Pith. sign in

REVIEW 5 major objections 5 minor 57 references

Correlated Synthetic Controls

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proposes a correlated synthetic control estimator for panels with many treated units and shows that when treatment assignment is correlated with unobservables it can deliver smaller ATT estimation error than…

desk verdict The CSC estimator is a genuinely new idea with an honest but incomplete theoretical case; the claimed dominance over DiD does not follow from the proofs, but the estimator and simulation deserve engagement. read the letter →

arxiv 2507.08918 v1 pith:CWPV3O2U submitted 2025-07-11 econ.EM

Tzvetan Moev This is my paper
classification econ.EM MSC 62P2091B82
keywords syntheticcontrolcorrelatedrandomcoefficientsmanytreatedunitsaveragetreatmenteffectonthedifference-in-differencesinteractivefixedeffectsMarielBoatliftpaneldata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes Correlated Synthetic Controls (CSC), an estimator for panels with many treated units and short pre-treatment periods. CSC builds a synthetic counterfactual for every treated individual from donor outcomes, with donor weights that follow a correlated random-coefficient model, so treated individuals with similar observables receive similar synthetic controls. The core theoretical claim is that when treatment assignment is correlated with unobserved factor loadings and the data follow an interactive fixed-effects model, the CSC estimate of the average treatment effect on the treated has smaller estimation error than the difference-in-differences estimate. A simulation study and a PSID-based reanalysis of the Mariel Boatlift support the ranking and illustrate how the estimator yields heterogeneous treatment effects.

What carries the argument

The central object is the correlated random-coefficient weight model $w_{ij}=\omega_j+x_i\alpha_j$, with the usual synthetic-control constraints that weights are non-negative and sum to one; it is estimated by a constrained quadratic program. This makes synthetic controls for treated units with similar covariates share donor weights, avoiding both the multiplicity of solutions and the overfitting of separate synthetic controls. The theoretical machinery is Abadie's representation (Lemma 1): under the Exact Fit assumption, the estimation error can be rewritten using only pre-treatment idiosyncratic errors and the common factors $\lambda_t$, eliminating the unobserved $\mu_i$; sub-Gaussian concentration inequalities then give the high-probability upper bound in Proposition 1.

What would settle it

Run the paper's simulation with an identical interactive fixed-effects DGP and non-random treatment assignment, but add small pre-treatment noise that makes exact fit impossible; if CSC's average ATT error then fails to beat feasible DiD across replications, the claimed dominance under selection on unobservables is refuted for realistic settings.

Watch

Extended reading notes

Core claim

Under an interactive fixed-effects data-generating process in which treatment assignment $D_{it}$ is correlated with the unobserved loadings $\mu_i$, the Correlated Synthetic Controls estimator's ATT error $|\hat{\tau}_{CSC}-\tau|$ admits an upper bound that does not depend on the covariance between treatment and unobservables, while the difference-in-differences estimator's asymptotic error is proportional to $(\bar{\lambda}_{pre}-\lambda_T)(\bar{\mu}_{don}-\mathbb{E}[\mu_i|D_{it}=1])/n_0$. The paper reads this as a dominance result: in the selection-on-unobservables cases that motivate many microeconometric applications, CSC is the safer default than DiD, even though its own bound is only an upper bound and requires an exact pre-treatment fit.

Load-bearing premise

The result collapses if the estimated weights do not exactly reproduce each treated unit's pre-treatment outcomes and covariates, a condition the author notes is unlikely to hold in real applications.

