Logit-based alternatives to two-stage least squares

arXiv:2312.10333 · econ.EM, math.ST, stat.TH · Submitted 2023-12-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Logit-based alternatives to two-stage least squares".

Warren: Here's the gist of the paper. We propose logit-based IV and augmented logit-based IV estimators that serve as alternatives to the traditionally used 2SLS estimator in the model where both the endogenous treatment variable and the corresponding instrument are binary.

Ruth: Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability.

Warren: In particular, in certain cases where the probability limits of both our estimators and the 2SLS estimator take the form of weighted-average treatment effects, our estimators are guaranteed to yield non-negative weights whereas the 2SLS estimator is not.

Ruth: The paper introduces a new instrumental variable estimator that consists of two steps, similar to the 2SLS estimator.

Title and authors: Ruth: The title itself, "Logit-based alternatives to two-stage least squares," sets the stage immediately by telling us exactly what this paper is offering: options that use logit models instead of just the standard two stage least squares approach.

Warren: I mean, it’s not a total overhaul of estimation techniques, but it’s tailoring the method to be more robust when the data structure involves binary variables for both the treatment and the instrument.

Iris: They are focusing specifically on settings where you have a binary treatment variable and a binary instrument variable present alongside controls in your model.

Ruth: That’s right, they aren't just looking at general IV estimation; they are zeroing in on that specific challenging scenario where the standard 2SLS might run into trouble <ref:2312.10333#pg1>.

Warren: And their main contribution is showing how these logit-based methods can provide a clearer picture of what you are actually estimating, which is something I always worry about when running complex models.

The paper's summary: Ruth: To summarize the paper "Logit-based alternatives to two-stage least squares," they introduce these two new estimators—the logit-based IV and the augmented logit-based IV—as substitutes for the traditional 2SLS estimator in this binary treatment and binary instrument setting <ref:2312.10333#pg1,Logit-based alternatives to two-stage least squares>.

Warren: The summary highlights that these new methods are as straightforward to compute as the 2SLS estimator, which is a big plus for practical application, but they offer an advantage in causal interpretability <ref:2312.10333#pg1,to compute as the 2SLS estimator>.

Iris: They point out that the key improvement comes when certain conditions are met, because under those conditions, the probability limits of their estimators take a specific form.

Ruth: That specific form is what gives them the causal interpretation we're looking for: it means they are guaranteed to yield non-negative weights in the complier term.

Warren: The paper explicitly states that the 2SLS estimator doesn't have this guarantee; it might assign negative weights to a significant fraction of people who only take treatment under certain conditions <ref:2312.10333#pg1>.

The paper's improvements: Iris: One major improvement they highlight is how these estimators handle the probability limits when they form a weighted-average treatment effect, ensuring those weights are always non-negative, which is crucial for causal interpretation.

Ruth: And they show that this advantage becomes particularly important when you have additional conditions in place—conditions that guarantee the always-taker, never-taker, and non-causal terms all vanish.

Warren: If those terms vanish, the logit-based IV estimator converges to a convex combination of treatment effects for compliers, while the 2SLS estimator’s limit isn't necessarily that same structure <ref:2312.10333#pg1>.

Iris: They also developed an augmented version of the logit-based IV estimator which adds an extra term in the initial logit regression, based on a binary model of the treatment variable on controls using a fixed value of the instrument.

Ruth: That augmented estimator is interesting because it’s shown to maintain causal interpretability even when that specific set of conditions isn't fully met, provided you use the correct link function.

Conclusion: Ruth: So, to wrap up on "Logit-based alternatives to two-stage least squares," the paper shows that these logit-based IV and augmented estimators are robust alternatives because they avoid the problem of negative weights in complier terms that you might see with 2SLS <ref:2312.10333#pg1,Logit-based alternatives to two-stage least squares>.

Warren: They confirm these estimators are nearly identical to some existing results, like those from Angrist and Evans or Angrist et al., but they differ significantly from other studies, like Dube and Harish.

