Sharp Bounds and Inference in Sample Selection Models with Treatment Endogeneity
Pith reviewed 2026-06-27 14:25 UTC · model grok-4.3
The pith
The paper derives sharp bounds on intensive margin treatment effects among compliers in nonparametric sample selection models with endogenous treatment under weak monotonicity and supplies debiased machine learning inference that achieves r
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
Under the weak sample selection monotonicity assumption, the proposed bounds for intensive margin treatment effects among compliers are sharp and tighter than those of Chen and Flores (2015). Semiparametrically efficient orthogonal moments and a debiased machine learning procedure permit valid root-n inference under high-dimensional covariates and flexible functional forms. Simulation results indicate good finite sample performance. Applications to Job Corps and the Oregon Health Insurance Experiment show that the method can deliver substantially tighter effect bounds and confidence intervals than existing alternatives.
What carries the argument
Sharp bounds on intensive margin complier treatment effects derived from weak sample selection monotonicity, implemented through semiparametrically efficient orthogonal moments and debiased machine learning for inference.
If this is right
- The bounds yield substantially tighter intervals than prior methods when applied to job training and health insurance experiments.
- Root-n inference remains valid even after using machine learning to estimate high-dimensional nuisance functions.
- Simulations confirm reliable coverage and length properties for the resulting confidence intervals in moderate samples.
- The partial identification strategy applies directly to other nonparametric models that combine endogenous treatment with nonrandom sample selection.
Where Pith is reading between the lines
- The same monotonicity-based bounding logic could be applied to labor-market or education settings that feature both noncompliance and selective observation.
- Orthogonal-moment constructions of this type may extend to other partial-identification problems that mix endogeneity with selection.
- Testing the monotonicity assumption itself with the same data would be a natural next diagnostic step left open by the identification result.
Load-bearing premise
The weak sample selection monotonicity assumption that the probability of selection into the observed sample is nondecreasing in treatment assignment.
What would settle it
Empirical evidence that the selection probability falls with treatment assignment in a manner that violates monotonicity, or data in which the proposed bounds are crossed while point estimates from alternative identification strategies lie outside them.
read the original abstract
This paper provides partial identification and inference for treatment effects in nonparametric sample selection models with endogenous treatment and (weak) sample selection monotonicity. Outcomes are observed only for a non-randomly selected subsample and treatment is endogenous because of noncompliance with assignment. The proposed bounds for intensive margin treatment effects among compliers are sharp and tighter than those of Chen and Flores (2015). For inference, we develop semiparametrically efficient orthogonal moments and a debiased machine learning procedure that permits valid root-$n$ inference under high-dimensional covariates and/or flexible functional forms. Simulation results indicate good finite sample performance. Applications to Job Corps and the Oregon Health Insurance Experiment show that the method can deliver substantially tighter effect bounds and confidence intervals than existing alternatives.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper develops partial identification and inference methods for treatment effects in nonparametric sample selection models with endogenous treatment (noncompliance) and weak sample selection monotonicity. The central contribution is a set of sharp bounds on intensive-margin treatment effects among compliers that are strictly tighter than the Chen-Flores (2015) bounds; these are paired with semiparametrically efficient orthogonal moments and a debiased machine-learning estimator that delivers root-n inference under high-dimensional covariates or flexible functional forms. Simulations and two empirical applications (Job Corps, Oregon Health Insurance Experiment) are provided.
Significance. If the sharpness result and the efficiency of the orthogonal moments hold, the paper supplies a practically useful tightening of the identified set for a policy-relevant parameter (complier intensive-margin effects) together with an implementable inference procedure that accommodates modern high-dimensional settings. The explicit use of orthogonal moments and DML is a methodological strength that facilitates credible empirical work.
major comments (2)
- [Abstract; §3 (Identification)] The claim that the bounds are sharp and strictly tighter than Chen and Flores (2015) rests entirely on the weak sample selection monotonicity assumption (never-selected units unaffected by treatment). The manuscript provides no sensitivity analysis showing how the identified set expands when this assumption is relaxed even mildly, nor primitive conditions under which the assumption is implied by an economic model. This is load-bearing for the headline result on tighter bounds.
