Heterogeneous Treatment Effects Analysis through

Distribution Regression based Changes-in-Changes

Matthew Hong∗

Nov. 5, 2024

Click Here for the Newest Version

Abstract

The changes-in-changes method, developed by Athey and Imbens (2006), is a powerful tool
for identifying the distributional effects of a policy intervention, allowing for endogenous
treatment assignment and full counterfactual distribution identification. However, challenges
with incorporating control variables to address concerns akin to differential parallel trends
in the difference-in-differences literature persist. In this paper, I propose a semiparametric
approach to changes-in-changes based on distribution regression that can flexibly account
for observed confounders. This approach can be applied to continuous and/or discrete out-
come variables. I derive functional central limit theorems for the distribution regression based
changes-in-changes estimator and for functionals thereof. These include unconditional distri-
butional and quantile treatment effects. Bootstrap validity result is also provided for conduct-
ing inference in practice. Lastly, I apply the approach to study the heterogeneous effects of
Earned Income Tax Credit on infant weights and find that the policy had higher concentrated
benefits for lower birth weights and more muted effects across the birth weight distribution
than previously reported.

∗
Email: hongmatt@[Link], Ph.D. Candidate, Department of Economics, Boston University.
I am grateful to my advisors, Ivan Fernandez-Val, Hiroaki Kaido and Jean-Jacques Forneron, for their continued
guidance and support. I would like to thank Zhongjun Qu, Kevin Lang, Shakeeb Khan, Stella Hong, Zhanyuan Tian,
and the Econometrics seminar at Boston University for their useful comments. All errors and omissions are my own.
1 Introduction
Much interest in economics research is directed towards quantifying heterogeneous impacts of
regulatory policies on outcomes of interest. As social and economic welfare policies often aim to
improve the outcomes for those at the lower end of the income distribution (e.g. minimum wage
policies) or for those with poor socioeconomic or health outcomes (e.g. food stamps, tax sub-
sidies), correctly identifying and estimating distributional causal effects is crucial for informing
future policy designs and decisions. For various regulatory policies, however, since randomized
control trials are often infeasible, the use of quasi-experimental methods such as difference-in-
differences (“DID”) have become prevalent, and several extensions of the DID method to the dis-
tributional setting have been proposed (Athey and Imbens, 2006; Bonhomme and Sauder, 2011;
Callaway and Li, 2019; Havnes and Mogstad, 2015; Kim and Wooldridge, 2023).
Among the alternatives, CIC stands out with many attractive features. Namely, CIC identifies
the full counterfactual outcome distribution for the treated group in the absence of treatment,
is scale-invariant to monotonic transformations of the outcome variable, and allows for endo-
geneous treatment assignment. Alternative approaches might yield counterfactual distributions
that lie outside the unit interval (especially for discrete outcomes), or might rely on random treat-
ment assignment assumptions which can be less plausible for welfare policies (Ghanem et al.,
2023).
Despite the numerous attractive features of CIC, difficulty in incorporating control variables
has stymied wider application of the method (Lechner, 2011; Melly and Santangelo, 2015). There
are two dimensions to the importance of accounting for potential observed confounders in such
a setting. First, in many economic applications, distribution of observed characteristics maybe
different across groups and the outcomes of individuals with different characteristics such as
age, race, and education levels may evolve differently over time. As such, implementing CIC
without explicitly accounting for potential confounders may misattribute effects due to policy
change. Secondly, said concerns akin to differential parallel trends in the difference-in-differences
literature can be exacerbated in a distributional setting, especially if the effect of controls on the
outcome variable can be heterogeneous across the outcome distribution. In a health economics
setting, for instance, varying levels of basic prenatal care access or quality across U.S. states could

1
differentially impact mothers who would be giving birth to healthier babies compared to those
who would be giving birth to less healthier, lower weight babies.
Existing remedies to account for controls are mainly threefold: (1) Nonparametric methods
which suffer from the curse of dimensionality as the number of support points for the controls in-
crease. (2) Partialling-out approach, where CIC is applied to the residuals from an ordinary least
squares estimation of regressing the outcome on control variables, but one that restricts the het-
erogeneity of controls’ effect on the outcome distribution.1 Lastly, (3) quantile regression-based
approach, which is similar to distribution regression for continuous outcomes, but may provide a
poor approximation for the conditional outcome distribution when the outcome is discrete or has
mass points (Chernozhukov et al., 2013, 2019). These cases arise frequently in relevant empirical
applications such as censoring, bunching or count data.
To overcome these challenges, I propose a distribution regression based approach for CIC.
The approach consists of modeling and estimating the conditional outcome distributions using
distribution regression for each group and time pair, then plugging in the estimates to the condi-
tional version of changes-in-changes. Distribution regression, as developed in Foresi and Peracchi
(1995) and Chernozhukov et al. (2013), can flexibly model the conditional outcome distribution
while allowing for the effect of potential observed confounders to vary across the outcome dis-
tribution. In practice, this ensures that individuals are compared not only with those who have
similar observed characteristics, but also those whose outcomes are affected similarly by the given
characteristics depending on their location in the outcome distribution. Furthermore, the distri-
bution regression can accommodate both continuous and discrete (or mixed continuous-discrete)
outcomes without adjustments to the regression estimator. Computationally, this approach is
feasible and straightforward to implement, as it involves estimating a series of logit or probit
regressions using off-the-shelf software packages over a grid of the outcome support.
Various economically meaningful functionals of the conditional CIC distribution function can
also be recovered. The marginal distribution (or unconditional quantile) function can be obtained
by integrating (and then inverting) the conditional outcome distribution over the empirical distri-
bution of the controls in the treated group. Unconditional quantile treatment effects can be recov-
1
In the empirical section at the end of this paper, I present qualitative and quantitative differences between
residualized CIC and CIC based on distribution regression resulting from a restriction in the heterogeneity. Available
software packages such as qte in R also offer only this option to incorporate covariates.

2
ered by taking the differences in the factual quantile function of the post-treatment outcomes in
the treated group (identified directly from data) and the counterfactual quantile function under no
treatment (identified by CIC). Additional functionals can be identified, including Gini coefficient,
Lorenz curves, average treatment effects and treatment effects on the components of a distribu-
tional decomposition analysis.2 Decomposition methods are useful for explaining outcome gaps
(e.g. income, mortgage decisions, and test score gaps across race or gender) and determining rel-
ative importance of observable components such as education or experience versus unobservable
components (often dubbed “discriminatory” factors). In the quasi-experimental setting, one can
use this to decompose the treatment effects to assess the relative importance of each component
and can provide insight into the mechanism through which outcome gaps change or persist.
This paper derives large sample theory for the distribution regression based CIC process. I
show that the process converges weakly to a tight mean-zero Gaussian process under standard
regularity conditions that ensure functional central limit theorem for the distribution regression
based estimator. I also show that functionals of the estimator are uniformly consistent and asymp-
totically Gaussian using functional delta method and Hadamard differentiability of the relevant
transformation maps. This paper builds on the theoretical results from (Chernozhukov et al., 2010,
2013) which establish functional central limit theorem for distribution regression based estima-
tor of conditional distribution functions, and the Hadamard differentiability of the counterfactual
operator and the monotone rearrangement operator.
I also extend the analysis for non-continuous outcome variables. As mentioned, although the
quantile regression-based CIC (Melly and Santangelo, 2015) and the distribution regression-based
CIC are similar in the continuous outcome case, the approach proposed in this paper can ac-
commodate mixed continuous/discrete outcome variables without adjusting the estimation step
for given group-time pairs, and can thus be applied to a wider range of important applications.
For mixed continuous/discrete outcomes, the CIC-identified counterfactual outcome distribution
maybe partially identified for sub-regions of the outcome support. I propose bounds on the par-
tially identified distribution function and characterize the settings under which the counterfactual
distribution remains point identified. These conditions are verifiable through data directly and
can provide guidance to the applied researcher as to when and where they may obtain point
2
See Fortin et al. (2021) for a detailed review of decomposition methods in economics.

3
estimates versus partially identified bounds.
Lastly, I apply the CIC based on distribution regression on an empirical application revisit-
ing Hoynes et al. (2015) to assess the heterogeneous effect of Earned Income Tax Credit (EITC)
on infant weights. I also present new results comparing the estimates from the more flexible
CIC-DR approach compared to the residualized CIC approach, which is the status quo option
for implementing CIC with control variables. Estimation results suggest that once heterogeneity
of control variables’ impact across the outcome distribution is allowed, EITC had higher concen-
trated benefits for lower birth weights and more muted effects across the birth weight distribution
than previously reported (30g per $1,000 of tax refund spending at the bottom of the birthweight
distribution, which is economically meaningful).

2 Model & Identification

2.1 Notation

Consider the canonical 2 × 2 quasi-experimental setting with data on individual i’s outcomes,
Yi ∈ R, and characteristics, Xi ∈ Rdx where dx denotes the dimension of X. Keeping in line with
the notation from Athey and Imbens (2006), a unit i belongs to either a control group (Gi = 0)
or a treatment group (Gi = 1), and pre- (Ti = 0) or post-treatment period (Ti = 1). Denote
the treatment status as Ii = Gi × Ti . Data on cross-sectional realizations of (Yi , Xi , Gi , Ti ) for
i = 1, . . . , Ngt for each (g, t) ∈ {(0, 0), (0, 1), (1, 0), (1, 1)} pair are observed. Let YiN be the
potential outcome for individual, i, under no treatment, and YiI be the potential outcome under
treatment. Also denote Igt = {i : Gi = g, Ti = t} to be the set of individuals that belong to
group g and time t. Then using the potential outcome framework, the observed outcome in a
given time period is,
Yi = (1 − Ii ) · YiN + Ii · YiI

This notation implicitly assumes that there is no anticipation of the treatment effects, i.e. that
the treatment has no effect prior to its implementation. The subscript, i, will be omitted unless
otherwise required for clarity.
Let Ygt ∼ Y |G = g, T = t be the conditional outcome distribution given the group and time

4
indicator variables, and Ugx ∼ U |G = g, X = x be the conditional distribution of the unobserved
component of the outcome variable. I denote FYgt |X (·|x) as the conditional distribution function
of Y |G = g, T = t, X = x and denote the support of X as X . Define Ygt ⊆ R as the region
of interest, contained in the support of Ygt and let T be the set of quantile indices. For a non-
decreasing function, F : Y → T ⊆ [0, 1], define F −1 : T → Y, where Y is the closure of Y,
as F −1 (u) = inf{y ∈ Y | F (y) ≥ u} with the convention that inf{∅} = sup Y and inf R =
inf Y. For a well-defined cumulative distribution function, F , this generalized (left) inverse is the
quantile function. Lastly, denote independence relation between random variables with ⊥.