Iris: The paper suggests that if you run a Hausman specification test on the conditional mean of Z given X and it doesn't reject the null hypothesis about assumption two point six being satisfied, then you can confirm those estimators are estimating the weighted-average treatment effect for compliers with weights given by that specific formula.

Ruth: It seems like a practical takeaway for anyone running these types of binary data models is to check those conditions before settling on 2SLS, because this paper gives us a clear pathway to causality when 2SLS might be misleading <ref:2312.10333#pg1>.

Warren: Yeah, it’s about having an alternative strategy that keeps the interpretation clear when the underlying assumptions are tricky.

DENIS CHETVERIKOV, JINYONG HAHN, ZHIPENG LIAO, SHUYANG SHENG

Department of Economics UCLA · Shenzhen Finance Institute, School of Management and Economics, The Chinese University of Hong Kong

econ.EM, math.ST, stat.TH

Submitted: 2023-12-16

Updated: 2026-02-18

Comments: 40 pages

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: We propose logit-based IV and augmented logit-based IV estimators that serve as alternatives to the traditionally used 2SLS estimator in the model where both the endogenous treatment variable and the…

Key concepts

Logit-based IV Estimator
This estimator uses a two-step process involving logit regressions. It first regresses the instrument on controls to get residuals, and then uses those residuals as an instrument in a subsequent instrumental variable regression for the outcome. This approach is designed to handle binary treatment variables effectively.
Causal Interpretability
This refers to whether the resulting probability limit of an estimator can be interpreted as a weighted-average treatment effect where all weights are non-negative. The logit-based IV estimator guarantees this, providing a clear causal interpretation that 2SLS often lacks.
Augmented Logit-based IV Estimator
This is an improved version of the main estimator that includes an extra term in the initial logit regression. This augmentation helps ensure causal interpretability even when standard assumptions are not perfectly met, offering greater robustness for causal inference.

Terminology

Summary

We propose logit-based IV and augmented logit-based IV estimators that serve as alternatives to the traditionally used 2SLS estimator in the model where both the endogenous treatment variable and the corresponding instrument are binary. p. 0

Where the paper says it

LOGIT-BASED ALTERNATIVES TO TWO-STAGE LEAST SQUARES DENIS CHETVERIKOV, JINYONG HAHN, ZHIPENG LIAO, AND SHUYANG SHENG ABSTRACT. We propose logit-based IV and augmented logit-based IV estimators that serve as alternatives to the traditionally used 2SLS estimator in the model where both the endogenous treatment variable and the corresponding instrument are binary.

Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability. p. 0
Where the paper says it

We propose logit-based IV and augmented logit-based IV estimators that serve as alternatives to the traditionally used 2SLS estimator in the model where both the endogenous treatment variable and the corresponding instrument are binary. Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability.

In particular, in certain cases where the probability limits of both our estimators and the 2SLS estimator take the form of weighted-average treatment effects, our estimators are guaranteed to yield non-negative weights whereas the 2SLS estimator is not. p. 0
Where the paper says it

Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability. In particular, in certain cases where the probability limits of both our estimators and the 2SLS estimator take the form of weighted-average treatment effects, our estimators are guaranteed to yield non-negative weights whereas the 2SLS estimator is not.

Introduction to Estimator Design

The paper introduces a new instrumental variable estimator that consists of two steps, similar to the 2SLS estimator. In the first step, it runs a logit regression of the instrument on the set of controls. p. 0

Where the paper says it

Like the 2SLS estimator, our estimator consists of two steps. In the first step, we run a logit regression of the instrument on the set of controls.

In the second step, it uses a classic instrumental variable estimator for the linear regression of the outcome on the treatment using residuals from the logit regression calculated as an instrument. p. 0
Where the paper says it

In the first step, we run a logit regression of the instrument on the set of controls. In the second step, we use a classic instrumental variable estimator for the linear regression of the outcome on the treatment using residuals from the logit regression calculated Date: February 19, 2026.

