Auto-Research PipelineA durable, human-gated pipeline for computational research

← All examples

Research runad07220e3e37· 24 October 2026

Why is the college-proximity IV estimate of the return to schooling nearly twice the OLS estimate? Testing whether the gap is an Angrist-Imbens margin-weighting artefact

CompleteArchived runEconomics · Modelled on American Economic Review

Paper

Where Proximity Bites: Margin Weighting and the IV–OLS Gap in the Return to Schooling

Abstract

Card's college-proximity 2SLS return to schooling, 0.132, is 1.76 times OLS. We test whether the gap arises because IV and OLS weight schooling margins differently. In an exact decomposition for 3,010 NLSYM men, reweighting the per-margin OLS slopes by the Angrist–Imbens weights lowers the estimate to 0.069. At the observational slopes, margin weighting contributes at most 0.0095 log points per year (95 percent bound). Covariate-cell weighting contributes little, and two-year-college proximity rejects the within-margin null. The point estimate is nearly twice OLS, a difference the data do not establish (Hausman p = 0.27).

JEL codes: C26, I26, J24

Keywordsreturn to schoolinginstrumental variablescollege proximityaverage causal responseweak instruments

§Ask about this paper

Grounded in this run

Questions are answered from this run's record — the paper, the executed analysis, the data provenance and the referee report. Nothing else is consulted.

This run is finished, so the recorded paper cannot be rewritten. Questions are still answered from its record.

Pipeline

1.Literature review

Attempt 3Done

Literature review complete (live:agent:claude-opus-5-5): 10 key papers, 30 grounded sources, 5 open gaps.

Summary

Instrumental-variables estimates of the return to schooling that exploit supply-side variation are typically as large as or larger than OLS (Card 2001), and Card's college-proximity design is the canonical case: in the National Longitudinal Survey of Young Men, men who grew up near a four-year college have more schooling and higher wages, the gains are concentrated among sons of poorly educated parents, and IV estimates exceed OLS by 25-60 percent across Card's specifications - by roughly three quarters (0.132 against 0.075) in the widely reproduced main specification. Three classes of explanation compete. The first is bias: ability bias should push OLS up, measurement error in schooling pushes it down, and an exclusion failure or a moderately strong first stage can push IV up, so weak-instrument-robust inference (Anderson and Rubin 1949; Lee, McCrary, Moreira and Porter 2022) is part of any honest comparison. The second is heterogeneity in returns: if the men whose schooling responds to proximity have unusually high marginal returns - Card's discount-rate or comparative-advantage interpretation - IV exceeds OLS without either being biased for its own estimand. The third, the subject of this study, is weighting along the schooling distribution. With years of schooling as a multivalued treatment, IV identifies the Angrist-Imbens (1995) average causal response, a weighted average of per-year effects whose weights are the instrument-induced shifts in the schooling distribution at each margin (Imbens and Angrist 1994 for the binary case), while OLS is a weighted average of the adjacent-level wage slopes with different, distribution-driven weights (Yitzhaki 1996). Because the observational wage-schooling profile is non-linear - sharp gains at high-school and college completion - the two estimators can differ purely because they weight different margins. Kling (2001) computed the Angrist-Imbens weights for Card's instrument and showed it acts mainly at the transition from high school to college and mainly on disadvantaged men, but did not compare those weights with OLS's. Ishimaru (2024) formalised the comparison as an exact decomposition of the IV-OLS gap into covariate-weight, treatment-level-weight and marginal-effect components using IV-weighted OLS, applying it to other datasets; Blandhol, Bonney, Mogstad and Torgovitsky (2022, revised 2025) showed that the linear covariate specification in Card's design fails their rich-covariates test yet moves the estimate only about 8 percent. What remains open is whether the margin dimension - where on the schooling distribution proximity bites - accounts for Card's gap, and how precisely that can be established when the first stage is only moderately strong. The direction is not obvious a priori: if proximity mainly moves men into the some-college years between the two diploma margins, reweighting the observational slopes toward those years could lower rather than raise the implied return.

Key papers (10)

  1. Using Geographic Variation in College Proximity to Estimate the Return to Schooling. David Card; NBER Working Paper 4483 (1993); published in L.N. Christofides, E.K. Grant and R. Swidinsky (eds.), Aspects of Labour Market Behaviour: Essays in Honour of John Vanderkamp, University of Toronto Press, 201-222, 1995.
    Originating study and source of the NLSYM data and the college-proximity instrument; documents that proximity raises schooling and earnings mainly for sons of less-educated parents and that IV exceeds OLS by 25-60 percent across specifications.
  2. Estimating the Return to Schooling: Progress on Some Persistent Econometric Problems. David Card; Econometrica 69(5): 1127-1160, 2001.
    Documents that supply-side IV estimates are typically as large as or larger than OLS and interprets the gap as high marginal returns among the low-education men whose schooling such instruments move - the heterogeneity explanation this study must distinguish from margin weighting.
  3. Identification and Estimation of Local Average Treatment Effects. Guido W. Imbens and Joshua D. Angrist; Econometrica 62(2): 467-475, 1994.
    Establishes that with a binary instrument and monotonicity, IV identifies an average effect for compliers, the foundation for interpreting the proximity estimate as local.
  4. Two-Stage Least Squares Estimation of Average Causal Effects in Models with Variable Treatment Intensity. Joshua D. Angrist and Guido W. Imbens; Journal of the American Statistical Association 90(430): 431-442, 1995.
    Shows TSLS with a multivalued treatment estimates a weighted average of per-unit causal effects (the average causal response) with weights given by instrument-induced shifts in the treatment CDF; supplies the IV margin weights and the monotonicity check used here.
  5. Interpreting Instrumental Variables Estimates of the Returns to Schooling. Jeffrey R. Kling; Journal of Business & Economic Statistics 19(3): 358-364, 2001.
    Closest applied paper: computes the Angrist-Imbens weights for Card's college-proximity instrument, finds its greatest impact at the transition from high school to college and most weight on disadvantaged men, but does not compare the weights with OLS's.
  6. On Using Linear Regressions in Welfare Economics. Shlomo Yitzhaki; Journal of Business & Economic Statistics 14(4): 478-486, 1996.
    Shows the OLS coefficient is a weighted average of slopes between adjacent support points, which yields the OLS margin weights against which the IV weights are compared.
  7. Empirical Decomposition of the IV-OLS Gap with Heterogeneous and Nonlinear Effects. Shoya Ishimaru; Review of Economics and Statistics 106(2): 505-520, 2024.
    Closest methodological paper: decomposes the IV-OLS gap into covariate-weight, treatment-level-weight and marginal-effect components using IV-weighted OLS; applied to NLSY79, UK and US Census data but not to Card's NLSYM design.
  8. When is TSLS Actually LATE?. Christine Blandhol, John Bonney, Magne Mogstad and Alexander Torgovitsky; NBER Working Paper 29709 (revised January 2025), 2022.
    Shows TSLS with non-saturated covariates is not in general a positively weighted average of LATEs; reanalysing Card (1995) they reject rich covariates but find the flexible-covariate estimand only about 8 percent below the linear IV estimate, so the covariate dimension explains little of the gap.
  9. Valid t-Ratio Inference for IV. David S. Lee, Justin McCrary, Marcelo J. Moreira and Jack Porter; American Economic Review 112(10): 3260-3290, 2022.
    Shows conventional 2SLS t-ratio inference is distorted unless the first-stage F is far above 10, motivating weak-instrument-robust intervals for the proximity estimate and the decomposition terms.
  10. Estimation of the Parameters of a Single Equation in a Complete System of Stochastic Equations. T. W. Anderson and Herman Rubin; Annals of Mathematical Statistics 20(1): 46-63, 1949.
    Source of the Anderson-Rubin test, whose inversion gives confidence sets for the IV coefficient that remain valid however weak the first stage.

Methods in use

  • OLS Mincer wage equations of log hourly wage on years of schooling with a quadratic in potential experience and race, region and urban controls, with heteroskedasticity-robust standard errors
  • 2SLS with the four-year-college-proximity indicator as the excluded instrument, reported with its first stage, reduced form and robust first-stage F-statistic
  • Weak-instrument-robust inference: Anderson-Rubin confidence sets and the tF adjustment of Lee et al. (2022) when the first-stage F is not large
  • Angrist-Imbens average-causal-response weights estimated as coefficients from regressions of 1{S >= j} on the instrument and covariates, with their non-negativity as a check on monotonicity (Kling 2001)
  • OLS margin weights for a discrete regressor (Yitzhaki 1996) and saturated-in-schooling (schooling-dummy) wage equations giving per-margin observational slopes
  • IV-weighted OLS and exact decomposition of the IV-OLS gap into weight-difference and marginal-effect components (Ishimaru 2024)
  • Durbin-Wu-Hausman (control-function) tests of the equality of IV and OLS
  • Specification checks: family-background controls, alternative experience/age controls, subsamples, and tests of the covariate specification (Blandhol et al. 2022)

Open gaps

  • No published study asks whether the near-doubling of Card's IV estimate relative to OLS is produced by margin weighting: Kling (2001) computed where the instrument acts but never set the IV weights against the OLS weights, and Ishimaru's (2024) decomposition has not been applied to Card's NLSYM design.
  • The direction of the margin-weighting term is unknown for this design: it depends on whether the some-college margins the instrument moves carry higher or lower observational returns than the margins OLS weights most, given sheepskin jumps at 12 and 16 years.
  • Inference on decomposition components with a moderate first stage (F in the low teens) is rarely weak-instrument robust, although each component is a ratio with the first stage in the denominator.
  • Whether the IV-OLS gap is statistically distinguishable from zero under weak-instrument-robust inference is seldom reported, so it is unclear how much of the gap needs an explanation at all.
  • Whatever margin weighting leaves unexplained mixes complier heterogeneity in returns, exclusion failures and biases in the observational slopes; how large this within-margin residual is, and how precisely it is measured, has not been quantified for Card's design.
Grounded search sources (30) · mode: live:agent:claude-opus-5-5

2.Ideation

Attempt 2Done

3 candidate ideas generated; selected: “Is the College-Proximity IV-OLS Gap a Margin-Weighting Artefact? An Exact Angrist-Imbens Decomposition of Card's Return to Schooling”.