2.2 Quantities of Interest: Functionals of CIC Distribution

I first highlight economically meaningful functionals that are identifiable through CIC conditional
on control variables. Once FY11N |X , the conditional distribution of untreated potential outcome for
the treated group in T = 1 is identified through conditional CIC, parameters such as conditional
and unconditional quantile/distributional treatment effects on the treated (UQTT/UDTT), condi-
tional average treatment effects (CATT), and average treatment effects on the treated (ATT) are
identified.
To study differential policy effects across the outcome distribution, quantile treatment effects
are often employed. In policy discussions, moreover, unconditional quantile treatment effects
might be preferred for ease of communication. UQTT can be identified by,

Z −1
δU QT T (τ ) = FYI,−1
11
(τ ) − FYN,−1
11
(τ ) = FYI,−1
11
(τ ) − FYN11 |X (y|x)dFX11 (x)
| {z } | {z } X11
Identified from data Identified from CIC

Note that the quantile function of Y11I is identified directly from the data, whereas the latter quan-
tity in the second equation is identified by inverting the conditional CIC distribution after inte-
grating over the control variable distribution. Relatedly, distributional treatment effects provide
an alternative means to present heterogeneous effects across the outcome distribution.
Z
δDT T (τ ) = FYI11 (y) − FYN11 (y) = FYI11 (y) − FYN11 |X (y|x)dFX11 (x)
X11

Since the entire counterfactual distribution for Y11N is identified, ATT is also identified. ATT

5
can be identified by, Z ∞ h i
δAT T = FY11N (y) − FY11I (y) dy
−∞

which makes use of the relationship between expectation and CDFs through integration by parts,
Z ∞ h i Z 0
E[Y11N ] = 1 − FY11N (y) dy − FY11N (y)dy
0 −∞

For conditional ATT (CATT), one could replace the marginal distribution of Y11I and Y11N with that
of the conditional distribution of Y11I |X and Y11N |X.

Remark 1 Note that this is a different expression than the one proposed in Athey and Imbens (2006)
for the average treatment effect, which makes use of the inverse of the CIC transformation map,
FY−1
01
◦ FY00 (·). Their expression is given by,

δAT T = E Y11I − E FY−1

   
01
(FY00 (Y10 ))

This transformation yields the post-treatment period outcome for an individual with outcome y in the
pre-treatment period (based on outcome changes observed in the control group over time). Although
the two are different ways of expressing the same quantity and thus are not consequential for iden-
tification, the first expression is more amenable to straightforward estimation based on distribution
regression.

Remark 2 The CIC implied ATT and CATT, in general, differ from those implied by standard DID
methods as recognized in Athey and Imbens (2006) and Roth and Sant’Anna (2023). Because the
identifying assumptions of CIC and DDID are non-nested, these can result in different values. Note
that although the expression for E[Y11I ] above is expressed as a functional of FYI11 |X (y|x) for consistent
notation with the counterfactual distribution, it is identified directly from the data. The details around
the difference between estimating the mean of Y11I nonparametrically and semiparametrically is
discussed during estimation in Section 3.

6
2.3 Conditional CIC

I now discuss the CIC framework with control variables along with the identifying assumptions.
The identification for the CIC conditional on controls follows a straightforward adaptation from
Athey and Imbens (2006) and has been considered in Melly and Santangelo (2015). I provide a brief
overview of the model and identification conditions below. For the main results of the current
paper, I focus on the case where Y is continuous and extend the analysis to non-continuous
outcomes in Section 5. Assumption 1 below introduces the model for the potential outcome under
no treatment and the restrictions required to identify the counterfactual outcome distribution for
the treated group in the post-treatment period, FY11N .

Assumption 1 1. (Potential Outcome Under No Treatment): The outcome under no treatment

is given by Y N = h(U, T, X), where U is a continuous scalar random variable and h : U ×
{0, 1} × X 7→ R.

2. (Strict Monotonicity): u 7→ h(u, t, x) is strictly monotonic in u for t = 0, 1 and ∀x ∈ X .

3. (Conditional Time-Invariance): U ⊥ T | G, X

4. (Support Overlap): UG=1,X=x ⊆ UG=0,X=x , ∀x ∈ X11

Assumption 1.1 and 1.3 are key conditions for CIC that play the role of a common trends
assumption in DID. With a production function that is allowed to evolve over time but is com-
mon across groups, unobserved distribution is assumed to be independent of time. This does
not require rank invariance among individuals across time, which is a stronger condition than
the one presented here. Assumption 1.3 is also referred to as the rank similarity condition as in
Chernozhukov and Hansen (2005).
Also note that Assumption 1.3 corresponds to a repeated cross-sectional sampling. When
there is access to panel data, where individual observations are tracked over time, such that
d
YitN = h(Uit , t, xi ) one can adjust the conditional time-invariance assumption to be Ui0 |Gi , Xi =
Ui1 |Gi , Xi . Assumption 1.2, which states that individuals with higher realizations of the unob-
served component, U will also have higher values of Y , enables the comparison of individuals
across groups with similar realizations of the outcome.

7
Lemma 1 Under Assumption 1, the conditional counterfactual distribution for the treated group in
post-treatment period is identified by,

FYN11 |X (y|x) = FY10 |X (FY−1

00 |X
(FY01 |X (y|x))) ∀(y, x) ∈ Y11 X11

The conditional CIC transformation takes the quantile change observed in the control group
between T = 0 and T = 1 for individuals with given characteristics, x, i.e. FY−1
00 |X
◦ FY01 |X (·|x),
and applies this inter-temporal quantile change to the corresponding sub-population (with the
same characteristics, x) in the treated group in T = 0. This is the counterfactual change in quan-
tiles that the treated group would have observed had the treatment not been implemented. The
composition mapping described above for the control group has an optimal transport interpre-
tation and have been explored in related contexts (Arkhangelsky et al., 2019; Gunsilius, 2023).
If DID takes arithmetic differences in control group’s mean outcome over time, CIC takes the
Wasserstein distance between the control group’s outcome distribution over time and applies the
distance to the treated group.
Note that the CIC-identified untreated potential outcome distribution is monotonic by con-
struction and is invariant to monotonic transformations of the outcome variable. Because the
CIC transformation mapping is a composition of monotonic transformations, the resulting coun-
terfactual distribution is a proper cumulative distribution function (CDF). DDID, in contrast, can
yield a counterfactual distribution that lie outside of the unit interval, or one that is not mono-
tonic in y (Athey and Imbens, 2006; Ghanem et al., 2023) depending on the application.3 Because
the identifying assumptions between CIC and DDID are non-nested, these will in general result
in different counterfactual distributions.

2.4 Decomposition of Heterogeneous Causal Effects

In conducting comparative sub-group analysis, there is an extensive literature on decomposition

methods both for mean and distributional outcomes (Fortin et al., 2011). These methods parti-
tion the outcome gap into portions explainable by observed compositional differences, such as
3
In fact, Ghanem et al. (2023) illustrates an example of DDID-identified counterfactual distribution violating the
logical monotonicity in their empirical application based on Cengiz et al. (2019).

8
differences in distribution of education or experience and those not explainable by observable dif-
ferences (often termed “discriminatory” or wage structure effects) in the spirit of Oaxaca (1973)
and Blinder (1973).
In the current setting, one can decompose the policy’s causal effect into each of these terms
to measure the extent of change in the relative importance in explaining the causal effect of the
policy. This could be of particular interest to a social planner that has in mind not only the
change in outcome distribution as a result of a regulatory policy, but also the channel through
which this occurred. For instance, the over-representation of black and/or female workers in low
and medium wage industry has been recognized as a justification for increasing minimum wage
to reduce wage disparities (Derenoncourt and Montialoux, 2021; Wursten and Reich, 2023; Blau
et al., 2023). If minimum wage increases induce a reduction in racial or gender wage gaps, this
maybe viewed as a positive outcome. If outcomes were improved through observable factors such
as education or experience (for which detailed decomposition is available in the literature), then
this decomposition of treatment effects can provide guidance for policymakers to look to invest
in more job training or schooling. On the other hand, if the change in the unobserved component
explained more of the treatment effects, then this could suggest a different policy or investment
strategy moving forward.
To formalize ideas, consider the counterfactual operator as proposed in Chernozhukov et al.
(2013) in a cross-sectional setting with two sub-groups. The decomposition for outcome distribu-
tion gaps between, say white and black workers can be expressed as,

FY <W |W > − FY <B|B> = (FY <W |W > − FY <W |B> ) + (FY <W |B> − FY <B|B> )
| {z } | {z }
△comp △returns

where Z
FY <W |W > (y) = FYW |XW (y|x)dFXW (x)
XW

is the stochastic assignment of outcomes to individuals with characteristics, x for white workers
(analogously defined for black workers), and
Z
FY <W |B> (y) = FYW |XW (y|x)dFXB (x)
XB

9
is the counterfactual outcome distribution for white workers had they possessed characteristics
observed in black workers. This is well defined under a support overlap condition, XB ⊆ XW .
Next, consider the following treatment effect on the outcome distribution gap e.g. wage gap,

(FYI <W,11|W,11> − FYI <B,11|B,11> ) − (FYN<W,11|W,11> − FYN<B,11|B,11> )

| {z } | {z }
Factual Wage Gap Counterfactual Wage Gap

which can also be interpreted as difference in sub-group specific treatment effects. Then, adapting
the notation to the current 2 × 2 setting, (g, t)-specific counterfactual outcome distribution for
sub-groups of interest, j and k, can be expressed as,
Z
FY <j,gt|k,gt> (y) = FYj,gt |Xj,gt (y|x)dFXk ,gt (x)

Decomposing each factual and counterfactual wage gap as above and rearranging terms,

= {(FYI <W,11|W,11> − FYI <W,11|B,11> ) − (FYN<W,11|W,11> − FYN<W,11|B,11> )}

+ {(FYI <W,11|B,11> − FYI <B,11|B,11> ) − (FYN<W,11|B,11> − FYN<B,11|B,11> )}

= (△Icomp − △N I N
comp ) + (△returns − △returns )
| {z } | {z }
δDT E,comp δDT E,returns

one can see that the treatment effect on the outcome distribution gap can be decomposed into
the change in components attributable to compositional differences and change in components
attributable to unobserved differences.

3 Estimation
A plug-in estimation approach based on distribution regression is proposed in this section. Distri-
bution regression, as previously introduced, is a flexible and computationally feasible approach
to modeling conditional outcome distributions given control variables. It is a semiparametric4
model where
FY |X (y|x) = Λ(P (x)′ β(y)) ∀x ∈ X and y ∈ Y (1)
4
β(y) is a function over y ∈ Y.

10
for some link function Λ(·) e.g. logit or probit and y 7→ β(y) is a function-valued parameter.
FY |X (y|x) can be approximated arbitrarily well for rich enough transformations, P (X), and it
coincides with the empirical cumulative distribution function of Y if P (x) is set as a constant.
Furthermore, one can add more structure on the coefficient function, β(y) depending on the
empirical application.5
I apply equation 1 in the quasi-experimental 2 × 2 setting to model the (g, t)-specific condi-
tional distribution functions under the following assumption. Although for ease of exposition, x
enters linearly inside the link function, more flexible specifications such as interaction terms or
higher order terms can be used under P (x). The following modeling assumption is the semipara-
metric structure that I introduce to be able to flexibly incorporate control variables.

