betaFeedback

WorksEconomics, Econometrics and Finance

Asymptotic Theory of Integrated Pairwise-Difference Functionals of Kernel Estimators, with Applications to Treatment Effect Heterogeneity

Abstract

This paper develops asymptotic theory for integrated pairwise-difference statistics constructed from kernel estimators, with applications to treatment effect heterogeneity. The statistics integrate absolute or positive-part differences between kernel estimates evaluated at pairs of points. Their two-point structure falls outside existing one-point L1L_1-type functional theory. We establish asymptotic normality using a Hájek-type projection of a Poissonized statistic and develop feasible centering and variance estimators. We consider two applications. First, we test whether treatment effects vary with one covariate block while holding another fixed. Second, we establish root-NN inference for the directional Gini mean difference of the conditional treatment effect function, which measures heterogeneity in outcome units. The framework also accommodates tests of monotone effect modification and estimated pseudo-outcomes. Simulations corroborate the asymptotic theory and show good finite-sample performance. An application to 401(k) eligibility and net financial assets illustrates the usefulness of the proposed tests.

Introduction

Many questions in econometrics concern how an unknown function varies across its arguments. Examples include whether treatment effects vary with individual characteristics (Crump et al., 2008), whether particular covariates contribute to a conditional mean function (Lavergne and Vuong, 2000), and how household demand responds to expenditure and demographic characteristics (Blundell et al., 2003). Integrated pairwise difference functionals naturally quantify such variation, capturing both magnitude and direction. When the underlying function is estimated nonparametrically, however, the asymptotic properties of these functionals are not immediate. Unlike conventional integrated functionals of kernel estimators evaluated at a single point, integrated pairwise-difference statistics involve jointly smoothed estimates at two points and therefore fall outside existing one-point L1L_1-type functional asymptotic theory. This paper develops asymptotic theory for integrated pairwise-difference statistics constructed from kernel estimators. The results provide tools to test and measure variation in unknown functions, with treatment effect heterogeneity as a leading application.

In the analysis of treatment effect heterogeneity, an important question is not simply whether treatment effects vary, but which covariates account for that variation. Identifying these covariates is important for constructing welfare-maximizing treatment assignment rules and assessing the external validity of treatment effect estimates. If treatment effects do not vary with one block of covariates conditional on another, a treatment rule that already uses the second block gains nothing from incorporating the first. Similarly, transporting an average treatment effect to another population requires reweighting only on the second block, provided the conditional treatment effect is transportable (Dahabreh et al., 2020). Let X=(X1,X2)X = (X_1, X_2) denote a partition of covariates and τ(x1,x2)\tau(x_1, x_2) the conditional average treatment effect (CATE). As an application of the asymptotic theory developed below, we test whether τ(x1,x2)\tau(x_1, x_2) varies with x1x_1 holding x2x_2 fixed, and conversely, denoting these hypotheses by H01H_0^1 and H02H_0^2. This conditional approach addresses a fundamental limitation of tests based on marginal treatment effects, such as τ1marg(x1)=E[τ(X1,X2)∣X1=x1]\tau^{\mathrm{marg}}_1(x_1)=E[\tau(X_1,X_2)\mid X_1=x_1], which average over one covariate block before assessing variation in the other. Such marginalization can conceal genuine treatment effect heterogeneity or attribute variation to covariates that do not independently modify treatment effects. For example, suppose τ(x1,x2)=1+β(x1−1/2)(x2−1/2)\tau(x_1,x_2)=1+\beta(x_1-1/2)(x_2-1/2), where β≠0\beta\ne0, and X1X_1 and X2X_2 are independent and uniformly distributed. Although treatment effects vary with both covariates, both marginal treatment effect functions are constant at one, so tests based on these functions, hereafter referred to as marginal tests, fail to detect this heterogeneity. Conversely, if treatment effects depend only on X2X_2 and the covariate blocks are correlated, τ1marg(x1)\tau^{\mathrm{marg}}_1(x_1) can vary with x1x_1 even though X1X_1 does not independently modify treatment effects. In our simulations, this leads a marginal test to spuriously attribute heterogeneity to X1X_1. The integrated pairwise difference statistics avoid these shortcomings by assessing treatment effect variation across values of one covariate block while holding the other fixed, rather than averaging over it.