Editorial extensions

If this is right

  • When treatment assignment is correlated with unobserved factor loadings, the paper's bound implies that CSC should be preferred to DiD even though DiD is the default in many-treated-unit panels.
  • Because CSC pools weights across treated units with similar observables, it inherits the usual SC benefit of excluding bad donors and can estimate individual treatment effects, so heterogeneous effects do not require running separate DiD designs.
  • The simulation shows that CSC approaches the infeasible oracle DiD when treatment is not randomly assigned, and that random assignment is the regime where feasible DiD outperforms CSC and the dominance logic reverses.
  • In the Mariel Boatlift application, CSC produces slightly better pre-treatment predictions than the penalised synthetic control estimator and finds a negative wage effect for low-skilled workers, with no detectable effect for high-skilled workers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The dominance claim is proven only under Exact Fit, an assumption the paper itself says is unlikely in applications; a natural next step is to replace it by an approximate-fit condition and check whether a meaningful gap over DiD survives.
  • Because the estimator currently requires discrete, time-invariant covariates and a single treatment date, its practical scope is narrower than the comparison suggests; extending it to continuous covariates and staggered adoption would test how far the theoretical result generalises.
  • The empirical application uses Florida rather than Miami identifiers in the PSID, so the reported low-skilled wage effect is a state-level average rather than a Miami effect; the paper implies the Miami effect would be larger, which a researcher could test with finer geographic data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes Correlated Synthetic Controls (CSC), a synthetic-control estimator for panels with many treated units and relatively short pre-treatment periods. The estimator lets donor weights depend linearly on treated-unit covariates, nesting both pooled and unit-specific synthetic controls. Under an interactive fixed effects model with treatment correlated with unobservables, the paper derives an upper bound on CSC's estimation error (Proposition 1) and an asymptotic bias expression for difference-in-differences (Proposition 2), and argues that CSC should be preferred to DiD in this setting. The paper also reports simulations comparing CSC, penalized synthetic control (PSC), feasible DiD, and an infeasible DiD benchmark, and applies CSC to the Mariel Boatlift using PSID data to estimate heterogeneous treatment effects.

Significance. If the theoretical comparison were established, the paper would offer a useful and non-obvious selection criterion for synthetic-control-type estimators versus DiD under selection on unobservables. The estimator is clearly motivated, the correlated-random-coefficients formulation is a sensible compromise between pooled and unit-specific synthetic controls, and the simulation design with an infeasible DiD benchmark is informative. The PSID application illustrates a credible use of the method, and the manuscript provides reproducible code. However, the central dominance claim is not currently proven: the paper compares a finite-sample, high-probability upper bound for CSC with an asymptotic bias expression for DiD, without any formal condition under which the former is smaller than the latter. The theoretical claim also rests on an Exact Fit assumption that the author explicitly concedes is unlikely to hold in applications.