Assumption 2

1. Samples {(Ygt,i , Xgt,i ) : i = 1, . . . , ngt } are i.i.d. copies of (Ygt , Xgt ) for all (g, t) pairs that
has probability law, Pgt .

2. The conditional distribution function takes the form,

FYgt |X (y|x) = Λ(x′ βgt (y))

for all (g, t, y, x) ∈ {0, 1}2 × Y × X where Λ(·) is a logit/probit link function.

The second part of Assumption 2 allows for the effect of controls to vary across the outcome
distribution as well as across groups and time periods. In the CIC framework, the unconditional
time-invariance assumption in Athey and Imbens (2006) is made more plausible by considering
the distribution of unobservables, net of the effects by potential observed confounders, to be
independent of time.

3.1 Estimation Algorithm

The estimation procedure consists of estimating the conditional outcome distribution for every
group and time period, then applying the CIC transformation conditional on each control variable
5
For instance, Fortin et al. (2021) uses linear-in-y restrictions to identify and estimate spillover effects of mini-
mum wage.

11
value. Once the conditional estimates have been obtained, one can further recover unconditional
estimates by weighted averages. Given Assumption 2, the estimation proceeds in the following
steps:

1. Specify a grid over the outcome support (e.g. 100 equidistant points across in the support
over which the distribution function is evaluated; these could be the quantiles of the out-
come variable or equidistant grid between the min and the max of the a sub-interval of the
outcome support) that the researcher is interested in estimating, S ⊆ Y01 .

2. For each y ∈ S, estimate conditional distribution functions for each (g, t) pair through
distribution regression using conditional Maximum Likelihood,

F̂Ygt |X (y|x) = Λ(x′ β̂gt (y)), ∀x ∈ X

X
β̂gt (y) ∈ arg max [1{Yi ≤ y} ln Λ(Xi′ b) + 1{Yi > y} ln (1 − Λ(Xi′ b))]
b∈Rp
i∈Igt

3. Rearrange potentially non-monotonic y 7→ F̂Y00 |X (y|x) using monotonic rearrangement

operator of Chernozhukov et al. (2010) and invert F̂Y00 |X (y|x) to obtain F̃Y−1
00 |X
(τ |x). Ana-
lytically, the inverse of the rearranged distribution function has the following representa-
tion,
Z ∞ n o Z 0 n o
F̃Y−1
00 |X
(τ |x) = 1 F̂Y00 |X (y|x) < τ dy − 1 F̂Y00 |X (y|x) > τ dy
0 −∞

4. Plug in distribution function estimates into CIC transformation to obtain counterfactual

probabilities at y ∈ T ,

  
F̂YN11 |X (y|x) = F̂Y10 |X F̃Y−1
00 |X
F̂ Y01 |X (y|x) ∀x ∈ X

The full expression after plugging in the rearranged distribution function is given by,
 Z ∞ Z 0 
′ ′ ′ ′ ′
F̂YN11 |X (y|x) = Λ x β̂10 1{x β̂00 (ỹ) < x β̂01 (y)}dỹ − 1{x β̂00 (ỹ) > x β̂01 (y)}dỹ
0 −∞

12
 5. For the unconditional distribution function, integrate over the empirical distribution of
controls in the treated group.

n11
1 X
F̂Y11N (y) = F̂ N (y|Xi )
n11 i=1 Y11 |X

Computationally, Step 2 consists of estimating a series of logit or probit regression using the
same set of controls over the grid of the outcome support. Although the computational burden
increases with the size of S, this process is straightforward and feasible, since it can be imple-
mented in parallel across the grid. It is also a flexible method compared to the existing partialling
out method that restrict treatment effect heterogeneity across different realizations of X.
Step 3 concerns the monotonic rearrangement of the estimated conditional distribution func-
tion. This maybe required in practice, because in finite samples, even under correct specification
of the distribution function, an estimated conditional distribution function may not be monotonic
in y for some x ∈ X in some regions of T . Computationally, this amounts to sorting the initially
predicted values of the conditional distribution function to obey logical monotonicity restrictions
of a distribution function. Note that this is an estimation adjustment and not an identification
restriction, unlike the non-monotonicity that may arise in the DDID framework due to the way
in which counterfactual distributions are identified and constructed. In this case, one can ap-
ply the monotone rearrangement procedure proposed in Chernozhukov et al. (2010) to recover a
monotonic distribution function. Under Assumption 2, the first-order asymptotic behavior of the
rearranged distribution function is equivalent to that of the estimated conditional CiC distribu-
tion function. Therefore, the large sample theory pertaining to the distribution regression based
CIC estimator will hold for the rearranged distribution function as well.
Once the counterfactual conditional distribution function has been estimated, functionals of
this counterfactual distribution function can also be estimated using a plug-in approach. Uncon-
ditional quantile treatment effects, for instance, can be obtained by integrate over the empirical

13
distribution of controls in the treated group and invert the unconditional distribution functions.

δ̂U QT E (τ ) = F̂YI,−1
11 |X
(τ ) − F̂YN,−1
11 |X
(τ )
n11
!−1 n11
!−1
1 X 1 X
= F̂ I (y|Xi ) − F̂ N (y|Xi )
n11 i=1 Y11 |X n11 i=1 Y11 |X

Note that the unconditional distribution function of Y11I can be estimated in two ways. Because
Y11I is observed directly in the data, one can take the empirical CDF or could use distribution
regression in the same way as rest of the (g, t) pairs and then aggregate it over the distribution
of the control variables.

Remark 3 The two approaches are equivalent if Λ is equal to the logit link function.6 Under
other link functions, the two might not be equal, but the resulting estimates of the conditional
distribution function can be smoother than the nonparametric CDF in many applications. This
can also be achieved if parametric restriction on the β(y) function itself is imposed, for instance
to be linear in y. In this case, there would be a robustness-precision tradeoff.

4 Asymptotic Theory
This section derives the asymptotic distribution of the distribution regression based CIC estimator
(henceforth “CIC-DR”) and shows that the functionals of the counterfactual distribution function
are asymptotically Gaussian. The results are derived under the standard regularity conditions
from Chernozhukov et al. (2013) that ensure the functional central limit theorem (FCLT) holds
for the DR-based (g, t)-conditional distribution estimators, F̂Ygt |X for all (g, t) pairs. Using the
FCLT of the plug-in ingredients along with the Hadamard differentiability of the CIC transfor-
mation map, ϕ(F1 , F2 , F3 ) = F1 ◦ F2−1 ◦ F3 , one can apply the functional delta method to derive
the asymptotic Gaussianity of the CIC-DR process. Furthermore, the asymptotic distribution of
smooth functionals such as unconditional quantile and distributional treatment effects can ob-
tained by using the functional delta method based on the FCLT of the CIC-DR process and the
Hadamard differentiability of the counterfactual operator from Chernozhukov et al. (2013), which
6
This can be verified by examining the first order condition of the log-likelihood function and using the fact that
λ(z) = Λ(z)(1 − Λ(z)).

14
is used for integrating over the distribution of the control together with the chain rule. Follow-
ing the derivation of the asymptotic distribution for the CIC-DR and the unconditional effects, I
also establish the validity of exchangeable bootstrap for the CIC-DR process, which include com-
monly used bootstrap schemes such as the empirical, weighted and m out of n bootstrap, using
the functional delta method for the bootstrap.

4.1 Asymptotic Distribution

To begin, I introduce necessary definitions for presentation of the main results and regularity
conditions that ensure that the DR-based conditional distribution functions obey the FCLT. Let
ℓ∞ (Ygt X ) be the set of bounded and measurable functions, g : Ygt × X → R for g, t, ∈ {0, 1}.

Assumption 3

1. (DR-i) For every (g, t) pair, Ygt is a compact interval in R and X is a compact subset of Rdx .
The conditional density function, fYgt |X (y|x) exists, is positive and uniformly bounded, and is
uniformly continuous in (y, x) on Ygt X .

2. (DR-ii) E [||X||2 ] < ∞ and the minimum eigenvalue of the Jacobian,

λ(X ′ βgt (y))2

 
Jgt (y) := E XX ′
Λ(X ′ βgt (y)) (1 − Λ(X ′ βgt (y))

is bounded away from zero uniformly in y ∈ Ygt for every (g, t) pair, where λ is the derivative
of Λ.

Under Assumptions 2 and 3, Corollary 5.4 of Chernozhukov et al. (2013) implies that as Ngt → ∞,
the following holds for the DR-based conditional distribution function, the conditional quantile
function and the inverse of the monotone rearranged conditional distribution function.

 
Ngt F̂Ygt |X (y|x) − FYgt |X (y|x) ⇝ Zgt (y, x) in ℓ∞ (Ygt X ) (2)
p
  −1
Ngt F̂Y−1 −1
⇝ ZFgt (u, x) in ℓ∞ (Ugt X ) (3)
p
gt |X
(u|x) − F Ygt |X (u|x)
  −1
−1,R −1,R
Ngt F̂Ygt |X (u|x) − FYgt |X (u|x) ⇝ ZFgt (u, x) in ℓ∞ (Ugt X ) (4)
p

15
where Zgt (y, x) = Ggt (ℓgt,y,x ) is a tight mean-zero Gaussian process indexed by (y, x) ∈ Ygt X
with the covariance function7 given by,

E [Ggt (ℓgt,y,x )Ggt (ℓgt,ỹ,x̃ )] = E [ℓgt,y,x ℓgt,ỹ,x̃ ] − E [ℓgt,y,x ] E [ℓgt,ỹ,x̃ ]

with,

Λ(X ′ βgt (y)) − 1{Ygt ≤ y}

ℓgt,y,x (Ygt , X) := −λ(x′ βgt (y))Jgt
−1
(y)x′ ′ ′
λ(X ′ βgt (y))X
Λ(X βgt (y))(1 − Λ(X βgt (y))

and where ZF,−1

gt (u, x) is a tight mean-zero Gaussian process with the covariance function given

by,
       
E Ggt (ℓ̄gt,u,x )Ggt (ℓ̄gt,ũ,x̃ ) = E ℓ̄gt,u,x ℓ̄gt,ũ,x̃ − E ℓ̄gt,u,x E ℓ̄gt,ũ,x̃

with,
1
ℓ̄gt,u,x (Ygt , X) := − ℓ −1 (Ygt , X)
fYgt |X (FY−1
gt |X
(u|x)) gt,FYgt |X (u|x),x

The result above for the inverse of the monotone rearranged conditional distribution function
demonstrates that the empirical process (3) above is asymptotically equivalent to the process (2).
This is a useful result, since it implies that the monotone rearrangement operator does not affect
the asymptotic distribution of the conditional CIC-DR process under the same conditions that
ensure the FCLT holds for the DR-based conditional distribution and quantile process.
I now take the FCLT results above as inputs for deriving the asymptotic distribution for the
conditional CIC-DR process. This is achieved by applying the functional delta method (Lemma
3.10.4 in Van Der Vaart and Wellner (1996)) based on the Hadamard differentiability of the CIC
transformation map,
ϕ(F1 , F2 , F3 ) = F1 ◦ F2−1 ◦ F3

The explicit expression for the derivative map of ϕ along with the relevant function space is con-
7
In the case of Λ being the logit link function, the expression for the covariance function is given by,
−1
λ(x′ βgt (y))λ(x̃′ βgt (ỹ))x′ Jgt (y)
−1
× E [(Λ(X ′ βgt (y ∧ ỹ)) − Λ(X ′ βgt (y))Λ(X ′ βgt (ỹ))) XX ′ ] Jgt (ỹ)x̃

and when Λ = Φ for probit link, a comparable expression can be derived with slight adjustments.