The authors also develop an augmented logit-based IV estimator that contains an extra term in the logit regression used in the first step, which originates from a binary regression model of the treatment variable on controls using a subsample of the data corresponding to an ex ante fixed value of the instrument. p. 1
Where the paper says it

This estimator is similar to the logit-based IV estimator itself but contains an extra term in the logit regression used in the first step of the logit-based IV estimator. This term in turn originates from the binary regression model of the treatment variable on controls using a subsample of the data corresponding to an ex ante fixed value of the instrument.

Causal Interpretability and Comparison

The logit-based IV estimator has a causal interpretation if its probability limit takes the form of a weighted-average treatment effect where the weights are non-negative. The 2SLS estimator is not guaranteed to yield such a form, and it may assign negative weights to compliers. p. 0

Where the paper says it

Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability. In particular, in certain cases where the probability limits of both our estimators and the 2SLS estimator take the form of weighted-average treatment effects, our estimators are guaranteed to yield non-negative weights whereas the 2SLS estimator is not.

Under specific conditions, the logit-based IV estimator has a causal interpretation while the 2SLS estimator does not. p. 1
Where the paper says it

Under these conditions, the probability limit of the logit-based IV estimator is represented by a convex combination of treatment effects for compliers and the probability limit of the 2SLS estimator is not. Thus, under these conditions, the logit-based IV estimator has a causal interpretation and the 2SLS estimator does not.

This advantage is particularly important under conditions guaranteeing that the always-taker, never-taker, and non-causal terms vanish. p. 1
Where the paper says it

Without further conditions, complier, always-taker, and never-taker terms take the form of weighted-average treatment effects but the advantage of the logit-based IV estimator is that the corresponding weights in the complier term are always non-negative, which is not necessarily the case for the 2SLS estimator. This advantage is particularly important under additional conditions guaranteeing that the always-taker, never-taker, and non-causal terms vanish.

Asymptotic Properties and Tests

The paper derives asymptotic normality results for both estimators, showing that the logit-based IV estimator converges to a weighted-average treatment effect under certain assumptions. A Hausman specification test is constructed to check some of the conditions underlying causal interpretability. p. 1

Where the paper says it

This term in turn originates from the binary regression model of the treatment variable on controls using a subsample of the data corresponding to an ex ante fixed value of the instrument. Moreover, we construct a Hausman specification test that can be used to check some of the conditions underlying causal interpretability of the logit-based IV estimator.

The authors demonstrate the usefulness of their estimators by applying them to three well-established empirical studies, finding that their estimates are nearly identical to 2SLS estimates in some cases but differ substantially in others where 2SLS may assign negative weights. p. 1
Where the paper says it

We find that our logit-based IV and augmented logit-based IV estimates are nearly identical to the 2SLS estimates reported in Angrist and Evans (1998) and Angrist et al. (2002), but differ substantially from the 2SLS estimates in Dube and Harish (2020).

Empirical Validation and Conclusion

The empirical applications show that the logit-based IV and augmented logit-based IV estimates are nearly identical to the 2SLS estimates reported in Angrist and Evans (1998) and Angrist et al. (2002), but differ substantially from the 2SLS estimates in Dube and Harish (2020). p. 1

Where the paper says it

We find that our logit-based IV and augmented logit-based IV estimates are nearly identical to the 2SLS estimates reported in Angrist and Evans (1998) and Angrist et al. (2002), but differ substantially from the 2SLS estimates in Dube and Harish (2020).

The results indicate that one of the reasons for this discrepancy is that the 2SLS estimator may be assigning negative weights to a significant fraction of compliers, which our estimators avoid. p. 1
Where the paper says it

(2002), but differ substantially from the 2SLS estimates in Dube and Harish (2020). Our results indicate that one of the reasons for this discrepancy in the latter case is that the 2SLS estimator may be assigning negative

The paper concludes by suggesting that in practice, one should run the Hausman test to check if the conditional mean of Z given X has a logit form. p. 1
Where the paper says it

This term in turn originates from the binary regression model of the treatment variable on controls using a subsample of the data corresponding to an ex ante fixed value of the instrument. Moreover, we construct a Hausman specification test that can be used to check some of the conditions underlying causal interpretability of the logit-based IV estimator.