The proposed statistics integrate absolute or positive-part differences between kernel estimates of τ\tau evaluated at different values of one covariate block while holding the other fixed. For absolute differences, the population counterpart is a directional Gini mean difference (Yitzhaki, 2003) of the functions τ(⋅,x2)\tau(\cdot,x_2), which equals zero if and only if H01H^1_0 holds. For positive-part differences over ordered pairs, it equals zero if and only if these functions satisfy the corresponding monotonicity restriction. Beyond hypothesis testing, the absolute-difference functional measures treatment-effect heterogeneity in outcome units and corresponds to a directional Gini mean difference of the CATE function.

The paper makes four main contributions. First, we establish asymptotic normality and consistency for the two-point L1L_1-type functionals. Because each integrand involves kernel estimates at two potentially distant points, the arguments used to establish existing one-point L1L_1 asymptotic results do not apply directly. The difficulty is analogous to the distinction between the asymptotic theory of sample averages and that of order-two U-statistics. Once the sample size is Poissonized, the integrands of a one-point L1L_1 statistic are independent whenever their kernel windows are disjoint (eg., see (eg., see Giné et al., 2003). Our statistics instead integrate over pairs of evaluation points, so two integrands can remain dependent when just one window from each pair overlaps, regardless of the separation between the remaining points, as with two U-statistic terms sharing one index. These pairs determine the leading variance and rule out the one-point theory’s blocking argument. We overcome this difficulty through a Hájek-type projection of the Poissonized statistic that reduces the pairwise structure to terms depending on individual evaluation points. A finite-sample centering correction for bias, combined with a variance estimator that uses realized kernel weights, improves the accuracy of the normal approximation. Second, we establish consistency and asymptotic normality for integrated pairwise-difference statistics based on pseudo-outcomes with estimated nuisance parameters, providing conditions under which first-stage estimation does not affect their limiting distributions under either H01H_0^1 or H02H_0^2. A drift lemma shows that, under these null hypotheses, first-stage estimation error affects the absolute-difference statistic only at quadratic order, whereas one-point positive-part statistics such as that of Shi et al. (2019, Theorem 4.3) are affected at first order. These results accommodate a root-NN-consistent parametric first-stage estimator on the same sample, provided either a doubly robust working model or the propensity model for the inverse probability weighted pseudo-outcome is correctly specified, without requiring an asymptotically linear representation of the first-stage estimator. They also accommodate nuisance functions estimated on an independent sample under a product-rate condition distinct from the root-NN requirement in double machine learning (Chernozhukov et al., 2018). In both settings, the resulting tests retain their asymptotic level. Third, we develop a multiple testing procedure to determine which covariate blocks account for variation in treatment effects by jointly testing H01H_0^1 and H02H_0^2, while controlling the familywise error rate (FWER). Rejecting H01H_0^1 provides evidence that X1X_1 modifies treatment effects conditional on X2X_2. Fourth, we develop bounds on the magnitude of treatment effect variation in outcome units and establish root-NN inference for the directional Gini mean difference.

The hypotheses H01H^{1}_{0} and H02H^{2}_{0} coincide with those studied in nonparametric significance testing (Delgado and González Manteiga, 2001; Fan and Li, 1996; Lavergne and Vuong, 2000; Lavergne et al., 2015) and conditional mean independence testing (Lundborg et al., 2024; Scheidegger et al., 2022), when the treatment effect pseudo-outcome is used as the response variable, as in the tests of effect invariance across environments of Saengkyongam et al. (2024). Among recent methods, the projected covariance measure (PCM) of Lundborg et al. (2024) and the weighted generalized covariance measure (WGCM) of Scheidegger et al. (2022) are state of the art in conditional mean independence testing. These methods can detect alternatives at faster rates in directions captured by their respective statistics. Both estimate a direction (a projection for PCM or a weight function for WGCM) on one subsample and evaluate the test statistic on another, using sample splitting to ensure the independence required for validity. Our proposed statistics instead have nondegenerate null limiting distributions without sample splitting, allowing every observation to contribute to both CATE estimation and testing, without making the result depend on a random partition of the sample. In simulations, our joint L1L_{1} pairwise absolute-difference test has competitive power in settings favorable to PCM and WGCM, while remaining well calibrated and powerful in settings where they perform less well. The WGCM, whose target is a weighted covariance between X1X_{1} and the CATE, has no power when the effect varies symmetrically in X1X_{1}, whereas our proposed joint test detects this heterogeneity and attributes it to X1X_{1}; and with the additive learners of its replication code, the PCM has no power against a pure interaction between the two blocks, which our test also detects. With both known and estimated pseudo-outcomes, the proposed test maintains rejection probabilities close to the nominal level under the null, whereas the PCM, even when aggregated over multiple random splits to reduce split dependence, is markedly conservative.