16
tained in Appendix A. The following theorem presents the formal result for the CIC-DR process.

Theorem 1 Suppose that Assumptions 1, 2 and 3 hold. Then as Ngt → ∞ and as N → ∞, where
N = g,t Ngt such that N/Ngt → sgt ∈ [0, ∞) for all (g, t) ∈ {0, 1}2 for g, t, = 0, 1, the following
P

result holds in ℓ∞ (Y11 X ):

p  
N11 F̂Y11I |X (y|x) − FY11I |X (y|x) ⇝ ZI11 (y, x)
√  
N F̂Y11N |X (y|x) − FY11N |X (y|x) ⇝ ZN11 (y, x)

11 (y, x) are independent tight mean-zero Gaussian processes indexed by (y, x) ∈

where ZI11 (y, x) and ZN
Y11 X . The covariance function for ZI11 (y, x) is given by E [ℓ11,y,x ℓ11,ỹ,x̃ ] and the expression for

11 (y, x) can be found in Appendix A.

ZN

Theorem 1 states that the centered and re-scaled DR-based conditional CIC process is asymp-
totically Gaussian. With this in hand, the asymptotic properties of smooth functionals of the
conditional CIC-DR process can also be derived. Before proceeding onto the more general result
on the asymptotic Gaussianity of smooth functionals of the CIC-DR process, I showcase results
on two treatment effects that researchers commonly study. The first functional to consider is the
UQTT. The asymptotic distribution can be obtained by applying the functional delta method in
combination with Hadamard differentiability of the counterfactual operator as defined in Cher-
nozhukov et al. (2013) as, Z
FYgt |X (y|x)dFXg′ t′ (x)
Xg′ t′

where possibly g ′ ̸= g and t′ ̸= t. In the case of marginal distribution functions or quantile

functions, one integrates over FX11 , where g = g ′ = t = t′ = 1. In the case of decompositional
treatment effects, however, one could consider integrating over the FXg′ ,t′ that belongs to another
sub-group of interest, under the appropriate common support condition as mentioned in Section
2. This leads to the following corollaries.

Corollary 1.1 Suppose that conditions for Theorem 1 hold. Then, as N → ∞, the UQTT and UDTT

17
processes converge weakly to the following limit processes:

√  
N δ̂U QT T (τ ) − δU QT T (τ ) ⇝ ZU QT T (τ ) in ℓ∞ (T )
√  
N δ̂U DT T (y) − δU DT T (y) ⇝ ZU DT T (y) in ℓ∞ (Y11 )

where δU QT T (τ ) = FY−1I (τ ) − FY−1N (τ ) and δU DT T (y) = FY11I (y) − FY11N (y). ZU11QT T (τ ) and
11 11

ZU11DT T (y) are tight mean-zero Gaussian processes indexed by τ ∈ T and y ∈ Y11 , with covariance
functions defined in Appendix A.

Next, I present a more general result for smooth functionals of the potential outcome distri-
butions such as Lorenz curves and Gini coefficients.

Theorem 2 Suppose that conditions for Theorem 1 hold and that the map indexed by z given
 
by, ϕ FY11 |X , FY11 |X , FX (z) is Hadamard differentiable with derivative maps, ϕ′F N ,FX and
N I
Y11 |X

ϕ′F I ,FX . Then the following holds.

11 |X
Y

√      
N ϕ F̂Y11I |X , F̂Y11N |X , F̂X (z) − ϕ FY11I |X , FY11N |X , FX (z)
   
′ ′
⇝ ϕF I ,FX ZFY I (z) + ϕF N ,FX ZFY N (z)
Y11 |X 11 Y11 |X 11

Here, ϕ′FY ,FX can be defined as a composition of Hadamard differentiable maps depending
gt |X

on the quantity of interest. For example, if unconditional distribution or quantile functions are
involved, the derivative map can be a composition of the outer map and of the counterfactual
operator (see proof of Corollary 1.1 as an example). If conditional effects are of interest, one can
set the map so that the conditional distribution function is left not integrated over the control
variable distribution.
Because the above results hold uniformly across indexes considered as an empirical process,
this implies that asymptotic normality holds pointwise for a given point in the index set. As
noted in Melly and Santangelo (2015), the asymptotic analysis in this paper is done uniformly for
the entire conditional distribution and quantile process based on distribution regression, thereby
nesting the pointwise case of Athey and Imbens (2006) as a special case where X is a constant.

18
4.2 Inference: Bootstrap Validity

In practice, the asymptotic variance function of the limiting Gaussian process maybe difficult to
estimate analytically. Therefore, this section establishes conditions under which bootstrap va-
lidity holds and suggests an asymptotically valid bootstrap inference procedure. As have been
considered in related contexts (Chernozhukov et al., 2013; Melly and Santangelo, 2015; Wüthrich,
2019), this paper proposes exchangeable bootstrap, which includes bootstrap schemes such as the
empirical bootstrap, weighted bootstrap and m-out-of-n bootstrap as special cases. The weighted
bootstrap, for example, is useful when the sample size is small and control variables are categor-
ical (which can cause “small cell” problems) since no observation is discarded in the resampling
process. When the sample is large, one can utilize subsampling for computational tractability.
The asymptotic distribution can be consistently estimated using exchangeable bootstrap by func-
tional delta method for bootstrap.
The following assumption around bootstrap weights establishes conditions under which boot-
strap validity is achieved (appears under Condition EB in Chernozhukov et al. (2013)).

Assumption 4 For each (g, t), (wgt,1 , wgt,2 , . . . , wgt,Ngt ) is an exchangeable, non-negative random
N
vector independent from the data, {(Ygt,i , Xi )}i=1
gt
such that for some ε > 0,

Ngt
1 X
(wgt,i − w̄gt )2 →p 1, and w̄gt →p 1
 2+ε 
sup E wgt,1 < ∞,
Ngt Ngt i=1

PNgt
where w̄gt = Ngt
−1
i=1 wgt,i is independent across (g, t).

For applications with “small cells’, one can use the weighted bootstrap with weights satisfying,
E [wgt,1 ] = V ar (wgt,1 ) = 1 such as the standard exponential distribution.

4.2.1 Bootstrap Algorithm

Any exchangeable bootstrap scheme obeying Assumption 4 will be valid, and a bootstrap proce-
dure for conducting inference for UQTT is provided below.

N
1. For each (g, t)-conditional sample {(Ygt,i , Xi )}i=1
gt
, draw bootstrap weights, (wgt,1 , . . . , wgt,Ngt )
that satisfy Assumption 4, and normalize weights to sum to 1 for every (g, t)-conditional

19
 sample.

2. Estimate the DR-based weighted conditional distribution function based on (wgt,1 , . . . , wgt,Ngt ),

∗
F̂Y∗gt |X (y|x) = Λ(x′ β̂gt (y)), ∀(y, x) ∈ Ygt X , g, t, ∈ {0, 1}
X
∗
β̂gt (y) = arg max wgt,i [1{ygt,i ≤ y} ln Λ(x′i b) + 1{ygt,i > y} ln (1 − Λ(x′i b))]
b
i∈Igt

3. Obtain bootstrap CIC-DR counterfactual distribution,

  
F̂Y∗ N |X (y|x) = F̂Y∗10 |X F̂Y∗,−1,R
00 |X
F̂Y∗01 |X (y|x)
11
 Z ∞ Z 0 
=Λ x′ β̂10
∗
1{x′ β̂00
∗
(ỹ) < x′ β̂01
∗
(y)}dỹ − 1{x′ β̂00
∗
(ỹ) > x′ β̂01
∗
(y)}dỹ
0 −∞

4. Estimate the treatment effect (or a functional of interest) by plugging in the relevant esti-
mates, e.g. for UQTT,

δ̂U∗ QT T (τ ) = F̂Y∗,−1 ∗,−1

I |X (τ ) − F̂Y N |X (τ )
11 11
!−1 !−1
X X
= w11,i F̂Y∗ I |X (y|Xi ) − w11,i F̂Y∗ N |X (y|Xi )
11 11
i∈I11 i∈I11

n o
∗,(b)
5. Repeat steps 1–4 for b = 1, . . . , B times and obtain bootstrap estimates, δ̂U QT T (τ ) .
τ ∈T ,b=1,...,B

6. Construct simultaneous (1 − α)-confidence bands based on maximal t-statistics across τ ∈

T as the end-point functions defined by,

√
CI1−α (τ ) = δ̂U QT T (τ ) ± t̂1−α · Σ̂U QT T (τ )1/2 / N

where t̂1−α a consistent

 estimator of the (1 − α)-th quantile of the Kolmogorov-Smirnov
√

−1/2 (b)
maximal t-statistic, of sup N Σ̂U QT T (τ ) |δ̂U QT T (τ ) − δ̂U QT T (τ )| : b = 1, . . . , B , and
τ ∈T
Σ̂U QT T (τ ) is a uniformly consistent estimator of ΣU QT T (τ ), the asymptotic variance func-
(b)
√  (b) 
tion of ẐU QT T (τ ) := N δ̂U QT T (τ ) − δ̂U QT T (τ ) . For instance, this could be the boot-
strap inter-quartile range re-scaled with the standard normal distribution such as, Σ̂U QT T (τ ) =

20
 n o
(b)
(q0.75 (τ )−q0.25 (τ ))/(z0.75 −z0.25 ), where qp (τ ) is the p-th quantile of ẐU QT T (τ ) : b = 1, . . . , B
and zp is the p-th quantile of N (0, 1).

If the main target of inference is the CIC-DR process itself, then the asymptotic simultane-
ous confidence bands as constructed in Step 6 can be adjusted to be centered around F̂Y11N |X (y|x)
instead and with analogous adjustments to t̂1−α and Σ̂(y). In this paper, I consider uniform infer-
ence methods that cover standard pointwise methods for real-valued parameters as a special case.
Specifically, the asymptotic simultaneous confidence band, CI1−α above, satisfies the coverage
guarantee given by,

n o
lim P δ̂U QT T (τ ) ∈ CI1−α (τ ), ∀τ ∈ T = 1 − α.
N →∞

This also makes it possible for considering inference on functionals of the CIC-DR distribution
function, such as the Lorenz curve, Gini coefficient and decompositional treatment effects. For
large sample sizes, one could consider computationally more tractable means such as subsampling
√  
or multiplier bootstrap that simulate the limit process of N F̂Y11I |X (y|x) − F̂Y11N |X (y|x) , or
the relevant centered and re-scaled empirical process (e.g. functional of the CIC-DR process)
as in Barrett and Donald (2003). Formally, the following theorem formulates the validity of the
exchangeable bootstrap.