Appendix A. p. 2

Where the paper says it

In Section 6, we present our empirical applications. In the Appendix, we provide all the proofs.

Regularity Conditions

The appendix provides regularity conditions for all theorems and corollaries, including Assumption 2.6 which states that the conditional mean function x 7→ E[ZX = x] takes the logit form. p. 8

Where the paper says it

Given that T(0) is a binary random variable, the conditional mean E[T(0)|X] takes values in the (0, 1) interval, and so this assumption may be less plausible than Assumption 2. However, we will use this assumption mainly to make the comparison between the logit-based IV and 2SLS estimators particularly transparent.

These conditions ensure that both estimators have a causal interpretation under appropriate regularity conditions. p. 6
Where the paper says it

2 are satisfied. Then under appropriate regularity conditions,3 for any s ∈ [0, 1] and a ∈ {Λ, 2SLS}, we have that βba →p βa, where βa = E[∆CP(X)ωCP(X)(sE[Z|X] + (1 − s)ha(X) − ha(X)E[Z|X])] E[T(Z − ha(X))] (6) + sE[∆AT(X)ωAT(X)(E[Z|X] − ha(X))] E[T(Z − ha(X))] (7) + (1 − s)E[∆NT(X)ωNT(X)(ha(X) − E[Z|X])] E[T(Z − ha(X))] (8) + E[E[sY(0) + (1 − s)Y(1)|X](E[Z|X] − ha(X))] E[T(Z − ha(X))] (9) and E[T(Z − ha(X))] = E[ωCP(X)(sE[Z|X] + (1 − s)ha(X) − ha(X)E[Z|X])] + sE[ωAT(X)(E[Z|X] − ha(X))] + (1 − s)E[ωNT(X)(ha(X) − E[Z|X])].

Appendix B. p. 2

Where the paper says it

In Section 6, we present our empirical applications. In the Appendix, we provide all the proofs.

Proofs for Section 2

The proof of Theorem 2.1 shows that under Assumption A. p. 9

Where the paper says it

1 take the following form: βΛ = E[∆CP(X)wΛ(X)] and β2SLS = E[∆CP(X)w2SLS(X)], where wa(x) = ωCP(x)E[Z|X = x](1 − ha(x)) E[ωCP(X)E[Z|X](1 − ha(X))] + E[ωAT(X)(E[Z|X] − ha(X))] (11) for all x ∈ X and a ∈ {Λ, 2SLS}. This corollary shows that without imposing Assumption 2.

p. 9
Where the paper says it

1 take the following form: βΛ = E[∆CP(X)wΛ(X)] and β2SLS = E[∆CP(X)w2SLS(X)], where wa(x) = ωCP(x)E[Z|X = x](1 − ha(x)) E[ωCP(X)E[Z|X](1 − ha(X))] + E[ωAT(X)(E[Z|X] − ha(X))] (11) for all x ∈ X and a ∈ {Λ, 2SLS}. This corollary shows that without imposing Assumption 2.

p. 9
Where the paper says it

1 take the following form: βΛ = E[∆CP(X)wΛ(X)] and β2SLS = E[∆CP(X)w2SLS(X)], where wa(x) = ωCP(x)E[Z|X = x](1 − ha(x)) E[ωCP(X)E[Z|X](1 − ha(X))] + E[ωAT(X)(E[Z|X] − ha(X))] (11) for all x ∈ X and a ∈ {Λ, 2SLS}. This corollary shows that without imposing Assumption 2.