major comments (5)
  1. [Section 3.f, Eq. (8)] The optimization problem defining the CSC weights is stated as a maximization of a sum of squared residuals. Since the objective is nonnegative and is intended to measure lack of fit, the formulation as written does not define a least-squares-type estimator; the surrounding text and simulations clearly intend minimization. The same issue appears in Appendix A.b, Eq. (23). Please correct the objective to minimization and re-state the constraints consistently.
  2. [Section 4.d, Eqs. (12) and (13)] The conclusion that CSC 'dominates' DiD does not follow from the stated results. Proposition 1 is a finite-sample high-probability upper bound on |τhat_CSC − τ|, while Proposition 2 is the probability limit of |τhat_DiD − τ| as n1 → ∞. Establishing dominance requires a theorem or explicit parameter condition showing that the right-hand side of (12) is smaller than the right-hand side of (13) for the same n0, n1, T0, F, λ, σ, and probability level. No such comparison is derived, and independence of the CSC bound from Cov(Dit, μi) does not imply that the bound is smaller than the DiD bias. This is the load-bearing step for the abstract's and conclusion's main claim.
  3. [Section 4.b, Assumption 2] Lemma 1 and Proposition 1 both rely on Exact Fit to eliminate the unobserved factor loadings μi from the estimation error. The author states in Section 4.b that Exact Fit 'in applications is unlikely to hold.' Because the bound collapses if Exact Fit fails, the theoretical result is conditional on a condition that is not credible in the target setting. The paper needs either an approximate-fit version with a formal bound on the additional error, or a clear statement that the theoretical comparison applies only under Exact Fit and that the practical case is supported only by simulation.
  4. [Appendix B.c, Claims 2 and 4] The proof of Proposition 1 applies concentration inequalities to quantities involving the estimated weights ŵ. For the pre-treatment term in Claim 2, ŵ depends on pre-treatment errors, so the vector to which the Hanson-Wright-type inequality is applied contains dependent products ŵ_ij ϵ_j,pre; the proof does not formally account for this dependence. Claim 4 similarly treats z_i = ϵ_is − Σ_j ŵ_ij ϵ_js as sub-Gaussian without conditioning on the estimated weights. Conditioning on the weights makes the post-treatment term manageable, but the pre-treatment term remains a gap. As written, the probability statement in (12) is not fully established.
  5. [Section 5.d, Table 1] The simulation is presented as confirming the theoretical result, but Table 1 reports average signed estimation error, (1/1000)Σ(τhat_r − τ), whereas Proposition 1 is a high-probability bound on the absolute value |τhat − τ|. Signed averages can be small because positive and negative errors cancel. Table 5 reports RMSE, which is closer in spirit, but the simulations do not compare the empirical CSC error to the analytical bound in (12) or to the DiD bias expression in (13). The simulation is informative descriptively but does not close the logical gap in the dominance claim.
minor comments (5)
  1. [Throughout] There are numerous typos and misspellings, including 'causal lint' (Section 2), 'obsevations' (Section 3.d), 'martial status' (Example 1), 'relatiely' and 'week correlation' (Section 5.d), and 'post-treatmnet' (Section 4.c). The manuscript would benefit from a careful proofreading pass.
  2. [Proposition 1, notation] The notation for λmin and λmax is confusing: the proposition text says they are the minimum and maximum common factor entries in absolute value, while the proof at the end of Appendix B.c says 'λmin ≡ λtilde and λmax ≡ λtilde,' which suggests a typo in the definitions or in the proof.
  3. [Equation (13)] Equation (13) has a stray comma after E[μi|Dit = 1] and the surrounding sentence is grammatically incomplete. Please clean up the display and the prose.
  4. [Section 5.d, Table 1 footnote] The footnote in Table 1 acknowledges that the analysis uses average estimation error rather than absolute error; this is a useful admission, but the same comment should be integrated into the main text where the table is discussed, because the table is otherwise easy to misread as evidence for the absolute-error bound.
  5. [References and dates] Several references are dated as 'very recent' or 'last month' (e.g., the Imbens Sargan lecture, Ben-Michael et al. 2021). These should be updated for a journal submission, and the reference list should be checked for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main comparison derives from the assumed DGP rather than from the paper's own conclusions.

full rationale

The paper's central comparison is not circular. Proposition 1's CSC error bound is derived from the interactive fixed-effects DGP, Assumption 2 (Exact Fit), sub-Gaussian errors, and a concentration argument; it does not use the relative performance of CSC versus DiD as an input. Proposition 2's DiD bias expression is obtained by Frisch-Waugh-Lovell algebra from the same DGP, not from the paper's conclusion. The claimed dominance rests on the observation that the CSC bound is independent of Cov(Dit, mu_i) while the DiD bias depends on it; whether the bound is actually smaller in relevant parameter regions is a logical gap in the comparison, not a circular reduction. The simulation generates data from a known DGP and does not fit constants to the theory. The paper cites prior SC literature for Exact Fit and Abadie's representation, but these are external, stated assumptions rather than self-citations carrying the conclusion, and the author explicitly notes that Exact Fit is unlikely to hold in applications, which weakens applicability but does not make the derivation circular.

Assumptions & free parameters 6 free parameters · 8 assumptions · 0 invented entities

The central theoretical result depends on a sequence of domain assumptions (interactive fixed effects, Exact Fit, subGaussian errors, invertibility, discrete covariates) and on simulation parameters chosen by hand. No new entities are postulated. The heaviest burden is Exact Fit, which the author acknowledges is unlikely to hold in applications.