- [§4.2] §4.2, the construction of the orthogonal moments: it is stated that the moments are semiparametrically efficient, but the efficiency bound is not explicitly derived or referenced under the joint maintained assumptions (including monotonicity and the selection model). Without this step it is difficult to verify that the proposed moments attain the bound.
minor comments (2)
- [Table 1] Table 1 (simulation design) reports coverage probabilities but does not include the width of the confidence intervals; adding this column would make the finite-sample comparison with existing methods more informative.
- [§2] The notation for the selection indicator and the observed outcome is introduced without an explicit comparison table to the Chen-Flores setup; a short notational concordance would improve readability.
Simulated Author's Rebuttal
We thank the referee for the constructive and detailed comments. We address each major comment below.
read point-by-point responses
-
Referee: [Abstract; §3 (Identification)] The claim that the bounds are sharp and strictly tighter than Chen and Flores (2015) rests entirely on the weak sample selection monotonicity assumption (never-selected units unaffected by treatment). The manuscript provides no sensitivity analysis showing how the identified set expands when this assumption is relaxed even mildly, nor primitive conditions under which the assumption is implied by an economic model. This is load-bearing for the headline result on tighter bounds.
Authors: The sharpness and tightening relative to Chen and Flores (2015) are established under the maintained weak sample selection monotonicity assumption, which is stated explicitly in the paper. We agree that robustness checks would strengthen the presentation. In the revision we will add a short subsection to §3 that (i) reports how the identified set expands under mild relaxations of the assumption and (ii) supplies primitive economic conditions (e.g., treatment does not affect selection for units that would never be selected under either treatment status) under which the assumption is implied by standard selection models. This addition directly addresses the load-bearing nature of the assumption while preserving the headline result under the stated conditions. revision: partial
-
Referee: [§4.2] §4.2, the construction of the orthogonal moments: it is stated that the moments are semiparametrically efficient, but the efficiency bound is not explicitly derived or referenced under the joint maintained assumptions (including monotonicity and the selection model). Without this step it is difficult to verify that the proposed moments attain the bound.
Authors: We thank the referee for this observation. The orthogonal moments were derived to attain the semiparametric efficiency bound under the joint maintained assumptions, but an explicit derivation of that bound was omitted from the current draft. In the revision we will add to §4.2 (i) the derivation of the efficiency bound under the nonparametric selection model together with the weak monotonicity restriction and (ii) a verification that the proposed moments achieve this bound. This will make the efficiency claim fully verifiable. revision: yes
Circularity Check
No circularity: bounds derived from explicit monotonicity assumption; inference uses independent orthogonal moments
full rationale
The paper states that sharp bounds on complier intensive-margin effects follow from the weak sample selection monotonicity assumption (never-selected units unaffected by treatment), which is an explicit maintained restriction rather than a self-referential definition. Tightness relative to Chen-Flores (2015) is obtained by adding this assumption to the model; sharpness is the standard partial-identification claim that the bounds are attained under the stated restrictions. The inference contribution (semiparametrically efficient orthogonal moments and debiased ML) is constructed separately from the identification step and does not reduce to fitted parameters or self-citations. No load-bearing self-citation, ansatz smuggling, or renaming of known results appears in the provided abstract or description. The derivation chain is therefore self-contained against external benchmarks.
Axiom & Free-Parameter Ledger
Reference graph
Works this paper leans on
-
[1]
Abadie, A. (2003). Semiparametric Instrumental Variable Estimation of Treatment Re- sponse Models. Journal of Econometrics, 113(2):231–263. Andersen, S. C., Beuchert, L., Heiler, P., and Nielsen, H. S. (2023). A guide to impact evaluation under sample selection and missing data: Teacher’s aides and adolescent mental health. Angrist, J. D. (2004). Treatmen...