  1. Idea 1Selected

    Is the College-Proximity IV-OLS Gap a Margin-Weighting Artefact? An Exact Angrist-Imbens Decomposition of Card's Return to Schooling

    Hypothesis. In Card's NLSYM design the 2SLS return to schooling (excluded instrument: a four-year college nearby, nearc4) exceeds OLS because the instrument weights schooling margins differently from OLS, not because returns at a given margin differ. Formally: re-weighting the covariate-adjusted per-margin OLS wage slopes by the instrument's Angrist-Imbens weights (the IV-weighted OLS estimand b_RW) closes at least half of the IV-OLS gap, theta = (b_RW - b_OLS)/(b_IV - b_OLS) >= 0.5, and the remaining within-margin difference b_IV - b_RW is not statistically different from zero (its 95 percent confidence set contains zero).

    Dataset. Card (1995) college proximity and returns to schooling, NLSYM card_college_proximity

    Test plan, novelty and risk

    Test plan. Sample: all 3,010 men in the 1976 NLSYM extract (no analysis variable is missing, so no rows are dropped). Outcome lwage (log of average hourly earnings in cents); treatment educ (years, 1-18); instrument nearc4; controls W = Card's main specification: exper, expersq, black, south, smsa, smsa66 and the 1966 region dummies reg662-reg669. Step 1 (benchmarks): OLS, first stage, reduced form and just-identified 2SLS with heteroskedasticity-robust (HC1) standard errors and R-squared; robust first-stage F; Anderson-Rubin 95 percent confidence set inverted analytically from the robust joint covariance of the reduced-form and first-stage coefficients; control-function Durbin-Wu-Hausman test of b_IV = b_OLS. Step 2 (margins): write schooling as S = min(S,8) + sum_{j=9..18} 1{S>=j} (years 1-8, 132 men, pooled linearly because cells are thin); a schooling-dummy OLS wage equation with W gives per-margin slopes d_k; OLS weights m_k are the coefficients on S in regressions of each component on S and W (Yitzhaki) and reproduce b_OLS = sum m_k d_k exactly; Angrist-Imbens IV weights w_k = pi_k/pi are the coefficients on nearc4 in regressions of each component on nearc4 and W, normalised by the first stage, and sum to one. Step 3 (exact decomposition): b_IV - b_OLS = sum_k (w_k - m_k) d_k [margin-weighting term] + (b_IV - b_RW) [within-margin term], with b_RW = sum_k w_k d_k; the same identity in reduced-form units, rho = pi*b_OLS + pi*(b_RW - b_OLS) + rho_e. Step 4 (inference): nonparametric pairs bootstrap with 1,999 replications (seed 20261024); Fieller/Anderson-Rubin-type 95 percent sets for every ratio term (b_RW, the weighting term, the within-margin term and the gap) from the bootstrap covariance of numerator and first stage; percentile intervals for reduced-form-unit terms and weight blocks; a direct-effect test (nearc4 added to the schooling-dummy OLS wage equation, HC1 t-test), which tests the within-margin null without dividing by the first stage; a joint Wald test that the IV and OLS weights coincide, in reduced-form units; and the total-variation distance between the two weight vectors, which bounds the weighting term for any common per-margin return profile. Step 5 (checks): Angrist-Imbens monotonicity, i.e. non-negative proximity effects on 1{S>=j} at every threshold j = 2..18; balance of nearc4 against father's and mother's education, a library card at 14 and living with both parents at 14, conditional on W. Step 6 (robustness, 999 bootstrap replications each): years 1-11 pooled into the linear piece; family-background controls added (parents' education with missing-value indicators, family structure at 14); age and age-squared in place of experience so that every control is predetermined; men enrolled in school in 1976 excluded (n = 2,732). Heterogeneity: separate decompositions for men with at least one parent with 12 or more years of schooling and for the rest, with the first-stage F reported. Refutation: the hypothesis is refuted if theta < 0.5 and the upper limit of the 95 percent set for the weighting term lies below half the estimated gap - margin weighting then cannot deliver half the gap even at the edge of its sampling uncertainty. It is supported only if theta >= 0.5 and the within-margin set contains zero.

    Novelty. Kling (2001) computed the Angrist-Imbens weights for this instrument but never set them against the OLS weights or asked whether reweighting the observational margin slopes reproduces the IV estimate. Ishimaru (2024) supplies the general IV-OLS decomposition but applies it to other datasets, and Blandhol et al. (2022, revised 2025) examine only the covariate dimension of Card's design. The study adds weak-instrument-robust (Fieller and Anderson-Rubin) inference for every decomposition term, a reduced-form formulation of the within-margin null that needs no division by the first stage, and a profile-free bound on how large any weighting term could be.

    Expected outcome. Kling finds the instrument acts mainly at the transition from high school into college, while sheepskin jumps at 12 and 16 years mean the intervening some-college years carry small observational slopes. Shifting weight toward those years should therefore leave the reweighted OLS estimate near, or even below, OLS. If so, the weighting term is small, the hypothesis is refuted, and the gap sits in the within-margin term - which may itself be imprecise with a first-stage F in the low teens.

    Main risk. The split between the weighting and within-margin terms takes the observational per-margin slopes as the benchmark. If selection bias differs across margins (for example at the diploma years), the causal profile can differ in shape from the observational one; the total-variation bound addresses this only at the point estimates, and the IV weights are noisy. The covariate adjustment is linear rather than saturated (Blandhol et al.). Schooling is self-reported and may be misclassified across adjacent years. And with a first stage this moderate, the within-margin term and the gap may be too imprecise to sign.

  2. Idea 2

    Covariate Weighting in Card's 2SLS: Do Instrument-Sensitive Subgroups Carry Higher Returns?

    Hypothesis. The IV-OLS gap reflects covariate weighting: in a saturate-and-weight 2SLS over cells defined by race, 1966 South residence and parental education, the IV group weights favour cells whose OLS returns are highest, so reweighting the cell-specific OLS returns by the 2SLS group weights closes at least half the gap.

    Dataset. Card (1995) college proximity and returns to schooling, NLSYM card_college_proximity

    Test plan, novelty and risk

    Test plan. Same sample, outcome, treatment and instrument. Define cells by black x south66 x (any parent with 12+ years of schooling). Estimate cell-specific OLS returns and first stages; build the 2SLS group weights (proportional to cell share x first-stage coefficient squared x within-cell variance of nearc4) and the OLS group weights (proportional to cell share x within-cell residual variance of schooling, as in Słoczyński); compute the covariate-reweighted OLS estimand and its gap to interacted 2SLS, with a pairs bootstrap (999 replications) and Anderson-Rubin sets for the interacted IV. Refuted if the reweighted OLS closes less than half the gap with an upper 95 percent bound below half the gap.

    Novelty. Blandhol et al. show that flexible covariate adjustment moves Card's IV estimate only modestly but do not ask whether the group weights themselves explain the gap; Kling (2001) reports which family-background groups the instrument weights but not their OLS returns.

    Expected outcome. Kling's finding that the instrument weights disadvantaged men suggests a positive covariate-weighting term, but its size depends on whether those men's OLS returns are higher; probably a modest share of the gap.

    Main risk. Cell-specific first stages are weak (some F well below 10), so interacted 2SLS is biased toward OLS and many cell weights are noisy; the cell definition is arbitrary.

  3. Idea 3

    A Zero-First-Stage Placebo for College Proximity: Does a Nearby Two-Year College Raise Wages Without Raising Schooling?

    Hypothesis. If college proximity affects wages only through schooling, then two-year-college proximity (nearc2), which barely moves completed schooling, has no reduced-form effect on log wages conditional on four-year proximity and Card's controls.

    Dataset. Card (1995) college proximity and returns to schooling, NLSYM card_college_proximity

    Test plan, novelty and risk

    Test plan. Same sample and controls, adding nearc4 as a control. Estimate the first stage of educ on nearc2 and the reduced form of lwage on nearc2 with HC1 standard errors, and compare the reduced form with the wage effect implied by the first stage times the OLS return. A significant reduced-form wage effect with a near-zero first stage, judged by a robust t-test on the reduced-form-minus-implied contrast, would indicate an exclusion failure common to proximity instruments. Not refuted if the contrast is within its 95 percent interval of zero.

    Novelty. Kling (2001) notes as a puzzle that two-year-college proximity does not affect schooling; using that fact as a placebo for the exclusion restriction of proximity instruments has not been done for this design.

    Expected outcome. A small, imprecise reduced form; the test will have limited power.

    Main risk. nearc2 is correlated with nearc4 and urban residence, so its conditional variation is limited and partly reflects local labour markets; a null result is weak evidence of validity.

Ranking rationale

Idea 0 answers the posed question directly. The decomposition is an exact in-sample identity, so every term adds up, and each term carries weak-instrument-robust intervals; the data support it at full sample size with a first-stage F in the low teens. Idea 1 targets the covariate dimension, which Blandhol et al. found moves Card's estimate little, and it would rest on cell-level first stages too weak for reliable inference. Idea 2 is a useful validity check but cannot by itself explain the gap and has low power.

3.Experimentation

Attempt 4Done

Experiment succeeded in 1 round (19.794s).