free parameters (6)
  • Simulation common factor matrix Lambda = 8 x 4 matrix drawn once from N(3,2), listed in Appendix A.f (26)
    Chosen by hand and held fixed across simulations so estimator comparisons are not affected by redraws; the simulation conclusions depend on this choice.
  • Simulation Gamma matrix = 6 x 4 matrix in Appendix A.f (27)
    Controls correlation between covariates and unobserved factor loadings; fixed across simulations.
  • Simulation Phi vector = (-1.12, -0.46, 3.12, 0.14)'
    Controls correlation between treatment assignment and unobservables in the logistic selection model; its nonzero value encodes the key selection-on-unobservables scenario.
  • Simulation Beta vector = (1, 0.4, 0.6, 0.8, 1, 1.2)
    Coefficients of covariates in the outcome DGP; fixed by hand.
  • Idiosyncratic shock variance sigma^2 = 1
    The DGP draws epsilon_iid N(0,1); the theoretical bound scales linearly in sigma, so this is a normalization.
  • Training period length for HTE figures = Ttrain=1 for wages, Ttrain=4 for labor supply
    Chosen based on lowest cross-validated RMSE in Table 3, not pre-specified; affects the empirical estimates reported in Section 6.d.
assumptions (8)
  • domain assumption Interactive fixed effects DGP: y_it = theta_t x_i' + D_it tau + lambda_t mu_i + epsilon_it (Section 4.a, Eq. 10-11).
    The paper assumes this DGP for the theoretical comparison; if outcomes are generated differently, the derived bounds and DiD bias expressions need not apply.
  • domain assumption Assumption 1 DGP restrictions: treatment not random and correlated with mu_i, errors iid, xi independent of D and epsilon but correlated with mu, mu stochastic with mean mu, lambda fixed, one post-treatment period.
    These restrictions define the selection-on-unobservables environment and simplify the algebra; the one-post-period assumption avoids dynamic ATT issues.
  • domain assumption Assumption 2 Exact Fit: there exists a unique weight matrix W with wij = omega_j + Sum_k alpha_kj x_i(k), columns summing to one and nonnegative, satisfying y_it = Sum_j wij y_jt for all pre-treatment t and x_i = Sum_j wij x_j for all covariates.
    This is the load-bearing premise for Lemma 1 and Proposition 1. The author acknowledges it is unlikely to hold in applications and that an approximate version is left to future work.
  • domain assumption Assumption 3 subGaussian errors epsilon with parameter sigma^2.
    Needed for the Hanson-Wright and Chernoff tail bounds that produce the CSC estimation error bound.
  • standard math Invertibility of lambda'_pre lambda_pre and T0 > F.
    Necessary for the Abadie representation and to recover loadings from pre-treatment factors.
  • domain assumption CSC feasibility requires discrete covariates only, or continuous covariates recoded into categories (Section 3.f, Eq. 9).
    The sum-to-one constraints force sums of alpha coefficients to match across categories; continuous covariates generally make the optimization infeasible. This restricts the estimator's applicability.
  • domain assumption For the empirical application, Florida state residence is used as a proxy for Miami treatment exposure because PSID lacks metro-area identifiers.
    The paper argues spillovers within Florida make this acceptable, but treatment effects may be diluted relative to a Miami-only analysis.
  • standard math For the DiD bias result, n1 goes to infinity with n0 fixed.
    The asymptotic expression in Proposition 2 only holds in this sequence; finite-sample DiD bias may differ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Correlated Synthetic Controls." pith.science (2026). https://pith.science/paper/CWPV3O2U

@misc{pith2026250708918,
  author       = {Pith},
  title        = {Pith review of: Correlated Synthetic Controls},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWPV3O2U}},
  note         = {Machine review of arXiv:2507.08918}
}
read the original abstract

Synthetic Control methods have recently gained considerable attention in applications with only one treated unit. Their popularity is partly based on the key insight that we can predict good synthetic counterfactuals for our treated unit. However, this insight of predicting counterfactuals is generalisable to microeconometric settings where we often observe many treated units. We propose the Correlated Synthetic Controls (CSC) estimator for such situations: intuitively, it creates synthetic controls that are correlated across individuals with similar observables. When treatment assignment is correlated with unobservables, we show that the CSC estimator has more desirable theoretical properties than the difference-in-differences estimator. We also utilise CSC in practice to obtain heterogeneous treatment effects in the well-known Mariel Boatlift study, leveraging additional information from the PSID.

Figures

Figures reproduced from arXiv: 2507.08918 by the authors.