2003
-
[2]
Bartalotti, O., Kédagni, D., and Possebom, V. (2023). Identifying marginal treatment effects in the presence of sample selection.Journal of Econometrics, 234(2):565–584. Bickel, P. J., Klaassen, C. A., Bickel, P. J., Ritov, Y., Klaassen, J., Wellner, J. A., and Ritov, Y. (1993).Efficientand AdaptiveEstimation for Semiparametric Models. Johns Hopkins Unive...
2023
-
[3]
A., Flores-Lagunes, A., Gonzalez, A., and Neumann, T
36 Flores, C. A., Flores-Lagunes, A., Gonzalez, A., and Neumann, T. C. (2012). Estimating the effects of length of exposure to instruction in a training program: The case of job corps. Review of Economics and Statistics, 94(1):153–171. Foresi, S. and Peracchi, F. (1995). The Conditional Distribution of Excess Returns: An Empirical Analysis.Journal of the ...
2012
-
[4]
truncation-by-death
Hines, O., Dukes, O., Diaz-Ordaz, K., and Vansteelandt, S. (2022). Demystifying statisti- callearningbasedonefficientinfluencefunctions. AmericanStatistician, 76(3):292–304. Horowitz, J.L.andManski, C.F.(1995). Identificationandrobustnesswithcontaminated and corrupted data.Econometrica, pages 281–302. Horowitz, J. L. and Manski, C. F. (2000). Nonparametri...
2022
-
[5]
Klein, N. (2024). Distributional regression for data analysis.Annual Review of Statistics and Its Application, 11(Volume 11, 2024):321–346. Lechner, M. and Melly, B. (2010). Partial identification of wage effects of training pro- grams. Lee, D. S. (2009). Training, wages, and sample selection: Estimating sharp bounds on treatment effects.Review of Economi...
-
[6]
any point within the interval[βL,1,βU,1]is feasible and compatible with our assumptions and observed data, suggestβL,1 (βU,1) is sharp, i.e.,βL,1 (βU,1) is the largest (smallest) lower (upper) bound forβ1 :=E[Y 1|ac]that is consistent with our assumptions and observed data. A.2. Supplementary Material for Section 2.2.1 A.2.1. Lee (2009) ITT bounds and SLA...
2009
-
[7]
IF ofβB,1 conditional onx
It assumes thatacandaa are at the bottompCF 2 := πac+πaa πac+πcc+πaa fraction of the conditional distributionY1|ac∪cc∪aa, i.e. FY1|ac∪aa(y) =F adj Y1|ac∪aa(y) := 1 pCF 2 FY1|ac∪cc∪aa(y)ify <Q cf 1 ( pCF 2 ) 1ify≥Q cf 1 ( pCF 2 ).(A.11) In contrast, we have shown thatβL,1 assumes FY1|ac(y) =F dh Y1|ac(y) := 1 pCF 1 FY1|ac∪cc∪aa(y)−πaa πac F...
2024
-
[8]
=W(Z,X)((1−D)SY−µ(Z,X)) +β0(X)W(Z,X)((1−D)S−m(Z,X)) −β0(X)W(Z,X)((1−D)S−m(Z,X))−β0(X)(m(1,X)−m(0,X))−N0 =W(Z,X)((1−D)SY−µ(Z,X)) + [µ(1,X)−µ(0,X)]−N0 and 60 IF(N L,1) =W(Z,X)(DS−r(Z,X))ψL(Z,X) −Q1(p(X),X)W(Z,X) [ ((1−D)S−m(Z,X)) + (DS−r(Z,X))F(Q1(p(X),X)|1,Z,X) +DS(1(Y≤Q 1(p(X),X))−F(Q1(p(X),X)|1,Z,X)) ] +W(Z,X)DS[Y1(Y≤Q 1(p(X),X))−ψL(Z,X)] +ψL(1,X)r(1,X)−...