Theorem 3 Suppose Assumptions 1, 2, 3 and 4 hold. Then the exchangeable bootstrap consistently
estimates the limit laws for the processes in Theorem 1 and in Corollaries 1.1.

5 Extension: CIC for Mixed Continuous-Discrete Outcomes

In this section, I extend the preceding framework developed for continuous outcomes to mixed
continuous and discrete outcomes where the potential outcome distributions can exhibit mass
points (e.g. censored outcomes or bunching). Censored outcomes, for example, can be highly
relevant in analyzing the distribution of wages or hours worked in minimum wage policy settings.
Such a setting highlights the general applicability of a DR-based approach for CIC. First, DR can
accommodate mixed continuous and discrete outcome variables without special adjustment to
the regression estimator, whereas QR does not. Second, although FY11N maybe partially identified

21
under the CIC framework, the bounds on FY11N will obey natural monotonicity restrictions and
will yield predictions within the closed unit interval. This is in contrast to DDID which can yield
counterfactual probabilities outside the unit interval in the lower or upper tails, precisely where
possibly heterogeneous effects are of interest.

Toy Example: Illustrating with a simple example can be instructive. Consider a setting with
left-censored outcomes, no control variables and censoring points (cG=0 = 2 and cG=1 = 3) that
are constant across T but different between G. This simulates a common setting when analyzing
the effect of a minimum wage increase on the distribution of wages between two neighboring
states. Panel (b) of Figure 1 shows that the counterfactual distribution identified by DDID (in

(a) Control Group (b) Treated Group

Figure 1: Example of censored outcomes at c0 = 2, c1 = 3. Point identified FY11N |X .

orange),
FYDDID
N (y) = FY10 (y) + FY01 (y) − FY00 (y)
11

has negative probabilities for y ∈ [c0 , c1 ], because FY10 (y) = 0 while FY01 (y) < FY00 (y) for y ∈
[c0 , c1 ]. This stands in contrast with the CIC identified FY11N (in blue), which is point identified, is
non-decreasing and has range inside the unit interval. In Figure 2, the censoring point is constant
across G and T . Here, FY11N is partially identified for y ∈ [2, 3) and the bounds for that region
are shown in blue. The main reason for this is due to the non-uniqueness of the inverse function

22
 (a) Control Group (b) Treated Group

Figure 2: Example of censored outcomes at c0 = c1 = 2. Partially identified FY11N |X .

of FY00 |X (·|x) at the censoring points. As will be made explicit in the subsequent section, for
censored outcomes, and more generally outcomes with finite number of discontinuities in the
conditional outcome distribution function, the generalized left and the right inverses no longer
agree with each other as in the continuous case.
Here, the upper bound is derived in the same manner as the point estimate for the counter-
factual distribution in the continuous outcome case, since the generalized left-inverse expression
is the same in the continuous outcome case and in this case. This upper bound is the blue line
corresponding to the higher horizontal line for [2, 3). The lower bound is derived by the expres-
sion to be explicitly stated in the following section, but essentially is taken by the generalized
right-inverse and corresponds to the lower horizontal line in Figure 2 panel (b) and has a jump
in the distribution function at y = 3.
The examples demonstrates that CIC does not necessarily lose point identification in a mixed
continuous-discrete outcome variable setting. As will be provided below, whether or not the
counterfactual distribution is point identified and to what degree it is partially identified depends
on the relationship between the ranges of the conditional distribution functions of Y01 |X, Y00 |X
and Y10 , the support of Y10 .

23
5.1 Identification for Conditional CIC with Mixed Continuous-DIscrete
Outcomes

To formalize concepts, I introduce additional notation. Namely, for a non-decreasing function,

F : Y → T ⊆ [0, 1], define the generalized (right) inverse, F † : T → Y, as F † (u) = sup{y ∈
Y ∪ {−∞} | F (y) ≤ u} maintaining the convention that sup{∅} = inf Y and sup R = sup Y.
For continuous outcomes, the two (left and right) inverses agree, but they disagree in general.
For a function f : R → R, moreover, let f (x−) = supz<x f (z). If f is left-continuous, then
f (x−) = f (x). Lastly, let Ran(F ) = {F (y) : y ∈ R} be the range of the function.
For mixed continuous-discrete outcomes, certain adjustments to the identifying assumption
can be considered. The assumption of strict monotonicity as introduced in Assumption 1 can
be overly restrictive and may not be practically well justified. Therefore, I relax Assumption 1
pertaining to the relationship between U and Y to be weakly monotonic instead.

Assumption 1′ (Weak Monotonicity): u 7→ h(u, t, x) is weakly monotonic in u for t = 0, 1 and

∀x ∈ X . Also, the remaining Assumption 1.1, 1.3, 1.4 hold.

With this modified assumption, I present a modified identification result. The identification of CIC
with mixed continuous-discrete outcomes and no control variables have recently been studied in
Ghanem et al. (2023), rationalized using a copula stability condition on the selection mechanism
rather than a conditional time invariance assumption as in Athey and Imbens (2006)8 . Adapting
their result to my case of conditional CIC with control variables, I restate the result as a lemma
below.

Lemma 2 Suppose Assumption 1′ holds. Then, FY11N |X (·) is partially identified with bounds,

lim sup{F LB (y|x) : y ≤ ỹ & y ∈ Y01 ∪ {−∞}}

ỹ↓y

≤FY11N |X (y|x)

≤ lim sup{F U B (y|x) : y ≤ ỹ & y ∈ Y01 ∪ {−∞}}

ỹ↓y

8
Ghanem et al. (2023) show that the two conditions are equivalent for continuous outcomes, but are non-nested
in general for mixed continuous-discrete or discrete outcome settings.

24
∀(x, y) ∈ Y11 X11 , where



F LB (y|x) := FY10 |X FY†00 |X FY01 |X (y|x) −


F U B (y|x) := FY10 |X FY−1
00 |X
FY01 |X (y|x)

The lim sup expression that wraps around F LB (y|x) and F U B (y|x) are made to ensure that the
bounds are themselves right-continuous since they are intended to cover distribution functions.
These bounds are valid for continuous, discrete and mixed continuous-discrete outcome variables.
Note that for continuous outcomes, the bounds collapse to a single point, specifically the upper
bound. Ghanem et al. (2023) also show that these bounds are sharp when Ran(FY00 ) is closed
and they numerically coincide with those proposed by Athey and Imbens (2006) when outcomes
are purely discrete.
Next, I specialize Lemma 2 to censored outcomes and provide conditions under which the
counterfactual distribution is point identified and where it is partially identified, along with their
corresponding regions. Note that these conditions are directly verifiable through data, providing
a useful check for the applied researcher.

Proposition 1 Suppose Assumption 1′ holds and FYgt |X has finite number of discrete components.
Then, FY11N |X is point identified for the set,

{(y, x) ∈ (Y01 ∩ Y10 ) × X : ∃y ′ ∈ Y00 s.t. FY10 |X (y|x) = FY00 |X (y ′ |x)}

and FY11N |X is partially identified for the set,

{(y, x) ∈ (Y01 ∩ Y10 ) × X : ∄y ′ ∈ Y00 s.t. FY10 |X (y|x) = FY00 |X (y ′ |x)}

and the bounds are given by Lemma 2.

The proposition implies that even if Ran(FY10 |X ) ⊈ Ran(FY00 |X ), which would result in par-
tial identification in the purely discrete outcome case, if the set Γyx := {(y, x) ∈ Y01 × X :
Ran(FY10 |X ) ⊈ Ran(FY00 |X )} is empty, then FY11N |X remains point identified.

Remark 4 The examples in Figures 1 and 2 highlight that based on the conditions regarding the

25
range of the distribution functions and the support of the outcome variables, it is possible for
the researcher to directly determine from the data whether they will be obtaining bounds on the
counterfactual distribution or will have point identification.

Example (Minimum Wage) To solidify intuition and to outline the usefulness of the result
for an applied researcher, consider an example of minimum wage in the 2 × 2 setting again, but
with no control variables. Denote the censoring points as wgt for g, t = 0, 1 where wgt = wg ∀t
for simplicity. Suppose that the control group’s baseline minimum wage is set at w0 and that of
the treated group is set at w1 . If a researcher observes in the control group that the fraction of
minimum wage workers increases over time i.e. FY01 (w0 ) > FY00 (w0 ), then the researcher will
obtain point identified counterfactual distribution, FY11N , for the entire region of interest Y11 .
Suppose, instead, that the researcher observes that the fraction of minimum wage workers
decreases over time in the control group i.e. FY01 (w0 ) < FY00 (w0 ). Then, if minimum wage is
constant across groups, i.e. w1 = w0 as in Figure 2 panel (b), then FY11N is partially identified for
w0 ≤ y ≤ FY−1
00
(FY01 (w0 )) and the counterfactual distribution defined on the set Y11 excluding
the interval [w0 , FY−1
00
(FY01 (w0 ))] will be point identified. If w1 > w0 , then it is less likely for
FY11N to be partially identified in general. In the running example, given that the distribution
function is discontinuous only at the censoring point, the counterfactual distribution will be point
identified, since one will be able to find a wage level, w′ possibly not equal to w1 , such that
FY00 (w′ ) = FY01 (w1 ) holds and the inverse of FY00 at FY01 (w1 ) is uniquely determined.

Generalization: Time-varying or Conditional Censoring Points One could also consider

more general settings, where for example, (1) a control group’s censoring point, c0t is allowed
to change over time or (2) allowed to vary across individual characteristics, cgtx more broadly.
(1) can allow for analyzing minimum wage policy effects even when the available control group
also experiences a minimum wage change during the period. This enables the identification of
the effect of minimum wage change in the treatment group beyond what would have occurred
had they experienced the same minimum wage path as the control group. Because the analysis
presented above holds conditionally on observables, the set-up also accommodates settings where
the censoring point depends on such observable characteristics. Such a feature allows for the

26
possibility that minimum wages can vary across industries as in Derenoncourt and Montialoux
(2021) or that tax schedules may change differently over time for people with some eligibility
status (e.g. disability insurance or having children).

