz continuity of potential outcomes in treatment, the effect is identified by

$A_0=0,\ A_1=G$9

where the quasi-stayer term $E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$0 reconstructs the missing stayer counterfactual (Chaisemartin et al., 2024). The paper’s practical message is that non-additive switcher logic survives even when exact stayers do not exist.

A later contribution generalizes continuous-treatment DID from deterministic dose contrasts to stochastic policy shifts. It defines the average stochastic dose effect among the treated,

$E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$1

and identifies it under untreated parallel trends plus dose-specific parallel trends. For the exponential tilt intervention,

$E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$2

the paper derives a nonparametric efficient one-step estimator with machine-learning nuisance estimation, root-$E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$3 consistency, asymptotic normality, and the nonparametric efficiency bound (Jetsupphasuk et al., 29 Nov 2025). Here the non-additive feature is not only nonlinear dose response, but also a policy estimand defined by a counterfactual dose distribution rather than a fixed dosage.

4. Interference, exposure mappings, and network dependence

Once outcomes depend on other units’ treatment, the classical DID contrast ceases to identify a single own-treatment effect. A technical note on unknown interference shows that, under no anticipation, consistency, and a no-treatment-for-anyone parallel trends condition, canonical DID identifies

$E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$4

where $E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$5 is a total average treatment effect on treated units and $E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$6 is an average spillover effect on control units. Without further assumptions, the DID estimand is informative only about their contrast, not about either quantity separately (Mealli et al., 24 Dec 2025). This is a direct non-additive result: the treatment-response surface depends on the full assignment environment, not only on own treatment.

More structured interference designs index effects by exposure. In one paper, exposure is summarized by $E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$7, and the main direct-effect estimand at exposure level $E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$8 is

$E[Y_1(0)-Y_0(0)\mid G=1]=E[Y_1(0)-Y_0(0)\mid G=0],$9

The same framework defines spillover contrasts $E[Y_1(1)-Y_1(0)\mid G=1].$0 and $E[Y_1(1)-Y_1(0)\mid G=1].$1, uses exposure-specific parallel trends, and develops IPW, regression, and doubly robust DID estimators. It later relaxes immediate-neighborhood interference and correct exposure-map specification through Approximate Neighborhood Interference and expected DATT objects (Xu, 2023).

Related work extends the estimand from static exposure states to exposure histories. One paper defines

$E[Y_1(1)-Y_1(0)\mid G=1].$2

derives a doubly robust estimator, and shows consistency, asymptotic normality, and semiparametric efficiency under conditional parallel trends and network dependence (Jetsupphasuk et al., 5 Feb 2025). Another paper defines

$E[Y_1(1)-Y_1(0)\mid G=1].$3

and

$E[Y_1(1)-Y_1(0)\mid G=1].$4

then estimates nuisance functions with graph neural networks because the relevant confounding may depend on the entire covariate matrix and network structure rather than on low-dimensional summaries (Sun et al., 29 Sep 2025). Across these papers, the treated-versus-control dichotomy is replaced by exposure-indexed direct and spillover effects.

5. Outcomes beyond Euclidean spaces and non-additive aggregation

A different meaning of NA-DiD appears when the outcome space itself is non-additive. For outcomes in a unique geodesic metric space $E[Y_1(1)-Y_1(0)\mid G=1].$5, arithmetic subtraction is unavailable, so one paper replaces differences by geodesics. With group-time Fréchet means $E[Y_1(1)-Y_1(0)\mid G=1].$6 and $E[Y_1(1)-Y_1(0)\mid G=1].$7, the causal estimand is the geodesic average treatment effect on the treated,

$E[Y_1(1)-Y_1(0)\mid G=1].$8

Identification uses a geometric parallel trends assumption,

$E[Y_1(1)-Y_1(0)\mid G=1].$9

so that the counterfactual untreated post-period Fréchet mean of the treated group is obtained by transporting the control geodesic trend to the treated group’s pre-period baseline. The plug-in estimator is

$\mu_i$0

and the paper derives convergence rates on a space of geodesics. The framework is applied to age-at-death distributions and electricity-generation compositions (Zhou et al., 29 Jan 2025). Here “non-additive” means that the outcome space lacks vector addition and subtraction.

Another paper uses “non-additive” literally at the level of effect aggregation. It rewrites classical DID as an integral of time-specific treated-control differences under an additive measure $\mu_i$1,

$\mu_i$2

and then replaces $\mu_i$3 with a capacity $\mu_i$4, yielding

$\mu_i$5

The integral is understood in the Choquet sense. This allows synergy, redundancy, ambiguity, and compliance erosion to enter through the aggregation rule rather than through a linear average. The paper states that standard DID is recovered when $\mu_i$6 is additive, but it also notes unresolved issues: capacity estimation did not work satisfactorily in the DID application, the Julia code implements a simplified proof-of-concept rather than the full time-indexed formula, and the identification theory is incomplete (Halkiewicz, 16 Jul 2025). This strand is conceptually distinct from scale-free or exposure-based DID, because the non-additivity lies in the measure used to aggregate effects.

6. Missingness, sensitivity, and interpretive cautions

Selective missing outcomes create a further non-additive mixture problem because observed DID contrasts aggregate latent response types. In a two-period design with post-treatment response indicator $\mu_i$7, one paper defines principal strata

$\mu_i$8

and allows both untreated trends and treatment effects to vary across the latent types $\mu_i$9, $\beta_i$0, $\beta_i$1, and $\beta_i$2. It shows that complete-case DID additionally requires

$\beta_i$3

a strong restriction because, for treated units, the left-hand side contains both untreated trend and treatment effect. The proposed alternative imposes principal-strata parallel trends,

$\beta_i$4

uses response rates over time to identify principal-stratum shares under monotonicity or related assumptions, and tailors Lee bounds to partially identify the always-responder effect ATT-AR (Shin, 2024). The practical lesson is that missingness can turn DID into a latent-mixture design even when treatment itself is binary.

Sensitivity analysis papers add a complementary warning. One matched-DID paper states standard DID through a model of additive bias, distinguishes hidden bias in treatment assignment from hidden bias due to departures from additive DID structure, and shows that a DID sensitivity analysis at $\beta_i$5 can be implemented using conventional paired-study sensitivity methods at

$\beta_i$6

(Keele et al., 2019). A later paper, described explicitly as not a direct NA-DiD identification contribution, uses Riesz representation and Double Machine Learning to quantify how far standard DiD estimates may move when conditional parallel trends fails because of latent pre-treatment confounders. In the canonical $\beta_i$7 case it derives

$\beta_i$8

and extends the same logic to $\beta_i$9 in staggered-adoption settings (Bach et al., 10 Oct 2025). These papers do not redefine the DID estimand, but they are part of the NA-DiD conversation because they clarify how fragile additive or mean-parallel-trends interpretations can be.

Several recurrent misconceptions are therefore rejected across the literature. The same DID algebra does not always identify the ATT; under non-canonical regimes it may identify effect heterogeneity, under interference it may identify a contrast between direct and spillover effects, and under selective response it may identify a selected latent-stratum mixture rather than a population effect. Likewise, the absence of exact stayers does not eliminate continuous-treatment DID if quasi-stayers exist, and the absence of linearity does not eliminate DID if the identifying restriction is reformulated on an odds-ratio, exposure-history, or geometric scale. This suggests that NA-DiD is best understood not as a single estimator, but as a systematic generalization of DID to settings in which additivity of outcomes, effects, exposure, or aggregation is no longer credible.