The L1L_{1} pairwise absolute-difference framework also yields a direct measure of the magnitude of treatment effect heterogeneity, a one-sided statistic for testing monotone effect modification, and asymptotic level control when nuisance parameters are estimated on the same sample. In an application to the effect of 401(k) eligibility on net financial assets (Chernozhukov and Hansen, 2004), treatment effects vary with income holding age fixed across all specifications, and with age holding income fixed at every recommended bandwidth. After accounting for the noise-induced bias, the estimated magnitude of income-related heterogeneity, approximately 2,700 USD, is distinguishable from zero, whereas that of age-related heterogeneity is not.

Our application to treatment effect heterogeneity attribution also relates to the broader literature on testing treatment effect heterogeneity. Most existing tests of treatment effect heterogeneity consider either an omnibus null, under which the CATE is constant across all covariates, or a marginal null, under which the CATE averaged over the remaining covariates is constant in a chosen subvector. The best linear predictor of Chernozhukov et al. (2025), whose proxy is learned on an auxiliary sample and whose inference is aggregated over many sample splits, and the rank-weighted average treatment effects of Yadlowsky et al. (2025), both built on a proxy learned from all covariates, are omnibus in that use. The integrated conditional moment tests of Lu and Song (2026) and Lapenta et al. (2026) and the linear-functional tests of Dukes et al. (2026) test constancy given a chosen subvector, with estimated nuisances; they are omnibus when the subvector is (X1,X2)(X_1, X_2) and marginal when it is X1X_1. Chang et al. (2015) and Hsu (2017) test zero, sign and dominance nulls. Leave-covariate-out importance measures (Hines et al., 2025 and Williamson et al., 2023) share the nulls H01H_0^1 and H02H_0^2 but have an influence function that degenerates at them, and the sign test of Shi et al. (2019), whose null is implied by H01H_0^1, has rejection probability tending to zero at every point of H01H_0^1. Level results with estimated nuisances exist for a one-point kernel statistic with a parametric first stage (Shi et al., 2019) and for a Kolmogorov–Smirnov statistic of inverse-probability-weighted moments indexed by hypercubes, with a series logit propensity score estimated on the same sample (Hsu, 2017). Estimators of the CATE on a subvector with an estimated first stage are given by Abrevaya et al. (2015), Fan et al. (2022), Lee et al. (2017) and Semenova and Chernozhukov (2021). Our general statistic also accommodates tests of monotone effect modification, relating it to tests of regression monotonicity (Chetverikov, 2019; Ghosal et al., 2000). Our asymptotic analysis builds on the theory of integrated L1L_1 statistics of kernel estimators evaluated at individual points (Giné et al., 2003; Chang et al., 2015; Lee et al., 2013), which we extend to pairwise differences. Our asymptotic proofs rely on the modern Poissonization technique employed in the one-point integrated L1L_1 theory of Giné et al. (2003), Chang et al. (2015) and Lee et al. (2013) and Hájek-type projection (van der Vaart, 2000, Chap. 11).

The remainder of the paper is organized as follows. § 2 sets out the setting, the hypotheses, and the test statistics. Section 3 develops the asymptotic theory: the centering, the limit variances, asymptotic normality, size and consistency, the extension to estimated nuisance parameters (§ 3.4), inference on the magnitudes (§ 3.5) and the implementation (§ 3.6). Section 4 reports Monte Carlo evidence on size, power, and attribution, and on the coverage of the magnitude intervals. § 5 applies the tests to the effect of 401(k) eligibility on net financial assets, and § 6 concludes. The proofs of the asymptotic results, the results for estimated nuisance parameters, and the magnitude results appear in Sections S2, S3, and § 13 of the Online Appendix. The Online Appendix contains auxiliary lemmas, further simulation designs and implementation details; its sections, equations, results, tables and figures are numbered with the prefix S.