Reviewer feedback applied: Independent referee (Claude Fable 5.1): major_revision. Required changes: (1) [analysis] Run the two-year-college proximity design from the same file: (a) nearc2 as the sole instrument, with its margin weights, b_RW and the full decomposition; (b) nearc2 and nearc4 jointly with a robust over-identification test. Add both to Table 5 and discuss whether an instrument with different margin weights produces a different IV-weighted OLS, as the weighting hypothesis predicts. (2) [analysis] Add the covariate dimension: either Ishimaru's three-way decomposition (covariate weights, margin weights, marginal effects) or a version of equation (2) with margin slopes estimated within coarse covariate cells (for example Black x South x SMSA), so that the within-margin term no longer absorbs covariate weighting. Report it next to Table 4 and revise the interpretation of the remainder accordingly. (3) [text] Report the bound honestly in Section IV.C and Table 4 Panel C: returns confined to the observed slope range could generate a weighting term of 0.0319, which is 0.561 of the gap, above the 0.5 threshold of the decision rule, and the range needed for half the gap (0.212) lies inside the observed range (0.238). Revise 'weighting alone produces the gap only if returns differ across margins by at least 0.424' and the conclusion to state that the refutation holds relative to the observational benchmark only. (4) [text] Quantify power for the within-margin term: report the minimum detectable effect at 80 percent power (about 2.8 x 0.0668 = 0.19 log points per year, roughly 2.5 times OLS; about 0.046 in reduced-form units, larger than the reduced form itself) and rewrite every sentence that calls the within-margin term 'large' or says it 'carries the whole gap' to say instead that it is not informative. (5) [text] In the abstract, Section I (paragraph 4) and Section VI, stop asserting that proximity places more weight on years 13-15 and less on the diploma years than OLS. Present these as point-estimate patterns whose intervals ([0.336, 0.760] and [0.080, 0.257]) contain the OLS values and which the Wald test does not distinguish from OLS (p = 0.62). (6) [text] Lead with the absolute bound on the weighting term (upper limit 0.0095 log points per year, 0.13 of the OLS return; in reduced-form units 0.0027 against a reduced form of 0.042) and demote theta and 'upper end / gap', stating explicitly that both divide by a point gap whose set contains zero. (7) [text] Table 2: remove the conventional stars from column 4 or replace them with Lee et al. tF-adjusted inference, and say in the text that t = 2.43 is below the tF 5 percent critical value at F = 14.1 while the Anderson-Rubin set excludes zero. (8) [analysis] For the heterogeneity split, report how many 'low background' men are classified through missing parental education (616 with at least one parent missing, 253 with both) and re-run the split on men with both parents' education observed, or define it on the observed parent only; state that the p = 0.048 equal-weights rejection is one of seven such tests. (9) [text] Cite Hungerford and Solon (1987) and Jaeger and Page (1996) where the diploma jumps are introduced; cite Heckman, Urzua and Vytlacil (2006) and Carneiro, Heckman and Vytlacil (2011) as the marginal-treatment-effect account of IV above OLS; cite Card (1999), Sloczynski (2022), Loken, Mogstad and Wiswall (2012) and Andrews, Stock and Sun (2019) where relevant. Give the exact table or section of Ishimaru (2024) that supports the claim that his NLSY79 margin term is positive, or drop the sign claim. (10) [text] Add a figure plotting the margin slopes d_k with the OLS and IV weights (and their intervals) by year of schooling; move Table 1 Panel B into its own table with its own column headings; state the software, the replication package, and that the equal-weights Wald test uses the bootstrap covariance of the reduced-form weight gaps. (11) [text] Replace 'Monotonicity survives' with 'the testable implication of monotonicity is not rejected', note that thresholds 2-8 rest on 132 men, and state how Card (1995) treated the NLS sampling weights and why unweighted estimation is appropriate here. (12) [text] Open the conclusion with 'the point estimate is nearly twice OLS, a difference the data do not establish' rather than 'is nearly twice OLS', and align the abstract's final sentence with the same wording.
Succeeded1 round · exit 0 · 19.794s

Data provenance

Datasetcard_college_proximity

Card (1995) college proximity and returns to schooling, NLSYM

Source
R package 'wooldridge' (Wooldridge, Introductory Econometrics), mirrored by the Rdatasets archive; original data from the National Longitudinal Survey of Young Men
Citation
Card, D. (1995) Using geographic variation in college proximity to estimate the return to schooling. In: Christofides, L.N., Grant, E.K. and Swidinsky, R. (eds) Aspects of Labour Market Behaviour: Essays in Honour of John Vanderkamp. Toronto: University of Toronto Press, 201-222. Data distributed with Wooldridge, J.M. Introductory Econometrics: A Modern Approach, via the Rdatasets archive.
Licence
Redistributed in the R package 'wooldridge' under GPL-3 via the Rdatasets archive; underlying survey data from the US National Longitudinal Survey of Young Men.
File
https://vincentarelbundock.github.io/Rdatasets/csv/wooldridge/card.csv
Observations
3,010
Variables (34)
id, nearc2, nearc4, educ, age, fatheduc, motheduc, weight, momdad14, sinmom14, step14, reg661, reg662, reg663, reg664, reg665, reg666, reg667, reg668, reg669, south66, black, smsa, south, smsa66, wage, enroll, KWW, IQ, married, libcrd14, exper, lwage, expersq
Units
wage is average hourly earnings in cents; lwage is its natural log. educ, exper and age are in years. nearc2 and nearc4 are the college-proximity indicators used as instruments. IQ and KWW are ability test scores. reg661-reg669 are region dummies; black, south and smsa are indicators.
Retrieved
24 Oct 2026, 02:16
Checksum
sha256 326dad817467b20d…

Real, publicly available data — downloaded and checksum-verified before the analysis ran, never simulated.

Result (RESULT_JSON)