Figure 1
Figure 1. The effect of immigration in the classical labour market model with inelastic labour supply and flexible wages. Note: We have inelastic LS, due to perfect competition between firms: if a firm decreases its wages, then all of its workers can leave and immediately find a job, meaning that wage elasticity of labour supply is infinite (Manning, 2006). home and arrived in the US before the policy was reversed in late Sep… view at source ↗
Figure 2
Figure 2. 95% CI for Mariel Boatlift’s effect on low-skilled and high-skilled workers. [PITH_FULL_IMAGE:figures/full_fig_p044_2.png] view at source ↗
Figure 3
Figure 3. Treatment Effects when Ttrain = 2 68 [PITH_FULL_IMAGE:figures/full_fig_p069_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Treatment Effects when Ttrain = 3 −1.5 −1.0 −0.5 0.0 1980 1981 1982 1983 1984 ATT a) Wage, Low−Skilled −1.5 −1.0 −0.5 0.0 1980 1981 1982 1983 1984 ATT b) Wage, High−Skilled −1500 −1000 −500 0 500 1980 1981 1982 1983 ATT c) LS, Low−Skilled −1500 −1000 −500 0 500 1980 19…
Figure 5
Figure 5. Figure 5: Treatment Effects when Ttrain = 4 for wages and Ttrain = 1 for LS 69 [PITH_FULL_IMAGE:figures/full_fig_p070_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 47 canonical work pages

  1. [1]

    A. Abadie. Using synthetic controls: Feasibility, data requirements, and methodological aspects. Forthcoming in Journal of Economic Literature, 2020 a . URL https://economics.mit.edu/files/17847

    work page 2020

  2. [2]

    A. Abadie. Discussion session. In Chamberlain Seminar. Symposium on Synthetic Control, 2020 b . URL https://www.youtube.com/watch?v=Wta8N6FaM4Y&t=5205s&ab_channel=ChamberlainSeminar

    work page 2020

  3. [3]

    Abadie and J

    A. Abadie and J. L'Hour. A penalized synthetic control estimator for disaggregated data. MIT Working Paper, 2020. URL http://economics.mit.edu/files/18642

    work page 2020

  4. [4]

    Abadie, A

    A. Abadie, A. Diamond, and J. Hainmueller. Synthetic control methods for comparative case studies: Estimating the effect of california’s tobacco control program. Journal of the American Statistical Association, 105 0 (490): 0 493--505, 2010

    work page 2010

  5. [5]

    Abadie, A

    A. Abadie, A. Diamond, and J. Hainmueller. Comparative politics and the synthetic control method. American Journal of Political Science, 59 0 (2): 0 495--510, 2015. doi:https://doi.org/10.1111/ajps.12116. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/ajps.12116

    doi:10.1111/ajps.12116 2015

  6. [6]

    Acemoglu and D

    D. Acemoglu and D. Autor. Skills, tasks and technologies: Implications for employment and earnings. In D. Card and O. Ashenfelter, editors, Handbook of Labor Economics, volume 4b, page 1043–1171. Elsevier, 2011

    work page 2011

  7. [7]

    Arkhangelsky, S

    D. Arkhangelsky, S. Athey, D. A. Hirshberg, G. W. Imbens, and S. Wager. Synthetic difference in differences. ArXiv preprint No. 1812.09970, 2020. URL https://arxiv.org/abs/1812.09970

    arXiv 2020

  8. [8]

    Ensemble Methods for Causal Effects in Panel Data Settings

    S. Athey, M. Bayati, G. Imbens, and Z. Qu. Ensemble methods for causal effects in panel data settings. AEA Papers and Proceedings, 109: 0 65--70, 2019. URL https://arxiv.org/abs/1903.10079

    work page Pith review arXiv 2019

Show all 57 references
  1. [9]

    Athey, M

    S. Athey, M. Bayati, N. Doudchenko, G. Imbens, and K. Khosravi. Matrix completion methods for causal panel data models. ArXiv preprint No. 1710.10251, 2021. URL https://arxiv.org/abs/1710.10251

    2021 arXiv

  2. [10]

    J. Bai. Panel data models with interactive fixed effects. Econometrica, 77 0 (4): 0 1229--1279, 2009. doi:https://doi.org/10.3982/ECTA6135. URL https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA6135

    2009 doi

  3. [11]

    Balestra and J

    P. Balestra and J. Krishnakumar. Fixed effects models and fixed coefficients models. In L. M\'aty\'as and P. Sevestre, editors, The Econometrics of Panel Data, chapter 2. Springer, 3rd edition, 2008