Estimation, Asymptotic Distribution and Inference Estimation consists of a plug-in strat-

egy similar to the one proposed for the continuous outcome case. Because the difference in the
bounds arise from non-uniqueness of inverses given finite number of jumps in the otherwise
continuous distribution function of the conditional outcome variable, the plug-in estimators for
the conditional distribution functions do not require any adjustments. The generalized left- and
right-inverses taken for FY00 |X is the only adjustment required in the implementation of condi-
tional CIC that is necessary. For QR-based CIC, the conditional distribution function estimator
itself would need to be modified in addition to the different definitions of generalized inverses.
Next, I outline how the asymptotic distribution for the plug-in estimates of the bounds based
on DR estimates of the conditional distribution function can be derived. The result can be es-
tablished similar to how FCLT was derived for the conditional CIC estimator in the continu-
ous outcome case. Note that the DR-based estimator of the upper bound is equivalent to the
point identified function in the continuous case, so no adjustment to the DR estimator is required
for continuous and mixed continuous-discrete outcomes. Therefore, only the case for the lower
bound remains to be shown.
For thew lower bound, one can use results from Cárcamo et al. (2020) to show that the right
inverse operator is Hadamard directionally differentiable, a notion that is different from (full)
Hadamard differentiability. Then the asymptotic distribution for the lower bound can be derived
by applying the functional delta method for Hadamard directionally differentiable functionals.
Note, however, that ? implies that the asymptotic distribution of the lower bound for conditional
CIC is no longer a Gaussian process since we do not have full Hadamard differentiability. Because
the asymptotic distribution is even more involved than the Gaussian process for the upper bound,
bootstrap inference based on ? can be implemented.

27
6 Effect of EITC on Birthweights
In this section, I apply the CIC-DR estimator developed in Section 3 to revisit the study of Hoynes
et al. (2015). I re-analyze the effect of Earned Income Tax Credit (“EITC”) on low birthweights
(≤ 2,500g or 5.5 lbs), which have been well recognized in the health economics literature to be
a consequential indicator of short and long-term health and economic outcomes (Almond et al.,
2005; Currie, 2011; Almond et al., 2011). Given the concentrated risk at the low end of the birth
weight (Y ) distribution, estimating heterogeneous effects of a welfare policy or a treatment across
the outcome distribution, while controlling for potential confounders (X) in a flexible manner is
of critical importance.
The importance of accounting for control variables for CIC can be highlighted by two issues
in this empirical context. First, birthweight trends may vary across geographic region that in-
validates the unconditional time-invariance condition as outlined in Athey and Imbens (2006) if
control variables are not accounted for. Socio-economic and health characteristics correlated with
race, age and education levels of mothers may also induce differential trends for birthweights over
time. Second, U.S. state of residence may matter more for mothers who would be giving birth to
low weight infants than healthier mothers, given that access to prenatal care and quality of care,
more generally, may vary geographically. Accordingly, assuming that race, age, education and/or
residence location matter equally across the potential outcome distribution may yield misleading
results. CIC-DR in this regard can provide reliable treatment effects estimates and confidence
bands compared to existing restrictive methods for informing future policy decisions.

6.1 Background & Data

EITC is a welfare policy targeting working low-income individuals in the form of additional
tax refunds. The maximum allowable amount of these credits increases in the number of pre-
existing children for the mother/family. Hoynes et al. (2015) study the impact of the largest EITC
expansion (Omnibus Reconciliation Act of 1993; OBRA93) using a DID design to identify the
impact of income gains (through EITC) on birth weights. They find significant effects of EITC
on reducing incidences of low birth weight and on increasing average birth weights among the
treated. Because OBRA93 generated large tax credit gains for those with two or more children

28
versus for those with one child, they define the control group as those with one pre-existing child
and the treated group as those with two or more children.9
Microdata from U.S. Vital Statistics Natality Data between 1989-1999 is used for analysis. Data
contains infant characteristics, such as birth weight, gender, birth order (parity) and birth date,
and mother’s demographic characteristics, such as age, race, education, state of residence and
marital status. To be in line with the original analysis, the sample is limited to single-mothers with
at least 18 years of age and at most 12 years of education, which yields 2-3 million observations per
(g, t) pair.10 The implementation in this section follows suit with the two identification conditions,
“cash-in-hand” and “sensitive developmental stage” proposed by Hoynes et al. (2015). The first
condition assumes that infant health is affected by the EITC through the cash payment in the
form of tax refunds and that this cash is spent in the subsequent 12 months of receipt. The
second condition assumes that the last three months of pregnancy before birth is important for
infant’s birth weight production. The paper also assumes full take-up of the treatment based
on high take-up evidence from the IRS around the time frame of analysis and I maintain this
assumption.11
Due to potential heterogeneity in treatment effects over time as well as the plausible pre-
trends based on the previous OBRA expansions, I consider treatment effects for pairs of samples in
1993 compared to those in 1994, 1993 compared to those in 1995, etc. I also compare my results to
the partialling out approach (via OLS) as proposed by Athey and Imbens (2006) and implemented
by Sasaki and Wang (2024) for the same empirical study.12 This approach significantly restricts
the heterogeneity in the controls’ effect on the outcome distribution. Note that Sasaki and Wang
(2024) focus on the extreme quantiles of the birthweight distribution, and as such they impose
such parametric restrictions on the effect of covariates on the outcome distribution. I extend their
analysis to the intermediate quantiles to showcase how a researcher would implement CIC with
9
They also consider the comparison between individuals with no child vs. those with one or more children, but
due to pre-existing trends in the gap between no children and one child, I focus on the one child vs. two or more
children case.
10
Sample selection in such a manner focuses the analysis on “high-impact sample” as described in Hoynes et al.
(2015).
11
See Footnote 19 of Hoynes et al. (2015) for details.
12
Sasaki and Wang (2024) considers a 2 × 2 setting by comparing observations between 1989–1993 to those in
1994–1999. The results for the pooled sample as well as the pairwise samples differ qualitatively, but the relative
differences between CIC-DR and the partialling out approach persist throughout the two cases.

29
control variables given available tools in the literature.
The restriction of the control variables’ effect on the outcome distribution can be seen by
letting Ỹgt = Y − X ′ βgt and observing that for a given (g, t) pair,

 
FỸgt |X (y) = P Ỹgt ≤ y|X = x = Λ (y + x′ βgt )

where βgt is assumed to be homogeneous across y ∈ Y and only the intercept is allowed to vary.
Contrasting this with the more general framework used under CIC-DR,

FYgt |X (y|x) = Λ (x′ βgt (y))

it is easy to see that the latter expression is much more flexible and includes the former specifi-
cation as a special case. Indeed one can test H0 : βgt (y) = 0 ∀y ∈ Ygt and I show an example in
Appendix B Figure 6, illustrating that the data rejects the more restrictive effect of control vari-
ables on the outcome distribution. In particular, the null hypothesis for the coefficient function
corresponding to mother’s race being black is rejected at the 5% significance level for low birth
weights, whereas the control variable’s effect is not significant for higher birth weights.

6.2 Estimation Results

Estimation results comparing CIC-DR, residualized CIC and the original results from Hoynes et al.
(2015) are presented in this section. Quantile and distributional treatment effects are presented
first, followed by average treatment effects implied by CIC-DR for comparison with alternative
methods.

Heterogeneous Treatment Effects Quantile treatment effects of EITC across the birth weight
distribution (grid of 50 equidistant points in [0.02, 0.98]-th quantiles) are reported for pairwise
samples (’93 vs. ’94, ’93 vs. ’95, etc.). Figure 3 below shows UQTT estimates of EITC on birth
weights based on CIC-DR and the OLS-residualized CIC based on Athey and Imbens (2006) and
Sasaki and Wang (2024). The uniform confidence bands for CIC-DR (in blue) are based on 500
bootstrap replications using standard exponential weights and the confidence bands for residu-
alized CIC (in red) are based on the expression for the asymptotic variance given by Athey and

30
Imbens (2006).

(a) Pre: 1993, Post: 1994 (b) Pre: 1993, Post: 1995

(c) Pre: 1993, Post: 1996 (d) Pre: 1993, Post: 1997

Figure 3: Pairwise UQTT of EITC on Birth Weight (g). Blue line plots estimates from CIC-DR and red line
plots estimates from OLS-residualized CIC. Blue shaded area is the uniform 90% confidence band obtained
from 500 weighted bootstrap replications and red shaded area is the confidence band obtained based on
analytical expression of asymptotic variance from Athey and Imbens (2006).

As the years pass from 1994 to 1997, and thereby the amount of EITC allowable increases,
the quantile treatment effects demonstrates an increasing pattern for the lower end of the in-
fant weight distribution. Note that in Figure 8 panel (d) for 1997, which is when the maximum
EITC allowable is the highest among the years considered, the effect of EITC on birth weight is
estimated to be more positive (albeit noisy) in the bottom decile than previously reported. This

31
is an approximately 50 g increase in birth weight at the 0.02-th quantile for the 1997 sample.
Also, although CIC-DR and the residualized CIC seems to be similar for 1994, as time passes, the
gap between the two demonstrably increases leading to statistically and economically significant
differences in the resulting estimates. Restricting heterogeneity of X’s impact across the distribu-
tion of birth weights would, therefore, seem to overstate the impact of EITC on the intermediate
and upper birth weight quantiles.

Interpretation of Heterogeneous Treatment Effects To contextualize these heterogeneous

estimates in light of Hoynes et al. (2015), I also present distributional treatment effects on the
treated. The aforementioned paper presents the treatment effects in terms of probability changes
of an infant weighing less than 2,500 g (or 5.5 lbs). In comparison, CIC-DR delivers a complete
picture across the entire distribution of birth weights with uniformly valid confidence bands as
opposed to only pointwise estimates. Because the original paper considered pooled sample letting
1992-1993 effective tax years as the pre-treatment period and 1994-1999 as the post-treatment
period, I also estimate UDTT using the same sample for comparison. Figure 4 shows the changes
in probability of birth weight being below a certain threshold, over a grid of thresholds. Since
the treatment effect is on the cumulative distribution, a negative UDTT can be interpreted as a
positive health outcome implying that more children are born with higher weights compared to
the control group. Similar to the pairwise results, the pooled results suggest that residualized
CIC would overstate the impact of EITC on birth weights. CIC-DR estimates a UDTT of -0.10
percentage points at y =2,500 g. DID results for y =2,500 g from Hoynes et al. (2015) on Table 2
suggests -0.36 percentage points; residualized CIC suggest a UDTT of -0.22 percentage points.
Two points are worth mentioning in interpreting these results. First, the estimation results
suggest that for the pooled sample, EITC had smaller-than-reported effects in reducing the in-
cidence of low birth weight i.e. proportion of those who had babies weight under 2,500 g. In
monetary terms, based on the overall after-tax income calculated in the original paper (Table 4 of
Hoynes et al. (2015)), this translates to 0.06 percentage point reduction in incidences of low birth
weights (in 2009 dollar terms) or a 2.75 g / $1,000 increase in birth weights at the quantile rank
corresponding to 2,500 g threshold. Recall that the pairwise result for the 1997 sample would
suggest a three times as large (6.42 g) gain at the same quantile compared to the pooled sample

32
 (a) UDTT for Pooled Sample (b) UQTT for Pooled Sample

Figure 4: Pooled UDTT and UQTT of EITC on Birth Weight (g). Pre: 1992-1993, Post: 1994-1999. Blue line
plots estimates from CIC-DR and red line plots estimates from OLS-residualized CIC. Shaded areas are the
uniform 90% confidence band obtained from 500 weighted bootstrap replications.

estimate. For the bottom decile, the benefits are much higher at approximately 32 g / $1,000.
Second, I argue that these results are nevertheless economically very meaningful improve-
ments. To see this, comparing EITC’s effect on low birth weight to that of smoking on low birth
weight can be instructive. A conservative estimate by Currie et al. (2009) suggests that the effect
of a mother being a smoker is estimated to reduce birth weight by 38.9 g, with each additional
cigarette reducing birth weight further by 2.2 g. Although the estimated effects of EITC around
the intermediate quantiles suggest 5-10 g of increased birthweight depending on the sample used,
a near 50 g increase in birth weight at the bottom of the weight distribution in 1997 is notable.
This is on par with the effect of being a smoker when translated into $1,000 terms. Furthermore,
as Hoynes et al. (2015) argue, additional cash could induce a negative income effect of smok-
ing behavior and through reduced stress and added employment effects of mothers, which can
further help improve birth outcomes through multiple channels.