{
  "dataset": "card_college_proximity",
  "n_obs": 3010,
  "seed": 20261024,
  "bootstrap_reps_main": 1999,
  "bootstrap_reps_other": 999,
  "chi2_crit_95": 3.84146,
  "mde_factor": 2.80159,
  "main": {
    "n": 3010,
    "F": 14.13867,
    "pi": 0.3199,
    "pi_se": 0.08508,
    "rho": 0.04207,
    "rho_se": 0.01752,
    "b_ols": 0.07469,
    "b_ols_se": 0.00365,
    "r2_ols": 0.29984,
    "r2_fs": 0.47712,
    "r2_rf": 0.19469,
    "b_iv": 0.1315,
    "b_iv_se": 0.05414,
    "ar_ci": [
      0.02818,
      0.28115
    ],
    "ar_p_at_ols": 0.27076,
    "b_rw": 0.06938,
    "gap": 0.05681,
    "weight": -0.00531,
    "within": 0.06212,
    "share_weight": -0.09348,
    "dwh_p": 0.27092,
    "direct": 0.02001,
    "direct_se": 0.01652,
    "direct_p": 0.22585,
    "b_rw_se": 0.00879,
    "weight_se": 0.00787,
    "within_se": 0.06682,
    "gap_se": 0.06687,
    "b_rw_ci": [
      0.04958,
      0.08658
    ],
    "weight_ci": [
      -0.02233,
      0.00954
    ],
    "within_ci": [
      -0.04137,
      0.2107
    ],
    "gap_ci": [
      -0.04771,
      0.20424
    ],
    "T1": 0.02389,
    "T1_ci": [
      0.01137,
      0.03676
    ],
    "T1_se": 0.00648,
    "T2": -0.0017,
    "T2_ci": [
      -0.00606,
      0.00265
    ],
    "T2_se": 0.00221,
    "T3": 0.01987,
    "T3_ci": [
      -0.01389,
      0.05235
    ],
    "T3_se": 0.01651,
    "T3_p": 0.22885,
    "weighting_ub_share": 0.16785,
    "weight_ub_over_ols": 0.12766,
    "mde_within": 0.18721,
    "mde_T3": 0.04627,
    "mde_within_over_ols": 2.50638,
    "mde_T3_over_rho": 1.09982,
    "t_iv": 2.4288,
    "tf_crit": 2.9255,
    "tv": 0.13398,
    "tv_pct": [
      0.09289,
      0.4665
    ],
    "wald_equal_weights": 8.08015,
    "wald_df": 10,
    "wald_p": 0.62101,
    "slope_range": 0.23777,
    "required_range": 0.42402,
    "required_range_half": 0.21201,
    "required_range_at_tv_p975": 0.12178,
    "required_range_over_ols": 5.67679,
    "max_weight_in_slope_range": 0.03186,
    "max_weight_share_of_gap": 0.56075,
    "max_weight_share_at_tv_p975": 1.95242,
    "required_range_half_at_tv_p975": 0.06089,
    "margins": {
      "label": [
        "1-8 (per year)",
        "9",
        "10",
        "11",
        "12",
        "13",
        "14",
        "15",
        "16",
        "17",
        "18"
      ],
      "n_at_level": [
        132,
        81,
        125,
        159,
        992,
        281,
        263,
        160,
        459,
        151,
        207
      ],
      "slope": [
        0.084,
        -0.03425,
        0.01489,
        0.11276,
        0.15718,
        0.08185,
        0.03612,
        -0.02179,
        0.20352,
        0.00648,
        0.14793
      ],
      "slope_se": [
        0.02209,
        0.05987,
        0.0531,
        0.04216,
        0.03272,
        0.02358,
        0.03157,
        0.03809,
        0.03407,
        0.03958,
        0.04531
      ],
      "w_ols": [
        0.04904,
        0.03153,
        0.04581,
        0.06102,
        0.07708,
        0.14525,
        0.15431,
        0.15125,
        0.14058,
        0.08571,
        0.05842
      ],
      "w_ols_se": [
        0.00798,
        0.00213,
        0.00247,
        0.00271,
        0.00287,
        0.00285,
        0.00286,
        0.00294,
        0.003,
        0.00334,
        0.00326
      ],
      "w_iv": [
        0.00839,
        0.04378,
        0.0758,
        0.0639,
        0.08334,
        0.19886,
        0.17755,
        0.13796,
        0.0945,
        0.05175,
        0.06417
      ],
      "w_iv_se": [
        0.06406,
        0.02722,
        0.03251,
        0.03593,
        0.04353,
        0.05523,
        0.04363,
        0.04066,
        0.04343,
        0.04267,
        0.03212
      ],
      "w_iv_ci": [
        [
          -0.1377,
          0.11331
        ],
        [
          -0.01101,
          0.09853
        ],
        [
          0.0208,
          0.15088
        ],
        [
          -0.00431,
          0.13644
        ],
        [
          0.00242,
          0.17515
        ],
        [
          0.11585,
          0.33523
        ],
        [
          0.10132,
          0.27666
        ],
        [
          0.05632,
          0.21172
        ],
        [
          -0.00367,
          0.15989
        ],
        [
          -0.05014,
          0.11866
        ],
        [
          0.00292,
          0.1335
        ]
      ],
      "w_ols_ci": [
        [
          0.03422,
          0.06516
        ],
        [
          0.02725,
          0.03571
        ],
        [
          0.0411,
          0.0506
        ],
        [
          0.05578,
          0.06638
        ],
        [
          0.07175,
          0.08294
        ],
        [
          0.13986,
          0.15091
        ],
        [
          0.14885,
          0.15995
        ],
        [
          0.14556,
          0.157
        ],
        [
          0.13478,
          0.14631
        ],
        [
          0.07902,
          0.09205
        ],
        [
          0.05199,
          0.06459
        ]
      ],
      "pi_k": [
        0.00269,
        0.01401,
        0.02425,
        0.02044,
        0.02666,
        0.06361,
        0.0568,
        0.04413,
        0.03023,
        0.01656,
        0.02053
      ],
      "pi_k_se": [
        0.01759,
        0.00839,
        0.00975,
        0.01122,
        0.01371,
        0.01802,
        0.01727,
        0.01679,
        0.01624,
        0.01275,
        0.01001
      ]
    },
    "m13_15_iv": 0.51437,
    "m13_15_iv_ci": [
      0.33591,
      0.75996
    ],
    "m13_15_ols": 0.45081,
    "m13_15_ols_ci": [
      0.4363,
      0.46611
    ],
    "m12_16_iv": 0.17784,
    "m12_16_iv_ci": [
      0.07951,
      0.25668
    ],
    "m12_16_ols": 0.21767,
    "m12_16_ols_ci": [
      0.21213,
      0.22355
    ],
    "rho_over_pred": 1.89535,
    "within_over_ols": 0.83169,
    "iv_over_ols": 1.76059
  },
  "monotonicity": {
    "min_t": -0.50594,
    "n_negative": 2,
    "thresholds": [
      [
        2,
        0.00136,
        1.01342
      ],
      [
        3,
        0.00104,
        0.64775
      ],
      [
        4,
        0.00013,
        0.06245
      ],
      [
        5,
        0.00027,
        0.10698
      ],
      [
        6,
        -0.00075,
        -0.17854
      ],
      [
        7,
        -0.0026,
        -0.50594
      ],
      [
        8,
        0.00323,
        0.51267
      ],
      [
        9,
        0.01401,
        1.67011
      ],
      [
        10,
        0.02425,
        2.48743
      ],
      [
        11,
        0.02044,
        1.82103
      ],
      [
        12,
        0.02666,
        1.94459
      ],
      [
        13,
        0.06361,
        3.52929
      ],
      [
        14,
        0.0568,
        3.28795
      ],
      [
        15,
        0.04413,
        2.62835
      ],
      [
        16,
        0.03023,
        1.86176
      ],
      [
        17,
        0.01656,
        1.29899
      ],
      [
        18,
        0.02053,
        2.05018
      ]
    ]
  },
  "balance": {
    "fatheduc": {
      "n": 2320,
      "mean": 10.00345,
      "coef": 0.20194,
      "se": 0.16502,
      "p": 0.22105
    },
    "motheduc": {
      "n": 2657,
      "mean": 10.34814,
      "coef": -0.08643,
      "se": 0.13471,
      "p": 0.5211
    },
    "libcrd14": {
      "n": 2997,
      "mean": 0.67434,
      "coef": 0.0373,
      "se": 0.01953,
      "p": 0.05617
    },
    "momdad14": {
      "n": 3010,
      "mean": 0.78937,
      "coef": -0.02241,
      "se": 0.01805,
      "p": 0.2145
    }
  },
  "summary": {
    "all": {
      "n": 3010,
      "wage_dollars": 5.77282,
      "lwage": 6.26183,
      "educ": 13.26346,
      "educ_sd": 2.67691,
      "exper": 8.85615,
      "age": 28.1196,
      "black": 0.23355,
      "south66": 0.41429,
      "smsa66": 0.6495,
      "enroll": 0.09236,
      "share_educ12": 0.32957,
      "share_educ16plus": 0.27143
    },
    "near": {
      "n": 2053,
      "wage_dollars": 6.05636,
      "lwage": 6.3114,
      "educ": 13.52703,
      "educ_sd": 2.58045,
      "exper": 8.68193,
      "age": 28.20896,
      "black": 0.21189,
      "south66": 0.32879,
      "smsa66": 0.79932,
      "enroll": 0.1018,
      "share_educ12": 0.32002,
      "share_educ16plus": 0.29323
    },
    "far": {
      "n": 957,
      "wage_dollars": 5.16457,
      "lwage": 6.15549,
      "educ": 12.69801,
      "educ_sd": 2.79152,
      "exper": 9.22989,
      "age": 27.9279,
      "black": 0.28004,
      "south66": 0.5977,
      "smsa66": 0.32811,
      "enroll": 0.0721,
      "share_educ12": 0.35005,
      "share_educ16plus": 0.22466
    }
  },
  "robustness": {
    "pool_1_11": {
      "n": 3010,
      "F": 14.13867,
      "pi": 0.3199,
      "pi_se": 0.08508,
      "rho": 0.04207,
      "rho_se": 0.01752,
      "b_ols": 0.07469,
      "b_ols_se": 0.00365,
      "r2_ols": 0.29984,
      "r2_fs": 0.47712,
      "r2_rf": 0.19469,
      "b_iv": 0.1315,
      "b_iv_se": 0.05414,
      "ar_ci": [
        0.02818,
        0.28115
      ],
      "ar_p_at_ols": 0.27076,
      "b_rw": 0.07279,
      "gap": 0.05681,
      "weight": -0.0019,
      "within": 0.05871,
      "share_weight": -0.03349,
      "dwh_p": 0.27092,
      "direct": 0.01889,
      "direct_se": 0.0165,
      "direct_p": 0.2522,
      "b_rw_se": 0.00841,
      "weight_se": 0.00748,
      "within_se": 0.06863,
      "gap_se": 0.06862,
      "b_rw_ci": [
        0.05596,