    2008

  4. [12]

    Ben-Michael, A

    E. Ben-Michael, A. Feller, and J. Rothstein. The augmented synthetic control method. ArXiv preprint No. 1811.04170, 2020. URL https://arxiv.org/abs/1811.04170

    2020 arXiv

  5. [13]

    Ben-Michael, A

    E. Ben-Michael, A. Feller, and J. Rothstein. Synthetic controls with staggered adoption. ArXiv preprint No. 1912.03290, 2021. URL https://arxiv.org/abs/1912.03290

    1912 arXiv

  6. [14]

    D. S. Bernstein. Matrix Mathematics. Princeton University Press, 2009. doi:doi:10.1515/9781400833344. URL https://doi.org/10.1515/9781400833344

    2009 doi

  7. [15]

    Bonhomme and E

    S. Bonhomme and E. Manresa. Grouped patterns of heterogeneity in panel data. Econometrica, 83 0 (3): 0 1147--1184, 2015. URL http://www.jstor.org/stable/43616962

    2015

  8. [16]

    G. J. Borjas. The wage impact of the marielitos: A reappraisal. ILR Review, 70 0 (5): 0 1077--1110, 2017. doi:10.1177/0019793917692945. URL https://doi.org/10.1177/0019793917692945

    2017 doi

  9. [17]

    B. Born, G. J. Müller, M. Schularick, and P. Sedláček. The Costs of Economic Nationalism: Evidence from the Brexit Experiment . Technical report, Centre for Macroeconomics Working paper, 2017. URL http://www.centreformacroeconomics.ac.uk/Discussion-Papers/2017/CFMDP2017-38-Paper.pdf

    2017

  10. [18]

    B. Born, G. J. Müller, M. Schularick, and P. Sedláček. The Costs of Economic Nationalism: Evidence from the Brexit Experiment . The Economic Journal, 129 0 (623): 0 2722--2744, 05 2019. doi:10.1093/ej/uez020. URL https://doi.org/10.1093/ej/uez020

    2019 doi

  11. [19]

    Botosaru and B

    I. Botosaru and B. Ferman. On the role of covariates in the synthetic control method . The Econometrics Journal, 22 0 (2): 0 117--130, 01 2019. doi:10.1093/ectj/utz001. URL https://doi.org/10.1093/ectj/utz001

    2019 doi

  12. [20]

    D. Card. The impact of the mariel boatlift on the miami labor market. ILR Review, 43 0 (2): 0 245--257, 1990. doi:10.1177/001979399004300205. URL https://doi.org/10.1177/001979399004300205

    1990 doi

  13. [21]

    Card and D

    D. Card and D. R. Hyslop. Female earnings inequality: The changing role of family characteristics on the extensive and intensive margins. Technical report, NBER Working Paper No. 25387, 2018. URL https://www.nber.org/system/files/working_papers/w25387/w25387.pdf

    2018

  14. [22]

    Carlin and D

    W. Carlin and D. W. Soskice. Macroeconomics: Institutions, instability, and the financial system. Oxford University Press, USA, 2014

    2014

  15. [23]

    Chamberlain

    G. Chamberlain. Multivariate regression models for panel data. Journal of econometrics, 18 0 (1): 0 5--46, 1982

    1982

  16. [24]

    Chernozhukov, K

    V. Chernozhukov, K. Wuthrich, and Y. Zhu. An exact and robust conformal inference method for counterfactual and synthetic controls. ArXiv preprint No. 1712.09089, 2020. URL https://arxiv.org/abs/1712.09089

    2020 arXiv

  17. [25]

    M. A. Clemens and J. Hunt. The labor market effects of refugee waves: Reconciling conflicting results. ILR Review, 72 0 (4): 0 818--857, 2019. doi:10.1177/0019793918824597. URL https://doi.org/10.1177/0019793918824597

    2019 doi

  18. [26]

    Cr\'epon and J

    B. Cr\'epon and J. Mairesse. The chamberlain approach to panel data: An overview and some simulations. In L. M\'aty\'as and P. Sevestre, editors, The Econometrics of Panel Data, chapter 5. Springer, 3rd edition, 2008