1, we have θb →p θ0 and γb →p γ0, for example, by Theorem 2. p. 6
Where the paper says it

LOGIT-BASED ALTERNATIVES TO TWO-STAGE LEAST SQUARES 7 be the probability limits of the estimators θb and γb appearing in Algorithms 2. Moreover, for all x ∈ X , let hΛ(x) = Λ(x ⊤θ0) and h2SLS(x) = x ⊤γ0.

7 in Newey and McFadden (1994). p. 2
Where the paper says it

We therefore provide here only a few key references that are particularly relevant for our work. The literature has been started by Imbens and Angrist (1994), who gave the local average treatment effect interpretation of the instrumental variable estimator in the case of a binary treatment and a binary instrument without controls allowing for heterogeneous treatment effects under the monotonicity (no defiers) assumption.

p. 2
Where the paper says it

We therefore provide here only a few key references that are particularly relevant for our work. The literature has been started by Imbens and Angrist (1994), who gave the local average treatment effect interpretation of the instrumental variable estimator in the case of a binary treatment and a binary instrument without controls allowing for heterogeneous treatment effects under the monotonicity (no defiers) assumption.

This implies that the probability limits appearing in the numerators of the estimators converge to those stated in Theorem 2. p. 6
Where the paper says it

In fact, the expression for the probability limit of the 2SLS estimator βb 2SLS in this theorem is closely related to that in Proposition 1 of Blandhol et al. 1 shows that without further assumptions, the probability limits of both logit-based IV and 2SLS estimators take the form of a sum of complier, alwaystaker, never-taker, and non-causal terms, appearing in expressions (6), (7), (8), and (9), respectively.

1.

Appendix C. p. 2

Where the paper says it

In Section 6, we present our empirical applications. In the Appendix, we provide all the proofs.

Proofs for Section 3

The proof of Theorem 3.1 shows that under Assumption A.2, we have ψb →p ψ0 by Theorem 2. p. 8

Where the paper says it

For some vector ψ0 ∈ R p , we have E[T(0)|X] = X ⊤ψ0 with probability one. This assumption ensures that the weights in Theorem 2.

7 in Newey and McFadden (1994). This convergence implies that the probability limit of the augmented logit-based IV estimator converges to the stated form. p. 6
Where the paper says it

In fact, the expression for the probability limit of the 2SLS estimator βb 2SLS in this theorem is closely related to that in Proposition 1 of Blandhol et al. 1 shows that without further assumptions, the probability limits of both logit-based IV and 2SLS estimators take the form of a sum of complier, alwaystaker, never-taker, and non-causal terms, appearing in expressions (6), (7), (8), and (9), respectively.

Appendix D. p. 2

Where the paper says it

In Section 6, we present our empirical applications. In the Appendix, we provide all the proofs.

Proofs for Section 4

The proof of Theorem 4. p. 6

Where the paper says it

Then under appropriate regularity conditions,3 for any s ∈ [0, 1] and a ∈ {Λ, 2SLS}, we have that βba →p βa, where βa = E[∆CP(X)ωCP(X)(sE[Z|X] + (1 − s)ha(X) − ha(X)E[Z|X])] E[T(Z − ha(X))] (6) + sE[∆AT(X)ωAT(X)(E[Z|X] − ha(X))] E[T(Z − ha(X))] (7) + (1 − s)E[∆NT(X)ωNT(X)(ha(X) − E[Z|X])] E[T(Z − ha(X))] (8) + E[E[sY(0) + (1 − s)Y(1)|X](E[Z|X] − ha(X))] E[T(Z − ha(X))] (9) and E[T(Z − ha(X))] = E[ωCP(X)(sE[Z|X] + (1 − s)ha(X) − ha(X)E[Z|X])] + sE[ωAT(X)(E[Z|X] − ha(X))] + (1 − s)E[ωNT(X)(ha(X) − E[Z|X])]. (10) The proof of this theorem, as well as of all other results in the main text, is rather simple and is provided in the Appendix.