        0.0902
      ],
      "weight_ci": [
        -0.01626,
        0.01319
      ],
      "within_ci": [
        -0.04473,
        0.20236
      ],
      "gap_ci": [
        -0.04635,
        0.20091
      ],
      "T1": 0.02389,
      "T1_ci": [
        0.01125,
        0.03566
      ],
      "T1_se": 0.00638,
      "T2": -0.00061,
      "T2_ci": [
        -0.00466,
        0.00333
      ],
      "T2_se": 0.00207,
      "T3": 0.01878,
      "T3_ci": [
        -0.01579,
        0.04958
      ],
      "T3_se": 0.01641,
      "T3_p": 0.25252,
      "weighting_ub_share": 0.23212,
      "weight_ub_over_ols": 0.17654,
      "mde_within": 0.19226,
      "mde_T3": 0.04599,
      "mde_within_over_ols": 2.57404,
      "mde_T3_over_rho": 1.09314,
      "t_iv": 2.4288,
      "tf_crit": 2.9255,
      "tv": 0.09334,
      "tv_pct": [
        0.06623,
        0.4172
      ],
      "wald_equal_weights": 5.13901,
      "wald_df": 7,
      "wald_p": 0.643,
      "slope_range": 0.22518,
      "required_range": 0.60866,
      "required_range_half": 0.30433,
      "required_range_at_tv_p975": 0.13617,
      "required_range_over_ols": 8.14884,
      "max_weight_in_slope_range": 0.02102,
      "max_weight_share_of_gap": 0.36997,
      "max_weight_share_at_tv_p975": 1.65368,
      "required_range_half_at_tv_p975": 0.06809
    },
    "family_background": {
      "n": 3010,
      "F": 16.73098,
      "pi": 0.32894,
      "pi_se": 0.08042,
      "rho": 0.04431,
      "rho_se": 0.01721,
      "b_ols": 0.07231,
      "b_ols_se": 0.00384,
      "r2_ols": 0.30262,
      "r2_fs": 0.52767,
      "r2_rf": 0.21392,
      "b_iv": 0.1347,
      "b_iv_se": 0.05309,
      "ar_ci": [
        0.03579,
        0.27679
      ],
      "ar_p_at_ols": 0.2131,
      "b_rw": 0.06692,
      "gap": 0.06239,
      "weight": -0.00538,
      "within": 0.06778,
      "share_weight": -0.0863,
      "dwh_p": 0.21325,
      "direct": 0.02247,
      "direct_se": 0.01653,
      "direct_p": 0.17392,
      "b_rw_se": 0.00876,
      "weight_se": 0.00769,
      "within_se": 0.06485,
      "gap_se": 0.06434,
      "b_rw_ci": [
        0.04798,
        0.0838
      ],
      "weight_ci": [
        -0.02172,
        0.00918
      ],
      "within_ci": [
        -0.03077,
        0.20535
      ],
      "gap_ci": [
        -0.0364,
        0.19844
      ],
      "T1": 0.02378,
      "T1_ci": [
        0.01178,
        0.03533
      ],
      "T1_se": 0.00584,
      "T2": -0.00177,
      "T2_ci": [
        -0.00622,
        0.0025
      ],
      "T2_se": 0.00228,
      "T3": 0.02229,
      "T3_ci": [
        -0.01261,
        0.05237
      ],
      "T3_se": 0.01639,
      "T3_p": 0.17368,
      "weighting_ub_share": 0.14708,
      "weight_ub_over_ols": 0.12691,
      "mde_within": 0.18168,
      "mde_T3": 0.04591,
      "mde_within_over_ols": 2.5127,
      "mde_T3_over_rho": 1.03618,
      "t_iv": 2.53709,
      "tf_crit": 2.75212,
      "tv": 0.13902,
      "tv_pct": [
        0.09157,
        0.41629
      ],
      "wald_equal_weights": 8.74536,
      "wald_df": 10,
      "wald_p": 0.55643,
      "slope_range": 0.23535,
      "required_range": 0.44879,
      "required_range_half": 0.22439,
      "required_range_at_tv_p975": 0.14988,
      "required_range_over_ols": 6.20677,
      "max_weight_in_slope_range": 0.03272,
      "max_weight_share_of_gap": 0.52441,
      "max_weight_share_at_tv_p975": 1.57029,
      "required_range_half_at_tv_p975": 0.07494
    },
    "age_controls": {
      "n": 3010,
      "F": 11.64446,
      "pi": 0.39238,
      "pi_se": 0.11499,
      "rho": 0.03246,
      "rho_se": 0.01708,
      "b_ols": 0.03438,
      "b_ols_se": 0.00277,
      "r2_ols": 0.28918,
      "r2_fs": 0.12389,
      "r2_rf": 0.25221,
      "b_iv": 0.08271,
      "b_iv_se": 0.04469,
      "ar_ci": [
        -0.00299,
        0.21631
      ],
      "ar_p_at_ols": 0.25386,
      "b_rw": 0.03136,
      "gap": 0.04833,
      "weight": -0.00303,
      "within": 0.05136,
      "share_weight": -0.0626,
      "dwh_p": 0.25341,
      "direct": 0.02028,
      "direct_se": 0.0166,
      "direct_p": 0.22175,
      "b_rw_se": 0.01726,
      "weight_se": 0.017,
      "within_se": 0.20539,
      "gap_se": 0.19166,
      "b_rw_ci": [
        0.01296,
        0.04744
      ],
      "weight_ci": [
        -0.01984,
        0.01152
      ],
      "within_ci": [
        -0.03276,
        0.19179
      ],
      "gap_ci": [
        -0.03605,
        0.18676
      ],
      "T1": 0.01349,
      "T1_ci": [
        0.00572,
        0.02147
      ],
      "T1_se": 0.00418,
      "T2": -0.00119,
      "T2_ci": [
        -0.00581,
        0.0037
      ],
      "T2_se": 0.00253,
      "T3": 0.02015,
      "T3_ci": [
        -0.01412,
        0.05087
      ],
      "T3_se": 0.01647,
      "T3_p": 0.22115,
      "weighting_ub_share": 0.23832,
      "weight_ub_over_ols": 0.33501,
      "mde_within": 0.57542,
      "mde_T3": 0.04614,
      "mde_within_over_ols": 16.73618,
      "mde_T3_over_rho": 1.42179,
      "t_iv": 1.85077,
      "tf_crit": 3.17707,
      "tv": 0.11329,
      "tv_pct": [
        0.07734,
        0.43255
      ],
      "wald_equal_weights": 8.09988,
      "wald_df": 10,
      "wald_p": 0.61908,
      "slope_range": 0.21984,
      "required_range": 0.42661,
      "required_range_half": 0.2133,
      "required_range_at_tv_p975": 0.11174,
      "required_range_over_ols": 12.40794,
      "max_weight_in_slope_range": 0.02491,
      "max_weight_share_of_gap": 0.51531,
      "max_weight_share_at_tv_p975": 1.96746,
      "required_range_half_at_tv_p975": 0.05587
    },
    "not_enrolled": {
      "n": 2732,
      "F": 12.43669,
      "pi": 0.31532,
      "pi_se": 0.08941,
      "rho": 0.04148,
      "rho_se": 0.01842,
      "b_ols": 0.07661,
      "b_ols_se": 0.00379,
      "r2_ols": 0.32208,
      "r2_fs": 0.47599,
      "r2_rf": 0.20874,
      "b_iv": 0.13156,
      "b_iv_se": 0.05736,
      "ar_ci": [
        0.02053,
        0.29627
      ],
      "ar_p_at_ols": 0.31594,
      "b_rw": 0.07028,
      "gap": 0.05495,
      "weight": -0.00633,
      "within": 0.06127,
      "share_weight": -0.11515,
      "dwh_p": 0.31603,
      "direct": 0.01948,
      "direct_se": 0.01737,
      "direct_p": 0.26221,
      "b_rw_se": 0.01203,
      "weight_se": 0.01116,
      "within_se": 0.09038,
      "gap_se": 0.09106,
      "b_rw_ci": [
        0.0479,
        0.08801
      ],
      "weight_ci": [
        -0.02608,
        0.00842
      ],
      "within_ci": [
        -0.0516,
        0.24272
      ],
      "gap_ci": [
        -0.0596,
        0.23306
      ],
      "T1": 0.02416,
      "T1_ci": [
        0.01009,
        0.0378
      ],
      "T1_se": 0.00712,
      "T2": -0.002,
      "T2_ci": [
        -0.00646,
        0.00202
      ],
      "T2_se": 0.00221,
      "T3": 0.01932,
      "T3_ci": [
        -0.01537,
        0.05596
      ],
      "T3_se": 0.01776,
      "T3_p": 0.27659,
      "weighting_ub_share": 0.1533,
      "weight_ub_over_ols": 0.10995,
      "mde_within": 0.2532,
      "mde_T3": 0.04975,
      "mde_within_over_ols": 3.30516,
      "mde_T3_over_rho": 1.19932,
      "t_iv": 2.29351,
      "tf_crit": 3.08358,
      "tv": 0.15523,
      "tv_pct": [
        0.10525,
        0.56681
      ],
      "wald_equal_weights": 11.76549,
      "wald_df": 10,
      "wald_p": 0.30106,
      "slope_range": 0.19109,
      "required_range": 0.35397,
      "required_range_half": 0.17699,
      "required_range_at_tv_p975": 0.09694,
      "required_range_over_ols": 4.62049,
      "max_weight_in_slope_range": 0.02966,
      "max_weight_share_of_gap": 0.53984,
      "max_weight_share_at_tv_p975": 1.97121,
      "required_range_half_at_tv_p975": 0.04847
    }
  },
  "heterogeneity": {
    "high_background": {
      "n": 1579,
      "F": 12.37712,
      "pi": 0.42361,
      "pi_se": 0.12041,
      "rho": 0.0436,
      "rho_se": 0.02531,
      "b_ols": 0.07852,
      "b_ols_se": 0.00529,
      "r2_ols": 0.25174,
      "r2_fs": 0.3215,
      "r2_rf": 0.13441,
      "b_iv": 0.10292,
      "b_iv_se": 0.05551,
      "ar_ci": [
        -0.0185,
        0.24429
      ],
      "ar_p_at_ols": 0.65812,
      "b_rw": 0.06501,
      "gap": 0.02441,
      "weight": -0.01351,
      "within": 0.03792,
      "share_weight": -0.55354,
      "dwh_p": 0.65708,
      "direct": 0.01632,
      "direct_se": 0.02338,
      "direct_p": 0.48517,
      "b_rw_se": 0.01473,
      "weight_se": 0.01429,
      "within_se": 0.06281,
      "gap_se": 0.06322,
      "b_rw_ci": [
        0.02815,
        0.08786
      ],
      "weight_ci": [
        -0.04855,
        0.00753
      ],
      "within_ci": [
        -0.07133,
        0.18232
      ],
      "gap_ci": [
        -0.09162,
        0.16159
      ],
      "T1": 0.03326,
      "T1_ci": [
        0.01519,
        0.0527
      ],
      "T1_se": 0.00973,
      "T2": -0.00572,
      "T2_ci": [
        -0.01452,
        0.00295
      ],
      "T2_se": 0.00451,
      "T3": 0.01606,
      "T3_ci": [
        -0.02701,
        0.05682
      ],
      "T3_se": 0.02206,
      "T3_p": 0.4665,
      "weighting_ub_share": 0.30856,
      "weight_ub_over_ols": 0.09592,
      "mde_within": 0.17596,
      "mde_T3": 0.0618,
      "mde_within_over_ols": 2.24113,
      "mde_T3_over_rho": 1.41741,
      "t_iv": 1.85411,
      "tf_crit": 3.08968,
      "tv": 0.12227,
      "tv_pct": [
        0.07774,
        0.42508
      ],
      "wald_equal_weights": 13.22062,
      "wald_df": 10,
      "wald_p": 0.2116,
      "slope_range": 0.7073,