CIC-implied Average and Median Treatment Effects Next, I present CIC-DR implied av-
erage (ATT) and median treatment effects (MTT) estimates. Although the main spirit of the
proposed method is to analyze heterogeneous beyond average effects, comparing those obtained

33
via CIC-DR with those found in Hoynes et al. (2015), which uses pointwise DDID13 can be an
instructive exercise. The CIC implied ATT can easily be obtained by using the expression from
Section 2.2. The results presented also include the residualized CIC. Both the CIC-DR implied ATT
and MTT are much smaller than those found by DDID or using residualized CIC. The ATT is no
longer statistically significant at the 5% significance level. Estimates from DDID and residualized

ATT MTT
DDID 9.95 –
(2.05) –
CIC-DR 2.67 3.60
(1.51) (1.46)
CIC-Resid 10.49 10.35
(1.03) (0.94)
DDID-DR 2.46 3.38
(1.47) (1.52)

Table 1: Average and Median Treatment Effects on the Treated. Numbers in parentheses are
bootstrapped standard errors based on 500 replications.

CIC are much closer to each other and are statistically significant at the 5% significance level.
Although the DDID design from Hoynes et al. (2015) control for potential observed confounders,
they do so by aggregating observations by cells induced by the combination of control variable
values. Because they do this for a series of thresholds, this could be considered to be more flexible
than the residualized CIC approach. The difference in the estimates between CIC-DR and DDID
could be attributed to largely two factors: First, a difference in the identifying assumptions and
the resulting estimate of the counterfactual distribution could lead to a divergence in the treat-
ment effects estimates. This could be especially salient if the marginal outcome densities exhibit
skewness or multi-modality and these patterns are different across group and time indicators.
Upon visual inspection (demonstrated in Appendix B), the CDFs and the quantile functions seem
similar in their shapes. The second factor could be attributed to distribution regression as a flex-
ible modeling tool which could be capturing heterogeneity of control variables effect across the
outcome distribution. If this is the case and the first posited factor is less relevant, then a DR-
13
In the sense that they consider a DID design on the probability of being low birthweight but only for several
thresholds, and thus not having simultaneously valid standard errors.

34
based DDID estimator could yield treatment effects estimates closer to those obtained using the
CIC-DR estimator.
Indeed, the results on the table appear to be consistent with the aforementioned narrative.
One can observe the DDID based on DR-estimates of the marginal outcome distributions (where
F̂Ygt |X (y|x) yields similar ATT and median treatment effects as those implied by CIC-DR. Here,
the potential outcome under no treatment in the treated group using DR-based DDID obtained
by,
 
F̂YDDID
N (y) = F̂ DR
Y10 (y) + F̂ DR
Y01 (y) − F̂ DR
Y00 (y)
11

where F̂YDR Λ(x′ β̂gt (y))dF̂X11 (x) is denoted to distinguish the DR-based estimates of
R
gt
(y) = X11

the distribution function. DDID and CIC will not be similar in general, especially if the distribu-
tions across groups are quite different, and DDID may even yield invalid CDFs that violate the
natural monotonicity condition when outcomes are discrete. In this empirical example, given that
the marginal distribution functions are smooth and are quite similar in their shape, the second
aforementioned factor appears to exhibit higher relevance. If the difference in the probability
gaps versus the quantile gaps across group and time indicators are large, this could increase the
relevance of the first factor in explaining the difference in the estimates.
Lastly, I report that I also find multiple violations of the natural monotonicity restriction of
conditional distribution functions for various control variables using DDID in this empirical ex-
ercise. This implies that the data clearly rejects the identifying assumption of DDID. Although
there are statistical tests available to test the violation of monotonicity as outlined in Roth and
Sant’Anna (2023), this could give rise to issues with pre-testing. On the other hand, the counter-
factual conditional distributions produced by CIC-DR do not violate monotonicity restrictions.

7 Conclusion
Heterogeneous treatment effects of a policy intervention are often of great interest, because poli-
cies are often motivated to improve the outcomes for those at the lower end of the outcome
distribution or to reduce inter-group outcome gaps. For example, welfare policies often target
those at the lower end of the income distribution or are intended to reduce labor market outcome
disparities between sub-groups of interest. To reliably estimate the differential effects of a given

35
treatment across the outcome distribution, controlling for the effect of potential observed con-
founders is crucial. Doing so in a restrictive manner e.g. assuming that a given control variable’s
effect on the entire outcome distribution is the same, however, can yield misleading estimates.
Towards this effort, I propose a distribution regression-based changes-in-changes estimator
and derive its asymptotic distributions along with bootstrap validity for inference. This paper
also examines the heterogeneous effects of EITC on infant weights across the entire weight dis-
tribution and find more concentrated benefits at the low end of the distribution, while more
muted effects across the rest of the distribution than previously reported. I also demonstrate that
average effects may mask important heterogeneity and deem policies unhelpful for improving
outcomes where they matter. For example, even though EITC served to help reduce low birth
weight incidences and increase the birth weights at the bottom of the weight distribution which
is the most consequential part of the distribution for downstream health outcomes, estimates of
average treatment effects on the treated suggest that the policy effect was statistically indistin-
guishable from null.
The current paper focuses on repeated cross sectional data in the canonical 2 × 2 setting with
low-dimensional controls. It would also be insightful to develop theory for when panel data is
available, when treatments are staggered in their adoption (along with a means to aggregate the
quantile or distributional treatment effects) and when high-dimensional controls are present. The
last direction of research can be useful when there are a large number of potential confounders,
but the researcher does not know which ones are the most relevant in determining conditional
outcome trends over time.

36
References
Almond, D., Chay, K. Y., and Lee, D. S. (2005). The costs of low birth weight. The Quarterly Journal of
Economics, 120(3):1031–1083.

Almond, D., Hoynes, H. W., and Schanzenbach, D. W. (2011). Inside the war on poverty: The impact of
food stamps on birth outcomes. The review of economics and statistics, 93(2):387–403.

Arkhangelsky, D. et al. (2019). Dealing with a technological bias: The difference-in-difference approach.
Centro de Estudios Monetarios y Financieros.

Athey, S. and Imbens, G. W. (2006). Identification and inference in nonlinear difference-in-differences

models. Econometrica, 74(2):431–497.

Barrett, G. F. and Donald, S. G. (2003). Consistent tests for stochastic dominance. Econometrica, 71(1):71–
104.

Blau, F. D., Cohen, I., Comey, M. L., Kahn, L., and Boboshko, N. (2023). The minimum wage and inequality
between groups. (31725).

Blinder, A. S. (1973). Wage discrimination: Reduced form and structural estimates. The Journal of Human
Resources, 8(4):436–455.

Bonhomme, S. and Sauder, U. (2011). Recovering distributions in difference-in-differences models: A com-

parison of selective and comprehensive schooling. Review of Economics and Statistics, 93(2):479–494.

Callaway, B. (2021). Bounds on distributional treatment effect parameters using panel data with an appli-
cation on job displacement. Journal of Econometrics, 222(2):861–881.

Callaway, B. and Li, T. (2019). Quantile treatment effects in difference in differences models with panel
data. Quantitative Economics, 10(4):1579–1618.

Cárcamo, J., Cuevas, A., and Rodrı́guez, L.-A. (2020). Directional differentiability for supremum-type func-
tionals: Statistical applications. Bernoulli, 26(3):2143 – 2175.

Cengiz, D., Dube, A., Lindner, A., and Zipperer, B. (2019). The effect of minimum wages on low-wage jobs.
The Quarterly Journal of Economics, 134(3):1405–1454.

Chernozhukov, V., Fernández-Val, I., and Galichon, A. (2010). Quantile and probability curves without
crossing. Econometrica, 78(3):1093–1125.

Chernozhukov, V., Fernández-Val, I., and Melly, B. (2013). Inference on counterfactual distributions. Econo-
metrica, 81(6):2205–2268.

Chernozhukov, V., Fernandez-Val, I., Melly, B., and Wüthrich, K. (2019). Generic inference on quantile and
quantile effect functions for discrete outcomes. Journal of the American Statistical Association.

Currie, J. (2011). Inequality at birth: Some causes and consequences. American Economic Review,
101(3):1–22.

Currie, J., Neidell, M., and Schmieder, J. F. (2009). Air pollution and infant health: Lessons from new jersey.
Journal of Health Economics, 28(3):688–703.

37
Derenoncourt, E. and Montialoux, C. (2021). Minimum Wages and Racial Inequality*. The Quarterly Journal
of Economics, 136(1):169–228.

Foresi, S. and Peracchi, F. (1995). The conditional distribution of excess returns: An empirical analysis.
Journal of the American Statistical Association, 90(430):451–466.

Fortin, N., Lemieux, T., and Firpo, S. (2011). Decomposition methods in economics. In Handbook of labor
economics, volume 4, pages 1–102. Elsevier.

Fortin, N. M., Lemieux, T., and Lloyd, N. (2021). Labor market institutions and the distribution of wages:
The role of spillover effects. Journal of Labor Economics, 39(S2):S369–S412.

Ghanem, D., Kédagni, D., and Mourifié, I. (2023). Evaluating the impact of regulatory policies on social
welfare in difference-in-difference settings. arXiv preprint arXiv:2306.04494.

Gunsilius, F. F. (2023). Distributional synthetic controls. Econometrica, 91(3):1105–1117.

Havnes, T. and Mogstad, M. (2015). Is universal child care leveling the playing field? Journal of public
economics, 127:100–114.

Hoynes, H., Miller, D., and Simon, D. (2015). Income, the earned income tax credit, and infant health.
American Economic Journal: Economic Policy, 7(1):172–211.

Kim, D. and Wooldridge, J. M. (2023). Difference-in-differences estimator of quantile treat-

ment effect on the treated. Working Paper. Available at: [Link]
vpKBdC4JddZX2OUacyaijje78eMWv/view (Accessed: May 23rd, 2024).

Lechner, M. (2011). The estimation of causal effects by difference-in-difference methods. Foundations and
Trends® in Econometrics, 4(3):165–224.

Melly, B. and Santangelo, G. (2015). The changes-in-changes model with covariates. Universität Bern, Bern,
1.

Oaxaca, R. (1973). Male-female wage differentials in urban labor markets. International Economic Review,
14(3):693–709.

Roth, J. and Sant’Anna, P. H. (2023). When is parallel trends sensitive to functional form? Econometrica,
91(2):737–747.