1 shows that under Assumption A.1, we have √n(βb − βΛ) = n−1/2 ∑n i=1 lΛi E[T(Z − Λ(X⊤θ0)] + o p(1) →d N(0, σ 2Λ), yielding the asserted claim. p. 6
Where the paper says it

Moreover, for all x ∈ X , let hΛ(x) = Λ(x ⊤θ0) and h2SLS(x) = x ⊤γ0. The following theorem derives the probability limits of both βb Λ and βb 2SLS.

The proof of Theorem 4.2 shows that under Assumption A. p. 9
Where the paper says it

1 take the following form: βΛ = E[∆CP(X)wΛ(X)] and β2SLS = E[∆CP(X)w2SLS(X)], where wa(x) = ωCP(x)E[Z|X = x](1 − ha(x)) E[ωCP(X)E[Z|X](1 − ha(X))] + E[ωAT(X)(E[Z|X] − ha(X))] (11) for all x ∈ X and a ∈ {Λ, 2SLS}. This corollary shows that without imposing Assumption 2.

3, the asymptotic distribution of βbAΛ is given by a normal distribution. p. 4
Where the paper says it

Let (Y1 , T1 , X1 , Z1 ), . , (Yn, Tn, Xn, Zn) be a random sample from the distribution of (Y, T, X, Z).

Appendix E. p. 2

Where the paper says it

In Section 6, we present our empirical applications. In the Appendix, we provide all the proofs.

Auxiliary Lemma

Lemma E.1 states that under Assumption 2. p. 2

Where the paper says it

We therefore provide here only a few key references that are particularly relevant for our work. The literature has been started by Imbens and Angrist (1994), who gave the local average treatment effect interpretation of the instrumental variable estimator in the case of a binary treatment and a binary instrument without controls allowing for heterogeneous treatment effects under the monotonicity (no defiers) assumption.

2 and that E[Y] < ∞, then E[(Y(1) − Y(0))(T(1) − T(0))X] = ∆CP(X)ωCP(X), etc. p. 5
Where the paper says it

To see why this is so, we first need to introduce some additional notation. For all x ∈ X , let ωCP(x) = P(G = CP|X = x), ωAT(x) = P(G = AT|X = x), ωNT(x) = P(G = NT|X = x) denote the x-conditional fractions of compliers, always-takers, and never-takers in the population, respectively.

, which are key for the proofs of the theorems. p. 2
Where the paper says it

In Section 6, we present our empirical applications. In the Appendix, we provide all the proofs.

References

The paper cites numerous relevant literature, including work by Angrist and Evans (1998), Imbens and Angrist (1995), and Dube and Harish (2020). The authors also reference foundational econometric works such as Hausman (1978) and Ramsey (1969). p. 1

Where the paper says it

This term in turn originates from the binary regression model of the treatment variable on controls using a subsample of the data corresponding to an ex ante fixed value of the instrument. Moreover, we construct a Hausman specification test that can be used to check some of the conditions underlying causal interpretability of the logit-based IV estimator.

The paper includes references to machine learning methods, such as Belloni et al. p. 4
Where the paper says it

Belloni et al. (2017), developed nonparametric and machine learning estimators of the LATE in the model with controls.

(2017) and Chernozhukov et al. p. 4
Where the paper says it

Belloni et al. (2017), developed nonparametric and machine learning estimators of the LATE in the model with controls.

(2018). The references provide a strong foundation for the work on causal inference in this setting. p. 2
Where the paper says it

Our paper contributes to the large literature discussing causal interpretability of various parametric IV estimators in the case of heterogeneous treatment effects. We therefore provide here only a few key references that are particularly relevant for our work.

Tables

Table 1 compares the estimators for Angrist and Evans (1998), Table 2 compares the estimators for Angrist et al. (2002), and Table 3 compares the estimators for Dube and Harish (2020). p. 1

Where the paper says it

We demonstrate the usefulness of our estimators by applying them to three wellestablished empirical studies: Angrist and Evans (1998), Angrist et al. (2002), and Dube and Harish (2020).