      "required_range": 0.19961,
      "required_range_half": 0.09981,
      "required_range_at_tv_p975": 0.05742,
      "required_range_over_ols": 2.54233,
      "max_weight_in_slope_range": 0.08648,
      "max_weight_share_of_gap": 3.54334,
      "max_weight_share_at_tv_p975": 12.31843,
      "required_range_half_at_tv_p975": 0.02871
    },
    "low_background": {
      "n": 1431,
      "F": 3.70129,
      "pi": 0.22082,
      "pi_se": 0.11478,
      "rho": 0.04202,
      "rho_se": 0.02395,
      "b_ols": 0.06772,
      "b_ols_se": 0.00521,
      "r2_ols": 0.31275,
      "r2_fs": 0.48427,
      "r2_rf": 0.23093,
      "b_iv": 0.19028,
      "b_iv_se": 0.12725,
      "ar_ci": null,
      "ar_p_at_ols": 0.24461,
      "b_rw": 0.06324,
      "gap": 0.12256,
      "weight": -0.00448,
      "within": 0.12704,
      "share_weight": -0.03655,
      "dwh_p": 0.24558,
      "direct": 0.02845,
      "direct_se": 0.02319,
      "direct_p": 0.21988,
      "b_rw_se": 1.30836,
      "weight_se": 1.30853,
      "within_se": 32.48763,
      "gap_se": 33.79053,
      "b_rw_ci": null,
      "weight_ci": null,
      "within_ci": null,
      "gap_ci": null,
      "T1": 0.01495,
      "T1_ci": [
        -0.00041,
        0.03138
      ],
      "T1_se": 0.00788,
      "T2": -0.00099,
      "T2_ci": [
        -0.00876,
        0.00825
      ],
      "T2_se": 0.00428,
      "T3": 0.02805,
      "T3_ci": [
        -0.01096,
        0.07205
      ],
      "T3_se": 0.0222,
      "T3_p": 0.20628,
      "weighting_ub_share": null,
      "weight_ub_over_ols": null,
      "mde_within": 91.01686,
      "mde_T3": 0.06218,
      "mde_within_over_ols": 1343.98913,
      "mde_T3_over_rho": 1.47995,
      "t_iv": 1.4953,
      "tf_crit": null,
      "tv": 0.53153,
      "tv_pct": [
        0.23874,
        5.25557
      ],
      "wald_equal_weights": 18.47182,
      "wald_df": 10,
      "wald_p": 0.04751,
      "slope_range": 0.40063,
      "required_range": 0.23057,
      "required_range_half": 0.11529,
      "required_range_at_tv_p975": 0.02332,
      "required_range_over_ols": 3.40471,
      "max_weight_in_slope_range": 0.21295,
      "max_weight_share_of_gap": 1.73757,
      "max_weight_share_at_tv_p975": 17.18044,
      "required_range_half_at_tv_p975": 0.01166
    },
    "high_background_obs": {
      "n": 1405,
      "F": 11.31352,
      "pi": 0.42145,
      "pi_se": 0.1253,
      "rho": 0.03411,
      "rho_se": 0.02625,
      "b_ols": 0.07527,
      "b_ols_se": 0.00566,
      "r2_ols": 0.23705,
      "r2_fs": 0.31035,
      "r2_rf": 0.12895,
      "b_iv": 0.08093,
      "b_iv_se": 0.05827,
      "ar_ci": [
        -0.05626,
        0.22489
      ],
      "ar_p_at_ols": 0.92254,
      "b_rw": 0.0675,
      "gap": 0.00566,
      "weight": -0.00777,
      "within": 0.01343,
      "share_weight": -1.37225,
      "dwh_p": 0.92207,
      "direct": 0.00574,
      "direct_se": 0.02463,
      "direct_p": 0.81571,
      "b_rw_se": 0.01277,
      "weight_se": 0.01175,
      "within_se": 0.07347,
      "gap_se": 0.07415,
      "b_rw_ci": [
        0.03893,
        0.09007
      ],
      "weight_ci": [
        -0.03292,
        0.01178
      ],
      "within_ci": [
        -0.11622,
        0.15904
      ],
      "gap_ci": [
        -0.12676,
        0.14844
      ],
      "T1": 0.03172,
      "T1_ci": [
        0.01397,
        0.05142
      ],
      "T1_se": 0.00945,
      "T2": -0.00328,
      "T2_ci": [
        -0.01139,
        0.00456
      ],
      "T2_se": 0.00388,
      "T3": 0.00566,
      "T3_ci": [
        -0.03971,
        0.05247
      ],
      "T3_se": 0.02437,
      "T3_p": 0.81628,
      "weighting_ub_share": 2.08008,
      "weight_ub_over_ols": 0.1565,
      "mde_within": 0.20582,
      "mde_T3": 0.06827,
      "mde_within_over_ols": 2.73457,
      "mde_T3_over_rho": 2.00164,
      "t_iv": 1.38883,
      "tf_crit": 3.22084,
      "tv": 0.10023,
      "tv_pct": [
        0.07142,
        0.41627
      ],
      "wald_equal_weights": 10.6046,
      "wald_df": 10,
      "wald_p": 0.38914,
      "slope_range": 0.49022,
      "required_range": 0.0565,
      "required_range_half": 0.02825,
      "required_range_at_tv_p975": 0.0136,
      "required_range_over_ols": 0.75062,
      "max_weight_in_slope_range": 0.04914,
      "max_weight_share_of_gap": 8.67695,
      "max_weight_share_at_tv_p975": 36.03503,
      "required_range_half_at_tv_p975": 0.0068
    },
    "low_background_obs": {
      "n": 815,
      "F": 0.42302,
      "pi": 0.09584,
      "pi_se": 0.14736,
      "rho": 0.01636,
      "rho_se": 0.03167,
      "b_ols": 0.07569,
      "b_ols_se": 0.00672,
      "r2_ols": 0.30181,
      "r2_fs": 0.45777,
      "r2_rf": 0.19089,
      "b_iv": 0.17066,
      "b_iv_se": 0.35109,
      "ar_ci": null,
      "ar_p_at_ols": 0.76213,
      "b_rw": 0.105,
      "gap": 0.09497,
      "weight": 0.02931,
      "within": 0.06566,
      "share_weight": 0.30864,
      "dwh_p": 0.76236,
      "direct": 0.00642,
      "direct_se": 0.0299,
      "direct_p": 0.82989,
      "b_rw_se": 1.8777,
      "weight_se": 1.8778,
      "within_se": 7.77469,
      "gap_se": 8.0386,
      "b_rw_ci": null,
      "weight_ci": null,
      "within_ci": null,
      "gap_ci": null,
      "T1": 0.00725,
      "T1_ci": [
        -0.01431,
        0.02884
      ],
      "T1_se": 0.01105,
      "T2": 0.00281,
      "T2_ci": [
        -0.01262,
        0.01881
      ],
      "T2_se": 0.0078,
      "T3": 0.00629,
      "T3_ci": [
        -0.05377,
        0.06325
      ],
      "T3_se": 0.0295,
      "T3_p": 0.83111,
      "weighting_ub_share": null,
      "weight_ub_over_ols": null,
      "mde_within": 21.78146,
      "mde_T3": 0.08266,
      "mde_within_over_ols": 287.76566,
      "mde_T3_over_rho": 5.05361,
      "t_iv": 0.48608,
      "tf_crit": null,
      "tv": 1.5036,
      "tv_pct": [
        0.34714,
        21.85857
      ],
      "wald_equal_weights": 18.05686,
      "wald_df": 10,
      "wald_p": 0.05401,
      "slope_range": 0.50812,
      "required_range": 0.06316,
      "required_range_half": 0.03158,
      "required_range_at_tv_p975": 0.00434,
      "required_range_over_ols": 0.83443,
      "max_weight_in_slope_range": 0.76401,
      "max_weight_share_of_gap": 8.045,
      "max_weight_share_at_tv_p975": 116.95377,
      "required_range_half_at_tv_p975": 0.00217
    }
  },
  "heterogeneity_counts": {
    "low": 1431,
    "low_any_missing": 616,
    "low_both_missing": 253,
    "both_obs": 2220
  },
  "second_instrument_and_weights": {
    "nearc2_alone": {
      "n": 3010,
      "F": 2.42896,
      "pi": 0.12162,
      "pi_se": 0.07803,
      "rho": 0.03565,
      "rho_se": 0.016,
      "b_ols": 0.07469,
      "b_ols_se": 0.00365,
      "r2_ols": 0.29984,
      "r2_fs": 0.47523,
      "r2_rf": 0.19458,
      "b_iv": 0.29317,
      "b_iv_se": 0.18625,
      "ar_ci": null,
      "ar_p_at_ols": 0.07469,
      "b_rw": 0.03888,
      "gap": 0.21848,
      "weight": -0.03581,
      "within": 0.25429,
      "share_weight": -0.16391,
      "dwh_p": 0.07401,
      "direct": 0.03117,
      "direct_se": 0.01482,
      "direct_p": 0.03542,
      "b_rw_se": 0.78223,
      "weight_se": 0.78218,
      "within_se": 2.56345,
      "gap_se": 1.94367,
      "b_rw_ci": null,
      "weight_ci": null,
      "within_ci": null,
      "gap_ci": null,
      "T1": 0.00908,
      "T1_ci": [
        -0.00242,
        0.0206
      ],
      "T1_se": 0.00583,
      "T2": -0.00436,
      "T2_ci": [
        -0.00879,
        -0.00026
      ],
      "T2_se": 0.00222,
      "T3": 0.03093,
      "T3_ci": [
        0.00479,
        0.05879
      ],
      "T3_se": 0.01396,
      "T3_p": 0.02672,
      "weighting_ub_share": null,
      "weight_ub_over_ols": null,
      "mde_within": 7.18173,
      "mde_T3": 0.03911,
      "mde_within_over_ols": 96.14968,
      "mde_T3_over_rho": 1.09681,
      "t_iv": 1.57413,
      "tf_crit": null,
      "tv": 0.57266,
      "tv_pct": [
        0.23426,
        9.94844
      ],
      "wald_equal_weights": 22.10659,
      "wald_df": 10,
      "wald_p": 0.01457,
      "slope_range": 0.23777,
      "required_range": 0.38152,
      "required_range_half": 0.19076,
      "required_range_at_tv_p975": 0.02196,
      "required_range_over_ols": 5.1078,
      "max_weight_in_slope_range": 0.13616,
      "max_weight_share_of_gap": 0.62322,
      "max_weight_share_at_tv_p975": 10.82669,
      "required_range_half_at_tv_p975": 0.01098,
      "margins": {
        "label": [
          "1-8 (per year)",
          "9",
          "10",
          "11",
          "12",
          "13",
          "14",
          "15",
          "16",
          "17",
          "18"
        ],
        "n_at_level": [
          132,
          81,
          125,
          159,
          992,
          281,
          263,
          160,
          459,
          151,
          207
        ],
        "slope": [
          0.084,
          -0.03425,
          0.01489,
          0.11276,
          0.15718,
          0.08185,
          0.03612,
          -0.02179,
          0.20352,
          0.00648,
          0.14793
        ],
        "slope_se": [
          0.02209,
          0.05987,
          0.0531,
          0.04216,
          0.03272,
          0.02358,
          0.03157,
          0.03809,
          0.03407,
          0.03958,
          0.04531
        ],
        "w_ols": [
          0.04904,
          0.03153,
          0.04581,
          0.06102,
          0.07708,
          0.14525,
          0.15431,
          0.15125,