    2008

  19. [27]

    D. R. Davis and D. E. Weinstein. Bones, bombs, and break points: the geography of economic activity. American economic review, 92 0 (5): 0 1269--1289, 2002

    2002

  20. [28]

    De Chaisemartin and X

    C. De Chaisemartin and X. d'Haultfoeuille. Two-way fixed effects estimators with heterogeneous treatment effects. American Economic Review, 110 0 (9): 0 2964--96, 2020

    2020

  21. [29]

    Doudchenko and G

    N. Doudchenko and G. Imbens. Balancing, regression, difference-in-differences and synthetic control methods: A synthesis. ArXiv preprint No. 1610.07748, 2018. URL https://arxiv.org/abs/1610.07748

    2018 arXiv

  22. [30]

    Dustmann, U

    C. Dustmann, U. Schönberg, and J. Stuhler. The impact of immigration: Why do studies reach such different results? Journal of Economic Perspectives, 30 0 (4): 0 31--56, November 2016. doi:10.1257/jep.30.4.31. URL https://www.aeaweb.org/articles?id=10.1257/jep.30.4.31

    2016 doi

  23. [31]

    B. Ferman. On the properties of the synthetic control estimator with many periods and many controls. ArXiv preprint No. 1906.06665, 2020. URL https://arxiv.org/pdf/1906.06665.pdf

    1906 arXiv

  24. [32]

    Ferman and C

    B. Ferman and C. Pinto. Synthetic controls with imperfect pre-treatment fit. ArXiv preprint No. 1911.08521, 2021. URL https://arxiv.org/pdf/1911.08521.pdf

    1911 arXiv

  25. [33]

    Ferman, C

    B. Ferman, C. Pinto, and V. Possebom. Cherry picking with synthetic controls. Journal of Policy Analysis and Management, 39 0 (2): 0 510--532, 2020. doi:https://doi.org/10.1002/pam.22206. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/pam.22206

    2020 doi

  26. [34]

    A. Fu, B. Narasimhan, and S. Boyd. Cvxr: An r package for disciplined convex optimization. Journal of Statistical Software, 2019. URL https://web.stanford.edu/ boyd/papers/cvxr_paper.html

    2019

  27. [35]

    D. L. Hanson and F. T. Wright. A Bound on Tail Probabilities for Quadratic Forms in Independent Random Variables . The Annals of Mathematical Statistics, 42 0 (3): 0 1079 -- 1083, 1971. doi:10.1214/aoms/1177693335. URL https://doi.org/10.1214/aoms/1177693335

    1971

  28. [36]

    M. Hlavac. Stargazer: beautiful latex, html and ascii tables from r statistical output. Technical report, CRAN Repository, 2018. URL https://cran.r-project.org/web/packages/stargazer/vignettes/stargazer.pdf

    2018

  29. [37]

    C. Hsiao. Panel models with interactive effects. Journal of Econometrics, 206 0 (2): 0 645--673, 2018. doi:https://doi.org/10.1016/j.jeconom.2018.06.017. URL https://www.sciencedirect.com/science/article/pii/S0304407618301131. Special issue on Advances in Econometric Theory: E...

    2018 doi

  30. [38]

    Hsiao and M

    C. Hsiao and M. H. Peseran. Random coefficients models. In L. M\'aty\'as and P. Sevestre, editors, The Econometrics of Panel Data, chapter 6. Springer, 3rd edition, 2008

    2008

  31. [39]

    D. Hsu, S. M. Kakade, and T. Zhang. A tail inequality for quadratic forms of subgaussian random vectors. ArXiv preprint No. 1110.2842v1, 2011. URL https://arxiv.org/abs/1110.2842

    2011 arXiv

  32. [40]

    S. M. Iacus, G. King, and G. Porro. Causal inference without balance checking: Coarsened exact matching. Political analysis, pages 1--24, 2012

    2012

  33. [41]

    G. Imbens. Causal data panel models, Aug 2021. URL https://www.youtube.com/watch?v=44rzE5dKg88

    2021

  34. [42]