These tables provide empirical evidence comparing the performance of the different estimation methods across different studies. p. 6
Where the paper says it

1 is not unique as different values of the parameter s give different representations. For example, it is always possible to get rid of the never-taker term by substituting 3To avoid distractions, we provide the list of regularity conditions for this theorem and all other results in the Appendix.

The fractions in these tables show where the first-stage predicted values exceed one or fall below zero for the 2SLS estimator. p. 7
Where the paper says it

Second, the always-taker term in (7) shows that the xconditional average treatment effect for always-takers ∆AT(x) has the weight ωAT(x)(E[Z|X = x] − ha(x)) E[T(Z − ha(x))] , which may be negative for both logit-based IV and 2SLS estimators, and the same applies to the never-taker term in (8). In addition, although the weights in the causal terms (6), (7), and (8) do integrate to one, as we show below (Corollary 2.

Summary

The logit-based IV and augmented logit-based IV estimators serve as alternatives to the traditionally used 2SLS estimator in settings with binary treatment and binary instrument. They offer an advantage in causal interpretability because their probability limits can be represented as convex combinations of treatment effects with non-negative weights, unlike the 2SLS estimator. p. 0

Where the paper says it

Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability. In particular, in certain cases where the probability limits of both our estimators and the 2SLS estimator take the form of weighted-average treatment effects, our estimators are guaranteed to yield non-negative weights whereas the 2SLS estimator is not.

p. 0
Where the paper says it

Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability. In particular, in certain cases where the probability limits of both our estimators and the 2SLS estimator take the form of weighted-average treatment effects, our estimators are guaranteed to yield non-negative weights whereas the 2SLS estimator is not.

The paper demonstrates that these estimators are robust and provide a causal interpretation under specific testable conditions, making them preferable in settings where the 2SLS estimator may assign negative weights. p. 2
Where the paper says it

Our estimators avoid this issue and are therefore preferable in settings like that in Dube and Harish (2020). Our paper contributes to the large literature discussing causal interpretability of various parametric IV estimators in the case of heterogeneous treatment effects.

p. 2
Where the paper says it

Our estimators avoid this issue and are therefore preferable in settings like that in Dube and Harish (2020). Our paper contributes to the large literature discussing causal interpretability of various parametric IV estimators in the case of heterogeneous treatment effects.

Pruning redundant reasoning tokens

The logit-based IV and augmented logit-based IV estimators serve as alternatives to the traditionally used 2SLS estimator in settings with binary treatment and binary instrument. They offer an advantage in causal interpretability because their probability limits can be represented as convex combinations of treatment effects with non-negative weights, unlike the 2SLS estimator. The paper demonstrates that these estimators are robust and provide a causal interpretation under specific testable conditions, making them preferable in settings where the 2SLS estimator may assign negative weights.

How the training signal is built

The logit-based IV estimator is constructed by running a logit regression of the instrument on controls to obtain residuals, which are then used as an instrument in a subsequent instrumental variable regression. The augmented estimator involves an extra covariate in the logit regression of Z on X. p. 1

Where the paper says it

In addition, we develop an augmented logit-based IV estimator that has a causal interpretation under conditions that are more plausible than those underlying causal interpretability of the logit-based IV and 2SLS estimators. This estimator is similar to the logit-based IV estimator itself but contains an extra term in the logit regression used in the first step of the logit-based IV estimator.

The construction of the estimators is detailed in Algorithms 2. p. 6
Where the paper says it

LOGIT-BASED ALTERNATIVES TO TWO-STAGE LEAST SQUARES 7 be the probability limits of the estimators θb and γb appearing in Algorithms 2. Moreover, for all x ∈ X , let hΛ(x) = Λ(x ⊤θ0) and h2SLS(x) = x ⊤γ0.

1 and 4.1.