          0.14058,
          0.08571,
          0.05842
        ],
        "w_ols_se": [
          0.00807,
          0.00208,
          0.00242,
          0.00263,
          0.0028,
          0.00291,
          0.00293,
          0.00295,
          0.00302,
          0.00332,
          0.00324
        ],
        "w_iv": [
          -0.086,
          -0.01284,
          0.05596,
          -0.14299,
          -0.11216,
          0.25586,
          0.26838,
          0.32859,
          0.17308,
          0.14401,
          0.12812
        ],
        "w_iv_se": [
          2.19628,
          0.77351,
          0.733,
          2.53234,
          2.23447,
          1.1159,
          1.61729,
          1.66916,
          1.6999,
          2.38248,
          2.24796
        ],
        "w_iv_ci": [
          [
            -1.36068,
            0.76203
          ],
          [
            -0.51035,
            0.37796
          ],
          [
            -0.34013,
            0.61164
          ],
          [
            -1.80852,
            1.37281
          ],
          [
            -1.67544,
            1.00683
          ],
          [
            -0.53054,
            1.22356
          ],
          [
            -0.63876,
            1.05944
          ],
          [
            -0.88679,
            1.63728
          ],
          [
            -0.42441,
            0.70028
          ],
          [
            -0.3832,
            0.72337
          ],
          [
            -0.33963,
            0.75599
          ]
        ],
        "w_ols_ci": [
          [
            0.0342,
            0.06472
          ],
          [
            0.02736,
            0.03553
          ],
          [
            0.04135,
            0.05043
          ],
          [
            0.05565,
            0.06605
          ],
          [
            0.07183,
            0.08269
          ],
          [
            0.13994,
            0.151
          ],
          [
            0.14869,
            0.16006
          ],
          [
            0.14555,
            0.15704
          ],
          [
            0.13466,
            0.14626
          ],
          [
            0.07895,
            0.09203
          ],
          [
            0.05226,
            0.06461
          ]
        ],
        "pi_k": [
          -0.01046,
          -0.00156,
          0.00681,
          -0.01739,
          -0.01364,
          0.03112,
          0.03264,
          0.03996,
          0.02105,
          0.01751,
          0.01558
        ],
        "pi_k_se": [
          0.01335,
          0.00697,
          0.00848,
          0.01008,
          0.0122,
          0.01638,
          0.01594,
          0.01537,
          0.015,
          0.01235,
          0.01005
        ]
      },
      "m13_15_iv": 0.85283,
      "m13_15_iv_ci": [
        -1.90149,
        4.28441
      ],
      "m13_15_ols": 0.45081,
      "m13_15_ols_ci": [
        0.43592,
        0.46638
      ],
      "m12_16_iv": 0.06092,
      "m12_16_iv_ci": [
        -1.1935,
        0.84784
      ],
      "m12_16_ols": 0.21767,
      "m12_16_ols_ci": [
        0.21233,
        0.22365
      ],
      "rho_over_pred": 7.54005,
      "within_over_ols": 3.40449,
      "iv_over_ols": 3.92505
    },
    "nearc2_nearc4_joint": {
      "n": 3010,
      "F": 8.31897,
      "pi": 1,
      "pi_se": 0.24522,
      "rho": 0.15706,
      "rho_se": 0.05089,
      "b_ols": 0.07469,
      "b_ols_se": 0.00365,
      "r2_ols": 0.29984,
      "r2_fs": 0.47756,
      "r2_rf": 0.1957,
      "b_iv": 0.15706,
      "b_iv_se": 0.05255,
      "ar_ci": [
        0.053,
        0.354
      ],
      "ar_p_at_ols": 0.08601,
      "b_rw": 0.06456,
      "gap": 0.08237,
      "weight": -0.01013,
      "within": 0.0925,
      "share_weight": -0.12301,
      "dwh_p": 0.08529,
      "direct": 0.09345,
      "direct_se": 0.04797,
      "direct_p": 0.05139,
      "b_rw_se": 0.00884,
      "weight_se": 0.00796,
      "within_se": 0.06294,
      "gap_se": 0.06175,
      "b_rw_ci": [
        0.04403,
        0.07993
      ],
      "weight_ci": [
        -0.02837,
        0.00312
      ],
      "within_ci": [
        0,
        0.23824
      ],
      "gap_ci": [
        -0.01065,
        0.22364
      ],
      "T1": 0.07469,
      "T1_ci": [
        0.0367,
        0.11162
      ],
      "T1_se": 0.01876,
      "T2": -0.01013,
      "T2_ci": [
        -0.02304,
        0.00281
      ],
      "T2_se": 0.00666,
      "T3": 0.0925,
      "T3_ci": [
        -0.00304,
        0.18229
      ],
      "T3_se": 0.04719,
      "T3_p": 0.04999,
      "weighting_ub_share": 0.03784,
      "weight_ub_over_ols": 0.04173,
      "mde_within": 0.17632,
      "mde_T3": 0.13221,
      "mde_within_over_ols": 2.36063,
      "mde_T3_over_rho": 0.84179,
      "t_iv": 2.98862,
      "tf_crit": null,
      "tv": 0.16307,
      "tv_pct": [
        0.09755,
        0.44337
      ],
      "wald_equal_weights": 16.26929,
      "wald_df": 10,
      "wald_p": 0.09218,
      "slope_range": 0.23777,
      "required_range": 0.50508,
      "required_range_half": 0.25254,
      "required_range_at_tv_p975": 0.18577,
      "required_range_over_ols": 6.76209,
      "max_weight_in_slope_range": 0.03877,
      "max_weight_share_of_gap": 0.47075,
      "max_weight_share_at_tv_p975": 1.27989,
      "required_range_half_at_tv_p975": 0.09289,
      "J": 1.26891,
      "J_df": 1,
      "J_p": 0.25997,
      "ar_interval": true,
      "direct_joint": 5.82083,
      "direct_joint_p": 0.05445,
      "direct_by_instrument": [
        0.03133,
        0.02034
      ]
    },
    "nls_weights": {
      "n": 3010,
      "F": 12.9627,
      "pi": 0.35076,
      "pi_se": 0.09742,
      "rho": 0.05536,
      "rho_se": 0.0201,
      "b_ols": 0.07526,
      "b_ols_se": 0.004,
      "r2_ols": 0.96245,
      "r2_fs": 0.84787,
      "r2_rf": 0.95678,
      "b_iv": 0.15782,
      "b_iv_se": 0.05785,
      "ar_ci": [
        0.05269,
        0.33162
      ],
      "ar_p_at_ols": 0.12097,
      "b_rw": 0.06703,
      "gap": 0.08256,
      "weight": -0.00823,
      "within": 0.09079,
      "share_weight": -0.09972,
      "dwh_p": 0.11936,
      "direct": 0.03217,
      "direct_se": 0.01858,
      "direct_p": 0.08343,
      "b_rw_se": 0.00999,
      "weight_se": 0.00913,
      "within_se": 0.07846,
      "gap_se": 0.07717,
      "b_rw_ci": [
        0.04374,
        0.08474
      ],
      "weight_ci": [
        -0.02846,
        0.00748
      ],
      "within_ci": [
        -0.01175,
        0.25686
      ],
      "gap_ci": [
        -0.02079,
        0.24493
      ],
      "T1": 0.0264,
      "T1_ci": [
        0.01224,
        0.04027
      ],
      "T1_se": 0.0074,
      "T2": -0.00289,
      "T2_ci": [
        -0.00788,
        0.00255
      ],
      "T2_se": 0.00266,
      "T3": 0.03184,
      "T3_ci": [
        -0.006,
        0.06524
      ],
      "T3_se": 0.01826,
      "T3_p": 0.08123,
      "weighting_ub_share": 0.09058,
      "weight_ub_over_ols": 0.09936,
      "mde_within": 0.21982,
      "mde_T3": 0.05117,
      "mde_within_over_ols": 2.92076,
      "mde_T3_over_rho": 0.92433,
      "t_iv": 2.72795,
      "tf_crit": 3.02973,
      "tv": 0.18549,
      "tv_pct": [
        0.10438,
        0.56849
      ],
      "wald_equal_weights": 12.10549,
      "wald_df": 10,
      "wald_p": 0.27806,
      "slope_range": 0.25425,
      "required_range": 0.44508,
      "required_range_half": 0.22254,
      "required_range_at_tv_p975": 0.14522,
      "required_range_over_ols": 5.91372,
      "max_weight_in_slope_range": 0.04716,
      "max_weight_share_of_gap": 0.57125,
      "max_weight_share_at_tv_p975": 1.75079,
      "required_range_half_at_tv_p975": 0.07261
    },
    "coarse_margins": {
      "n": 3010,
      "F": 14.13867,
      "pi": 0.3199,
      "pi_se": 0.08508,
      "rho": 0.04207,
      "rho_se": 0.01752,
      "b_ols": 0.07469,
      "b_ols_se": 0.00365,
      "r2_ols": 0.29984,
      "r2_fs": 0.47712,
      "r2_rf": 0.19469,
      "b_iv": 0.1315,
      "b_iv_se": 0.05414,
      "ar_ci": [
        0.02818,
        0.28115
      ],
      "ar_p_at_ols": 0.27076,
      "b_rw": 0.06983,
      "gap": 0.05681,
      "weight": -0.00487,
      "within": 0.06168,
      "share_weight": -0.08567,
      "dwh_p": 0.27092,
      "direct": 0.01984,
      "direct_se": 0.01652,
      "direct_p": 0.22973,
      "b_rw_se": 0.00747,
      "weight_se": 0.00642,
      "within_se": 0.06908,
      "gap_se": 0.06862,
      "b_rw_ci": [
        0.05344,
        0.08377
      ],
      "weight_ci": [
        -0.01843,
        0.00641
      ],
      "within_ci": [
        -0.04123,
        0.20781
      ],
      "gap_ci": [
        -0.04635,
        0.20091
      ],
      "T1": 0.02389,
      "T1_ci": [
        0.01125,
        0.03566
      ],
      "T1_se": 0.00638,
      "T2": -0.00156,
      "T2_ci": [
        -0.00502,
        0.00147
      ],
      "T2_se": 0.00172,
      "T3": 0.01973,
      "T3_ci": [
        -0.0151,
        0.05022
      ],
      "T3_se": 0.01642,
      "T3_p": 0.22937,
      "weighting_ub_share": 0.1128,
      "weight_ub_over_ols": 0.08579,
      "mde_within": 0.19352,
      "mde_T3": 0.04599,
      "mde_within_over_ols": 2.59092,
      "mde_T3_over_rho": 1.0932,
      "t_iv": 2.4288,
      "tf_crit": 2.9255,
      "tv": 0.0743,
      "tv_pct": [
        0.04959,
        0.3733
      ],
      "wald_equal_weights": 3.75179,
      "wald_df": 4,
      "wald_p": 0.44064,
      "slope_range": 0.13704,