Sasaki, Y. and Wang, Y. (2024). Extreme changes in changes. Journal of Business & Economic Statistics,
42(2):812–824.

Van Der Vaart, A. W. and Wellner, J. A. (1996). Weak convergence and Empirical Processes. Springer.

Wursten, J. and Reich, M. (2023). Racial inequality in frictional labor markets: Evidence from minimum
wages. Labour Economics, 82:102344.

Wüthrich, K. (2019). A closed-form estimator for quantile treatment effects with endogeneity. Journal of
Econometrics, 210(2):219–235.

38
A Appendix: Proofs

A.1 Notation and Statistical Definitions

To present the proofs for the theorems and corollaries in the main text, I formally introduce
notation to describe the mathematical set up and the relevant function spaces. Let ℓ∞ (YX ) denote
the set of all bounded and measurable mappings YX 7→ R. YX is a subset of R̄1+dx with topology
induced by the standard metric, ρ, on R̄1+dx , where R̄ is the extended real line. Also let λ(f, f˜) =
  2 1/2
f − f˜ dFX be a metric on F. Here, F is defined as a class of measurable functions that
R

includes {FYgt |X (y|·) : y ∈ Ygt , (g, t) ∈ {0, 1}2 } as well as the indicators of all rectangles in R̄dx
such that F is totally bounded under λ. The preceding definition ensures that the uniform CLT
holds for the estimators of the conditional distribution function and control variable distributions.

A.2 Proof of Theorem 1

I utilize the derivative map obtained in Appendix SE.3 of Callaway (2021) to apply the functional
delta method. Specifically, let Ȳgt X be the support of Ygt conditional on X, D := ℓ∞ (Y10 X ) ×
ℓ∞ (Y00 X ) × ℓ∞ (Y01 X ) and consider the map, ϕ : Dϕ ⊂ D 7→ ℓ∞ (Ȳ01 X ) defined by

ϕ(F1 , F2 , F3 ) = F1 ◦ F2−1 ◦ F3

for F = (F1 , F2 , F3 ) ∈ Dϕ . Here, Dϕ = E 3 where E denotes the set of all distribution functions
with density function that is uniformly bounded above and bounded away from zero. Then, ϕ is
Hadamard differentiable at F0 = (FY10 |X , FY00 |X , FY01 |X ) ∈ D with the derivative map given by,

  γ3 − γ2 ◦ F −1 ◦ FY01 |X
ϕ′F0 (γ) = γ1 ◦ FY−1
00 |X
◦ FY01 |X + fY10 |X FY−1
00 |X
◦ FY01 |X  Y00 |X 
fY00 |X FY−1
00 |X
◦ F Y01 |X

tangentially to Dϕ in γ = (γ1 , γ2 , γ3 ) ∈ Dϕ . This can be derived from the functional chain rule
and from the result that composition of Hadamard differentiable maps are themselves Hadamard

39
differentiable (Lemmas 3.10.24 and 3.10.28 of Van Der Vaart and Wellner (1996)). Then taking the
derivative map and applying the functional delta method as in Theorem 3.10.4 of Van Der Vaart
and Wellner (1996) based on the limit laws of the input conditional distribution processes (i.e.
Z10 (y, x), Z00 (y, x), Z01 (y, x)) yields the limit law for the conditional CIC-DR process,



−1
ZN
11 (y, x) =Z10 F Y00 |X F Y01 |X (y|x)|x ,x


+ fY10 |X FY−1 00 |X
F Y01 |X (y|x)|x |x Z00,01 (y, x)

where, 

Z01 (y, x) − Z00 FY−1 00 |X
F Y01 |X (y|x)|x ,x
Z00,01 (y, x) =  
fY10 |X FY−1


00 |X
F Y 01 |X (y|x)|x |x

The result for the asymptotic distribution of F̂Y11I |X follows directly from Corollary 5.4 of Cher-
nozhukov et al. (2013) as shown in the expression following Assumption 3. This completes the
proof of Theorem 1.

A.3 Proof of Corollary 1.1

To show that the DR-based UQTT process is asymptotically Gaussian, note that the UQTT process
is the difference in the conditional quantile process integrated over the control variable distribu-
tion between the treated and the untreated potential outcomes.

Z −1 Z −1
δ̂U QT T (τ ) = F̂YI11 |X (y|x)dF̂X11 (x) − F̂YN11 |X (y|x)dF̂X11 (x)
X11 X11

This is a smooth functional of the respective distribution functions, and therefore the functional
delta method can be applied to the following map based on Theorem 1. To be concrete, let Dφ
be the product space of measurable functions, Γ : Ygt X 7→ [0, 1] defined by (y, x) 7→ Γ(y, x)
and Π : F 7→ R defined by f 7→ f dΠ, where Π is bounded and restricted to be a probability
R

40
measure on X . Consider the map φ : Dφ ⊂ D = ℓ∞ (Ygt X )2 × ℓ∞ (F) 7→ E = ℓ∞ (T ) defined by

Z −1 Z −1
φ(Γ1 , Γ2 , Π) := Γ1 (·, x)dΠ(x) − Γ2 (·, x)dΠ(x) .

Then, φ can be defined as a composition of two maps, φ1 : Dφ 7→ Dφ2 and φ2 : Dφ2 7→ ℓ∞ (T )

where Dφ2 = ℓ∞ (Ygt ) defined by,

φ(Γ1 , Γ2 , Π) = φ2 (φ1 (Γ1 , Π), φ1 (Γ2 , Π))

where
Z
φ1 (Γ, Π) := Γ(·, x)dΠ(x) and φ2 (Ψ1 , Ψ2 ) := Ψ−1 −1
1 − Ψ2

for Γ ∈ Dvarphi and By Lemma D.1. of Chernozhukov et al. (2013), ϕ1 is Hadamard differentiable
at (FYgt |X , FX ) tangentially to the subset, D0 = U C(Ygt X , ρ) × U C(F, λ) with the derivative
map,

Z
ϕ′1,FY |X ,FX (γ, π)(y) = γ(y, x)dFX (x) + π(FYgt |X (y|·)) (5)
gt

ϕ2 is also a difference of inverse maps, which are Hadamard differentiable by Lemma 3.10.21 of
Van Der Vaart and Wellner (1996) with the derivative map given by,

k1 k2
ϕ′2,Ψ1 ,Ψ2 (k1 , k2 ) = − ◦ Ψ−1
1 + ◦ Ψ−1
2
ψ1 ψ2

41
Then, by Lemmas 3.10.24 and 3.10.28 of Van Der Vaart and Wellner (1996), ϕ is also Hadamard
differentiable at (FY11I |X , FY11N |X , FX ) with the derivative map given by,

ϕ′FI |X ,FY N |X ,FX (γ1 , γ2 , π)

Y11 11
Z Z 
= ϕ′2,ϕ1 (F I γ1 (y, x)dFX (x) + π(FY11I |X (y|·)),
,FX ),ϕ1 (FY N |X ,FX ) γ2 (y, x)dFX (x) + π(FY11N |X (y|·))
11 |X
Y 11
R R
γ1 (y, x)dFX (x) + π(FY11I |X (y|·)) γ2 (y, x)dFX (x) + π(FY11N |X (y|·))
=− ◦ FY−1I + ◦ FY−1N
fY11I 11 fY11N 11

Next, defining the empirical processes for the conditional distribution process and the control
variable distribution as,

√  
Ẑgt (y, x) := N F̂Ygt |X (y|x) − FYgt |X (y|x)
Z  
p
Ĝgt (f ) := Ngt f d F̂X − FX
Xgt

where F̂X is the empirical distribution of X. Under Assumption 2 and 3, and as N → ∞ with
N/Ngt → sgt ∈ [0, ∞), along with the definition of F in Appendix A.1, the following holds

 
I N
(y, x), Ĝ11 (f ) ⇝ ZI11 (y, x), ZN


Ẑ11 (y, x), Ẑ11 11 (y, x), G11 (f )

as stochastic processes indexed by (y, x, g, t, f ) ∈ YX {0, 1}2 F. The limit process is a tight mean-
zero Gaussian process, where Zgt and Ggt almost surely have uniformly continuous paths with
respect to the metric ρ and λ respectively.
Finally, apply the functional delta method to the composition of Hadamard differentiable maps
and obtain the following limit laws for the UQTT.

√  
N δ̂U QT T − δU QT T ⇝ ZU QT T (τ )

as a stochastic process indexed by τ ∈ T , and ZU QT T (τ ) is a tight mean-zero Gaussian process

42
defined by,    
ZI11,U FY−1I (τ ) ZN
11,U F −1
Y N (τ )
ZU QT T (τ ) = −  11  +  11 
fY11I FY−1I (τ ) fY11N FY−1N (τ )
11 11

where

Z  
ZI11,U (y) := ZI11 (y, x)dFX (x) + G FY11I |X (y|·)
Z X11  
ZN
11,U (y) := ZN
11 (y, x)dFX (x) + G F N
Y11 |X (y|·)
X11

Note that for UDTT, the one need only consider the first Hadamard differentiable map, ϕ1 , and
apply the functional delta method directly to obtain the limit law,

ZU DT T (y) = ZI11,U (y) − ZN

11,U (y)

This completes the proof of Corollary 1.1.

A.4 Proof of Theorem 2

This follows directly from the functional delta method. For quantities involving the integration
over the control variable distribution, the same map found in expression (5) can be used.

A.5 Proof of Theorem 3

Under the stated assumptions and results found in (2), by Corollary 5.4 of Chernozhukov et al.
(2013), the exchangeable bootstrap is valid for the the DR-based estimators for the conditional
distribution and quantile functions. Combining this with Theorem 1 and the Hadamard differen-
tiability of the functionals involved, one can apply the functional delta method for the bootstrap
based on section 3.10.3 of Van Der Vaart and Wellner (1996) and the result holds.

43
B Appendix: Figures

(a) Pre: 1993, Post: 1994 (b) Pre: 1993, Post: 1995

(c) Pre: 1993, Post: 1996 (d) Pre: 1993, Post: 1997

Figure 5: Pairwise UDTT of EITC on Birth Weight (g). Blue line plots estimates from CIC-DR and red line
plots estimates from OLS-residualized CIC. Shaded areas are the uniform 90% confidence band obtained
from 500 weighted bootstrap replications.

44
(a) Pre: 1993, Post: 1997 (b) Pre: 1993, Post: 1997

(c) Pre: 1993, Post: 1997

Figure 6: β01 (y) for select mothers’ characteristics.

45
 (a) Pre: 1993, Post: 1994 (b) Pre: 1993, Post: 1994

(c) Pre: 1993, Post: 1995 (d) Pre: 1993, Post: 1995

Figure 7: Marginal distribution and quantile functions of the pairwise samples.

46
 (a) Pre: 1993, Post: 1996 (b) Pre: 1993, Post: 1996

(c) Pre: 1993, Post: 1997 (d) Pre: 1993, Post: 1997

Figure 8: Marginal distribution and quantile functions of the pairwise samples.

47