Where the gains come from

The main advantage of the logit-based IV estimator is that its probability limit can be written as a convex combination of treatment effects with non-negative weights when Assumption 2.4 is satisfied. p. 6

Where the paper says it

Suppose that Assumptions 2. 2 are satisfied.

This ensures that the logit-based IV estimator has a causal interpretation, whereas the 2SLS estimator may have negative weights. p. 9
Where the paper says it

At the same time, the weights of the logit-based IV estimator are always non-negative, as the logit function Λ(·) takes values in the (0, 1) interval. Thus, under our assumptions, the logit-based IV estimator has a causal interpretation and the 2SLS estimator does not.

The augmented estimator improves upon the original by incorporating an extra term in the logit regression of Z on X. p. 1
Where the paper says it

In addition, we develop an augmented logit-based IV estimator that has a causal interpretation under conditions that are more plausible than those underlying causal interpretability of the logit-based IV and 2SLS estimators. This estimator is similar to the logit-based IV estimator itself but contains an extra term in the logit regression used in the first step of the logit-based IV estimator.

The augmented logit-based IV estimator has a causal interpretation even if Assumption 2.4 is not satisfied, as long as the correct link function Φ is being used. p. 5

Where the paper says it

In particular, it is easy to check that the logit-based IV estimator is consistent for ρ under Assumption 2. 1 as long as the function x 7→ E[Y(0)|X = x] is linear.

This suggests that the augmented estimator improves upon the 2SLS estimator in terms of causal interpretability. p. 0
Where the paper says it

We propose logit-based IV and augmented logit-based IV estimators that serve as alternatives to the traditionally used 2SLS estimator in the model where both the endogenous treatment variable and the corresponding instrument are binary. Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability.

The paper recommends running the Hausman test to check if the conditional mean of Z given X has a logit form. p. 1
Where the paper says it

This term in turn originates from the binary regression model of the treatment variable on controls using a subsample of the data corresponding to an ex ante fixed value of the instrument. Moreover, we construct a Hausman specification test that can be used to check some of the conditions underlying causal interpretability of the logit-based IV estimator.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved causal inference for binary treatment variable settings: The logit-based IV estimator guarantees that the corresponding weights in the complier term are always non-negative whereas the 2SLS estimator is not, which is crucial when treatment effects are heterogeneous, preventing issues where the 2SLS estimator may be assigning negative weights to a significant fraction of compliers.

  2. Enhanced robustness against estimation bias: The augmented logit-based IV estimator can be used when Assumption 3.1 holds, ensuring the estimator has a causal interpretation even when the more general Assumption 2.4 is not satisfied, providing a pathway to causality under more plausible assumptions.

  3. Model selection and identification: A Hausman specification test can be constructed by comparing the logit-based IV and augmented logit-based IV estimators; if this test does not reject the null hypothesis of Assumption 2.6 being satisfied, it confirms that both estimators estimate the weighted-average treatment effect for compliers (call it W-LATE) in (16) with weights given by (17).

  4. Adaptive estimation strategy: For empirical applications where the 2SLS estimator lacks a causal interpretation, the system should switch to the logit-based IV estimator, as it is preferable in settings like that in Dube and Harish (2020) because it avoids assigning negative weights.

  5. Asymptotic distribution modeling: The asymptotic variance formulas for both estimators can be computed using a plug-in method, allowing for the estimation of the asymptotic variances sigma 2 Λ and sigma 2 AΛ, which is vital for constructing confidence intervals around the causal estimates.

Abstract

We propose logit-based IV and augmented logit-based IV estimators that serve as alternatives to the traditionally used 2SLS estimator in the model where both the endogenous treatment variable and the corresponding instrument are binary. Our novel estimators are as easy to compute as the 2SLS estimator but have an advantage over the 2SLS estimator in terms of causal interpretability. In particular, in certain cases where the probability limits of both our estimators and the 2SLS estimator take the form of weighted-average treatment effects, our estimators are guaranteed to yield non-negative weights whereas the 2SLS estimator is not.

Related papers