Setting, hypotheses and statistics

Setup and estimators

The data are Wi=(Yi,Ti,X1i,X2i,Vi)W_i = (Y_i, T_i, X_{1i}, X_{2i}, V_i), i≤N<∞i \le N < \infty, with a binary treatment Ti∈{0,1}T_i \in\{0,1\}, an real-valued outcome Yi∈Y⊂RY_i \in\mathcal{Y} \subset\mathbb{R}, covariates Xi=(X1i,X2i)∈X1×X2⊂Rd1×Rd2X_i = (X_{1i}, X_{2i}) \in\mathcal{X}_1 \times\mathcal{X}_2 \subset\mathbb{R}^{d_1} \times\mathbb{R}^{d_2} along which heterogeneity is assessed, partitioned by the analyst into two continuously distributed blocks that need not be independent, and further covariates ViV_i needed for identification, possibly none. Let d:=d1+d2d := d_1 + d_2, and suppose X1,X2\mathcal{X}_1, \mathcal{X}_2 are compact with nonempty interior.

With Y(t)Y(t) the potential outcome under treatment tt, the object of interest is

τ(x1,x2):=E[Y(1)−Y(0)∣X1=x1,X2=x2],(x1,x2)∈X1×X2.(1)\tau(x_1,x_2) := E[Y(1) - Y(0) \mid X_1 = x_1, X_2 = x_2], \qquad(x_1,x_2) \in\mathcal{X}_1 \times\mathcal{X}_2. \tag*{(1)}

Assumption 2.1 (Sampling). {Wi}i=1N\{W_i\}_{i=1}^{N} are independent and identically distributed draws from the law of WW.

Assumption 2.2 (SUTVA and unconfoundedness). (a) Y=TY(1)+(1−T)Y(0)Y = TY(1) + (1 - T)Y(0); (b) (Y(1),Y(0))(Y(1), Y(0)) is independent of TT conditionally on (X1,X2,V)(X_1, X_2, V).

Assumption 2.3 (Overlap). (X1,X2)(X_1, X_2) has a Lebesgue density ff, and for some p∈(0,1/2)p \in(0, 1/2), rt(x,v):=Pr⁡(T=t∣X=x,V=v)≥pr_t(x,v) := \Pr(T = t \mid X = x, V = v) \ge p, t∈{0,1}t \in\{0,1\}, and f(x)≥pf(x) \ge p for all xx in a neighborhood of X1×X2\mathcal{X}_1 \times\mathcal{X}_2.

Imposing Assumption 2.3 on a neighborhood of the evaluation region avoids boundary bias and the ensuing corrections. With μt(x,v):=E[Y∣T=t,X=x,V=v]\mu_t(x,v) := E[Y \mid T = t, X = x, V = v], the inverse probability weighted (IPW) and doubly robust (DR) pseudo-outcomes are

ζIPW:=[1(T=1)r1−1(T=0)r0]Y,ζDR:=μ1−μ0+1(T=1)(Y−μ1)r1−1(T=0)(Y−μ0)r0,(2)\zeta^{\mathrm{IPW}} := \left[\frac{1(T = 1)}{r_1} - \frac{1(T = 0)}{r_0}\right]Y,\qquad \zeta^{\mathrm{DR}} := \mu_1 - \mu_0 + \frac{1(T = 1)(Y - \mu_1)}{r_1} - \frac{1(T = 0)(Y - \mu_0)}{r_0}, \tag*{(2)}

with rt,μtr_t,\mu_t evaluated at (X,V)(X,V) and ζ\zeta denotes either.

Proposition 2.1 (Identification). Under Assumptions 2.1–2.3, τ(x)=E[ζ∣X=x]\tau(x) = E[\zeta\mid X = x] for every x∈X1×X2x \in\mathcal{X}_1 \times\mathcal{X}_2, and without VV, τ(x)=E[Y∣T=1,X=x]−E[Y∣T=0,X=x]\tau(x) = E[Y \mid T = 1, X = x] - E[Y \mid T = 0, X = x].

The proof is in Section S1 of the Online Appendix. Because the CATE conditional on XX averages over VV, variation in this CATE implies that the CATE conditional on (X,V)(X,V) is not constant, but does not establish which covariates drive that variation. Conversely, a constant CATE conditional on XX does not rule out heterogeneity conditional on (X,V)(X,V), since averaging over VV may conceal it.