    A. Kaul, S. Klöbner, G. Pfeifer, and M. Schieler. Synthetic control methods: never use all pre-intervention outcomes together with covariates. Saarland Working Paper, 2018. URL https://www.gregor-pfeifer.net/files/SCM_Predictors_MC.pdf

    2018

  35. [43]

    P. Krugman. Increasing returns and economic geography. Journal of political economy, 99 0 (3): 0 483--499, 1991

    1991

  36. [44]

    A. Manning. A generalised model of monopsony*. The Economic Journal, 116 0 (508): 0 84--100, 2006. doi:https://doi.org/10.1111/j.1468-0297.2006.01048.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1468-0297.2006.01048.x

    2006

  37. [45]

    McCoy and D

    C. McCoy and D. Gonzalez. Cuban immigration and immigrants in florida and the united states: Implications for immigration policy. Technical report, University of Florida, Gainesville: Bureau of Economic and Business, 1985. referenced in Skop (2001)

    2001

  38. [46]

    H. R. Moon and M. Weidner. Linear regression for panel with unknown number of factors as interactive fixed effects. Econometrica, 83 0 (4): 0 1543--1579, 2015. doi:https://doi.org/10.3982/ECTA9382

    2015 doi

  39. [47]

    F. Oswlad. psidR: Build panel data sets from psid raw data. Technical report, CRAN Repository, 2020. URL https://cran.r-project.org/web/packages/psidR/psidR.pdf

    2020

  40. [48]

    Peri and V

    G. Peri and V. Yasenov. The labor market effects of a refugee wave. Journal of Human Resources, 54 0 (2): 0 267--309, Jan. 2018. doi:10.3368/jhr.54.2.0217.8561r1. URL https://doi.org/10.3368/jhr.54.2.0217.8561r1

    2018

  41. [49]

    M. H. Pesaran and T. Yamagata. Testing slope homogeneity in large panels. Journal of Econometrics, 142 0 (1): 0 50--93, 2008. ISSN 0304-4076. doi:https://doi.org/10.1016/j.jeconom.2007.05.010. URL https://www.sciencedirect.com/science/article/pii/S0304407607001224

    2008 doi

  42. [50]

    S. J. Redding and D. M. Sturm. The costs of remoteness: Evidence from german division and reunification. American Economic Review, 98 0 (5): 0 1766--97, 2008

    2008

  43. [51]

    Rigollet

    P. Rigollet. MIT OpenCourseWare 18.S997: High Dimensional Statistics Chapter 2 , 2015. URL https://ocw.mit.edu/courses/mathematics/18-s997-high-dimensional-statistics-spring-2015/lecture-notes/MIT18_S997S15_CourseNotes.pdf

    2015

  44. [52]

    E. H. Skop. Race and place in the adaptation of mariel exiles. The International Migration Review, 35 0 (2): 0 449--471, 2001

    2001

  45. [53]

    T. Suri. Selection and comparative advantage in technology adoption. Econometrica, 79 0 (1): 0 159--209, 2011. doi:https://doi.org/10.3982/ECTA7749. URL https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA7749

    2011 doi

  46. [54]

    US Department of Commerce

    U. US Department of Commerce. 1980 census of population: Florida. Technical report, US Department of Commerce: Census Bureau, 1982. URL https://www2.census.gov/prod2/decennial/documents/1980a_flABCs1-01.pdf

    1980

  47. [55]

    Wooldridge

    J. Wooldridge. Two-way fixed effects, the two-way mundlak regression, and event study estimators. Technical report, Working paper, 2021. URL https://www.dropbox.com/sh/zj91darudf2fica/AADj_jaf5ZuS1muobgsnxS6Za?dl=0&preview=two_way_mundlak_20210405.pdf

    2021

  48. [56]

    J. M. Wooldridge. Further results on instrumental variables estimation of average treatment effects in the correlated random coefficient model. Economics Letters, 79 0 (2): 0 185--191, 2003. doi:https://doi.org/10.1016/S0165-1765(02)00318-X. URL https://www.sciencedirect.com/s...

    2003 doi

  49. [57]

    Y. Xu. Generalized synthetic control method: Causal inference with interactive fixed effects models. Political Analysis, 25 0 (1): 0 57–76, 2017. doi:10.1017/pan.2016.2

    2017 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.