2024
-
[9]
augmented IPW
The second-order term is Rn(βB) := Ψ(ˆη)−Ψ(η). The cross-fitted stabilized estimating equation uses the EIF IF(β B) = IF(N + 0 ) +IF(N − 0 ) +IF(N + B,1) +IF(N − B,1)−βB IF(π ac) E[πac(X)] . We now denote shorthand in what followsIFβB =IF(β B,η)and ˆIF(β B) =IF(β B,ˆη)be- ing the EIF with estimated nuisances. By construction, our EIF are Neyman-orthogonal...
1993
-
[10]
(2024), Table 6.1
This is exactly the denominator in Heiler et al. (2024), Table 6.1. We now consider the numerator given byN1 +N
2024
-
[11]
(2024), Table 6.1., see also Semenova (2025), Section 6.6
+N 0 = DS e(X)Y1{Y≤Q1(p(X),X)}−(1−D)S 1−e(X)Y −DS e(X)Q1(p(X),X) [1{Y≤Q1(p(X),X)}−p(X)] +Q 1(p(X),X) [ 1−D 1−e(X)(S−s(0,X))−p(X)D e(X) (S−s(1,X)) ] +s(0,X) [ β1,1(X,p(X)) ( 1−D e(X) ) −β0,0(X,0) ( 1−1−D 1−e(X) )] , which is exactly the corresponding component of the influence function proposed in Heiler et al. (2024), Table 6.1., see also Semenova (2025),...
2024
-
[12]
Combining results with the denominator yield the well-known EIF for the LATE (Frölich, 2007; Chernozhukov et al., 2018; Heiler, 2022)
+N 0 =W(Z,X)[DY−ψL(Z,X)r(Z,X)] +ψL(1,X)r(1,X)−ψL(0,X)r(0,X) +W(Z,X)[(1−D)Y−µ(Z,X)] +µ(1,X)−µ(0,X) =W(Z,X)Y+ [ψ L(1,X)r(1,X) +µ(1,X)] ( 1− Z P(Z= 1|X) ) −[ψL(0,X)r(0,X) +µ(0,X)] ( 1− (1−Z) 1−P(Z= 1|X) ) =W(Z,X)Y+E[Y|Z= 1,X] ( 1− Z P(Z= 1|X) ) −E[Y|Z= 0,X] ( 1− (1−Z) 1−P(Z= 1|X) ) = Z P(Z= 1|X)(Y−E[Y|Z= 1,X])−(1−Z) 1−P(Z= 1|X)(Y−E[Y|Z= 0,X]) +E[Y|Z= 1,X]−E[...
2007
-
[13]
+N 0 = D P(D= 1|X)(Y−E[Y|D= 1,X])−(1−D) 1−P(D= 1|X)(Y−E[Y|D= 0,X]) +E[Y|D= 1,X]−E[Y|D= 0,X], which is the usual efficient AIPW influence function for the ATE (Hahn, 1998). D. Monte Carlo Simulations D.1. Design In this section we report a simple Monte Carlo design to assess coverage and power properties of the suggested confidence intervals in Section 4.3...
1998
-
[14]
Cross-fitting seems to be required to have the minimal rejection area close to the true null
Given the design, this corresponds to an effective sample size of approximately600and1500always- selected compliers respectively. Cross-fitting seems to be required to have the minimal rejection area close to the true null. The confidence intervals have conservative coverage at and close to the null. This is expected as the effect is not at the boundary o...
2020
-
[15]
n= 2000andn= 5000observations andM= 200replications
Regular is equivalent but without cross-fitting. n= 2000andn= 5000observations andM= 200replications. E. Empirical Study II: Oregon Health Insurance Experiment E.1. Data and Methods In this section, we study the effects of Medicaid coverage on healthcare utilization in the Oregon Health Insurance Experiment (OHIE), following Finkelstein et al. (2012). In ...
2012
-
[16]
A consistent pattern emerges
Tables E.2–E.4 compare baseline characteristics of always-selected compliersacand the full sample across the three utilization outcomes. A consistent pattern emerges. On average, for outpatient visits and prescription drugs,acare somewhat older (about 1.7 years), more likely to be White, and substantially less likely to be Hispanic, while gender differenc...
2019
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.