References

  1. [1]Dai, M., W. Shen, and H. S. Stern (2023). Nonparametric tests for treatment effect heterogeneity in observational studies. Canadian Journal of Statistics 51, 531–58.arxiv.org/abs/2103.15023
  2. [2]de Boer, C. (2001). A Practical Guide to Splines (Revised ed.). New York: Springer.
  3. [3]Ding, P., A. Feller, and L. Miratrix (2016). Randomization inference for treatment effect variation. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 78, 655–71.arxiv.org/abs/1412.5000
  4. [4]Ding, P., A. Feller, and L. Miratrix (2019). Decomposing treatment effect variation. Journal of the American Statistical Association 114, 304–17.
  5. [5]Efron, B. and C. Stein (1981). The jackknife estimate of variance. Annals of Statistics 9, 586–96.DOI
  6. [6]Gelber, A. M. (2011). How do 401(k)s affect saving? Evidence from changes in 401(k) eligibility. American Economic Journal: Economic Policy 3, 103–22.
  7. [7]Gunter, L., J. Zhu, and S. A. Murphy (2011). Variable selection for qualitative interactions. Statistical Methodology 8, 42–55.
  8. [8]Hall, P. and N. E. Heckman (2000). Testing for monotonicity of a regression mean by calibrating for linear functions. Annals of Statistics 28, 20–39.DOI
  9. [9]Hochberg, Y. (1988). A sharper Bonferroni procedure for multiple tests of significance. Biometrika 75, 800–802.DOI
  10. [10]Hsu, Y.-C., M. Huber, Y.-Y. Lee, and C.-A. Liu (2026). Testing monotonicity of mean potential outcomes in a continuous treatment with high-dimensional data. Review of Economics and Statistics 108, 792–806.
  11. [11]Hsu, Y.-C. and S. Shen (2021). Testing monotonicity of conditional treatment effects under regression discontinuity designs. Journal of Applied Econometrics 36, 346–66.DOI
  12. [12]Lee, S., O. Linton, and Y.-J. Whang (2009). Testing for stochastic monotonicity. Econometrica 77, 585–602.
  13. [13]Lee, S. and A. M. Shaikh (2014). Multiple testing and heterogeneous treatment effects: re-evaluating the effect of progresa on school enrollment. Journal of Applied Econometrics 29, 612–26.
  14. [14]Lee, S., K. Song, and Y.-J. Whang (2018). Testing for a general class of functional inequalities. Econometric Theory 34, 1018–64.
  15. [15]Poterba, J. M., S. F. Venti, and D. A. Wise (1994). 401(k) plans and tax-deferred saving. In D. A. Wise (Ed.), Studies in the Economics of Aging, pp. 105–42. Chicago: University of Chicago Press.
  16. [16]Poterba, J. M., S. F. Venti, and D. A. Wise (1996). How retirement saving programs increase saving. Journal of Economic Perspectives 10(4), 91–112.
  17. [17]Reynaud-Bouret, P. (2003). Adaptive estimation of the intensity of inhomogeneous Poisson processes via concentration inequalities. Probability Theory and Related Fields 126, 103–53.
  18. [18]Romano, J. P., A. M. Shaikh, and M. Wolf (2010). Hypothesis testing in econometrics. Annual Review of Economics 2, 75–104.
  19. [19]Romano, J. P. and M. Wolf (2005). Stepwise multiple testing as formalized data snooping. Econometrica 73, 1237–82.
  20. [20]Sant’Anna, P. H. (2021). Nonparametric tests for treatment effect heterogeneity with duration outcomes. Journal of Business & Economic Statistics 39, 816–32.
  21. [21]Sarkar, S. K. (1998). Some probability inequalities for ordered MTP₂ random variables: a proof of the Simes conjecture. Annals of Statistics 26, 494–504.DOI
  22. [22]Shah, R. D. and J. Peters (2020). The hardness of conditional independence testing and the generalised covariance measure. Annals of Statistics 48, 1514–38.
  23. [23]Shergin, V. V. (1990). The central limit theorem for finitely dependent random variables. In B. Grigelionis, Y. V. Prohorov, V. V. Sazonov, and V. Statulevičius (Eds.), Probability Theory and Mathematical Statistics: Proceedings of the Fifth Vilnius Conference, Volume II, pp. 424–31. Utrecht and Vilnius: VSP and Mokslas.DOI
  24. [24]Sperlich, S., D. Tjøstheim, and L. Yang (2002). Nonparametric estimation and testing of interaction in additive models. Econometric Theory 18, 197–251.DOI
  25. [25]Sweeting, T. J. (1977). Speeds of convergence for the multidimensional central limit theorem. Annals of Probability 5, 28–41.DOI
  26. [26]Zimmert, M. and M. Lechner (2019). Nonparametric estimation of causal heterogeneity under high-dimensional confounding. Working paper, arXiv:1908.08779.arxiv.org/abs/1908.08779

Paper details

vlnc:W_hqbvff8a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Claude Opus 5.5, GPT-6 Astra, Grammarly

View statement
Looking for related papers…
Contents