      "required_range": 0.76463,
      "required_range_half": 0.38231,
      "required_range_at_tv_p975": 0.15218,
      "required_range_over_ols": 10.23692,
      "max_weight_in_slope_range": 0.01018,
      "max_weight_share_of_gap": 0.17923,
      "max_weight_share_at_tv_p975": 0.90052,
      "required_range_half_at_tv_p975": 0.07609,
      "margins": {
        "label": [
          "1-11 (per year)",
          "12",
          "13-15 (per year)",
          "16",
          "17-18 (per year)"
        ],
        "n_at_level": [
          497,
          992,
          704,
          459,
          358
        ],
        "slope": [
          0.04905,
          0.18109,
          0.04405,
          0.15174,
          0.07183
        ],
        "slope_se": [
          0.01,
          0.02575,
          0.00948,
          0.02754,
          0.01596
        ],
        "w_ols": [
          0.18739,
          0.07708,
          0.45081,
          0.14058,
          0.14414
        ],
        "w_ols_se": [
          0.01111,
          0.0028,
          0.00779,
          0.00302,
          0.00606
        ],
        "w_iv": [
          0.19187,
          0.08334,
          0.51437,
          0.0945,
          0.11592
        ],
        "w_iv_se": [
          0.11016,
          0.04324,
          0.10694,
          0.04314,
          0.06358
        ],
        "w_iv_ci": [
          [
            -0.04725,
            0.40061
          ],
          [
            0.00456,
            0.17515
          ],
          [
            0.33511,
            0.76806
          ],
          [
            -0.00373,
            0.15791
          ],
          [
            -0.02787,
            0.23324
          ]
        ],
        "w_ols_ci": [
          [
            0.16616,
            0.2089
          ],
          [
            0.07183,
            0.08269
          ],
          [
            0.43592,
            0.46638
          ],
          [
            0.13466,
            0.14626
          ],
          [
            0.13206,
            0.15549
          ]
        ],
        "pi_k": [
          0.06138,
          0.02666,
          0.16455,
          0.03023,
          0.03708
        ],
        "pi_k_se": [
          0.03507,
          0.01371,
          0.04605,
          0.01624,
          0.02093
        ]
      },
      "rho_over_pred": 1.8833,
      "within_over_ols": 0.82575,
      "iv_over_ols": 1.76059
    }
  },
  "covariate_cells": {
    "pooled_coarse": {
      "n": 3010,
      "F": 13.83006,
      "pi": 0.31641,
      "pi_se": 0.08508,
      "rho": 0.04187,
      "rho_se": 0.01754,
      "b_ols": 0.07447,
      "b_ols_se": 0.00365,
      "r2_ols": 0.30257,
      "r2_fs": 0.47755,
      "r2_rf": 0.19813,
      "b_iv": 0.13234,
      "b_iv_se": 0.05477,
      "ar_ci": [
        0.02777,
        0.28493
      ],
      "ar_p_at_ols": 0.26717,
      "b_rw": 0.06964,
      "gap": 0.05787,
      "weight": -0.00484,
      "within": 0.0627,
      "share_weight": -0.08357,
      "dwh_p": 0.26725,
      "direct": 0.01995,
      "direct_se": 0.0165,
      "direct_p": 0.2268,
      "b_rw_se": 0.00767,
      "weight_se": 0.00663,
      "within_se": 0.06818,
      "gap_se": 0.06779,
      "b_rw_ci": [
        0.053,
        0.08385
      ],
      "weight_ci": [
        -0.01868,
        0.00666
      ],
      "within_ci": [
        -0.04068,
        0.21109
      ],
      "gap_ci": [
        -0.04587,
        0.20426
      ],
      "T1": 0.02356,
      "T1_ci": [
        0.0108,
        0.03562
      ],
      "T1_se": 0.00636,
      "T2": -0.00153,
      "T2_ci": [
        -0.0051,
        0.00156
      ],
      "T2_se": 0.00173,
      "T3": 0.01984,
      "T3_ci": [
        -0.01489,
        0.05006
      ],
      "T3_se": 0.01631,
      "T3_p": 0.22373,
      "weighting_ub_share": 0.11512,
      "weight_ub_over_ols": 0.08944,
      "mde_within": 0.19101,
      "mde_T3": 0.04568,
      "mde_within_over_ols": 2.56477,
      "mde_T3_over_rho": 1.09099,
      "t_iv": 2.41622,
      "tf_crit": 2.95027,
      "tv": 0.08312,
      "tv_pct": [
        0.05172,
        0.39575
      ],
      "wald_equal_weights": 3.96718,
      "wald_df": 4,
      "wald_p": 0.41047,
      "slope_range": 0.13574,
      "required_range": 0.69618,
      "required_range_half": 0.34809,
      "required_range_at_tv_p975": 0.14622,
      "required_range_over_ols": 9.34802,
      "max_weight_in_slope_range": 0.01128,
      "max_weight_share_of_gap": 0.19497,
      "max_weight_share_at_tv_p975": 0.9283,
      "required_range_half_at_tv_p975": 0.07311
    },
    "cells_coarse": {
      "n": 3010,
      "F": 13.83006,
      "pi": 0.31641,
      "pi_se": 0.08508,
      "rho": 0.04187,
      "rho_se": 0.01754,
      "b_ols": 0.07447,
      "b_ols_se": 0.00365,
      "r2_ols": 0.30257,
      "r2_fs": 0.47755,
      "r2_rf": 0.19813,
      "b_iv": 0.13234,
      "b_iv_se": 0.05477,
      "ar_ci": [
        0.02777,
        0.28493
      ],
      "ar_p_at_ols": 0.26717,
      "b_rw": 0.05419,
      "gap": 0.05787,
      "weight": -0.02028,
      "within": 0.07815,
      "share_weight": -0.35053,
      "dwh_p": 0.26725,
      "direct": 0.02528,
      "direct_se": 0.01652,
      "direct_p": 0.12598,
      "b_rw_se": 0.01651,
      "weight_se": 0.01614,
      "within_se": 0.07073,
      "gap_se": 0.06779,
      "cov": 0.00142,
      "cov_se": 0.00575,
      "mar": -0.0217,
      "mar_se": 0.01501,
      "C2": 0.00045,
      "C2_ci": [
        -0.0027,
        0.00361
      ],
      "M2": -0.00687,
      "M2_ci": [
        -0.01359,
        0.00007
      ],
      "cells": {
        "cell": [
          0,
          1,
          2,
          3,
          5,
          6,
          7
        ],
        "n": [
          387,
          1202,
          254,
          464,
          206,
          218,
          279
        ],
        "w_ols": [
          0.10901,
          0.40377,
          0.10884,
          0.16297,
          0.06034,
          0.07929,
          0.07578
        ],
        "w_iv": [
          0.01011,
          0.59736,
          -0.06143,
          0.16849,
          0.05699,
          -0.0094,
          0.23789
        ],
        "w_iv_se": [
          0.15992,
          0.21996,
          0.17643,
          0.16459,
          0.1008,
          0.14862,
          0.16399
        ],
        "slope": [
          0.06102,
          0.06778,
          0.08187,
          0.07961,
          0.08368,
          0.08056,
          0.09413
        ]
      },
      "b_rw_ci": [
        0.01423,
        0.07905
      ],
      "weight_ci": [
        -0.0591,
        0.00351
      ],
      "within_ci": [
        -0.02073,
        0.23471
      ],
      "gap_ci": [
        -0.04587,
        0.20426
      ],
      "cov_ci": [
        -0.0102,
        0.01279
      ],
      "mar_ci": [
        -0.05823,
        0.00003
      ],
      "T1": 0.02356,
      "T1_ci": [
        0.0108,
        0.03562
      ],
      "T1_se": 0.00636,
      "T2": -0.00642,
      "T2_ci": [
        -0.01383,
        0.00116
      ],
      "T2_se": 0.00383,
      "T3": 0.02473,
      "T3_ci": [
        -0.00782,
        0.05369
      ],
      "T3_se": 0.01587,
      "T3_p": 0.11914,
      "weighting_ub_share": 0.06066,
      "weight_ub_over_ols": 0.04713,
      "mde_within": 0.19816,
      "mde_T3": 0.04445,
      "mde_within_over_ols": 2.66082,
      "mde_T3_over_rho": 1.06162,
      "t_iv": 2.41622,
      "tf_crit": 2.95027,
      "tv": 0.64601,
      "tv_pct": [
        0.46825,
        1.76933
      ],
      "wald_equal_weights": 55.41247,
      "wald_df": 34,
      "wald_p": 0.01162,
      "slope_range": 0.3266,
      "required_range": 0.08957,
      "required_range_half": 0.04479,
      "required_range_at_tv_p975": 0.03271,
      "required_range_over_ols": 1.20277,
      "max_weight_in_slope_range": 0.21098,
      "max_weight_share_of_gap": 3.64606,
      "max_weight_share_at_tv_p975": 9.98603,
      "required_range_half_at_tv_p975": 0.01635
    }
  },
  "software": {
    "python": "3.12.13",
    "numpy": "2.5.2",
    "pandas": "3.0.5",
    "scipy": "1.18.1"
  },
  "weighting_ub_below_half_gap": true,
  "profile_free_bound_exceeds_half": true,
  "hypothesis_supported": false
}
Experiment script · experiment_v1.py
"""Is the college-proximity IV-OLS gap in the return to schooling an Angrist-Imbens margin-weighting artefact?
Real data: Card (1995) NLSYM extract, 3,010 men, 1976 wages. Exact in-sample decomposition
  b_IV - b_OLS = sum_k (w_k - m_k) d_k   [margin weighting]   +   (b_IV - b_RW)   [within-margin residual]
d_k : per-margin slopes from a schooling-dummy OLS wage equation (years 1-8 pooled linearly, 1{S>=j} for j=9..18)
m_k : OLS weights (coef of S in a regression of component k on S and W); sum to one, give b_OLS = sum m_k d_k
w_k : Angrist-Imbens IV weights pi_k / pi (coef of Z in a regression of component k on Z and W); sum to one
b_RW: IV-weighted OLS = sum w_k d_k.  In reduced-form units: rho = pi*b_OLS + pi*(b_RW-b_OLS) + rho_e.
Revision: nearc2 alone and jointly with nearc4 (Hansen J, 2-df AR set, 2-df direct-effect test); a three-way split
with margin slopes inside race x South x SMSA cells (cell weights, within-cell margin weights, remainder); the
family-background split on men with both parents' schooling observed; NLS sampling weights; MDEs; tF critical value."""
import json
import platform
import numpy as np
import pandas as pd
import scipy
from scipy import stats
from dataset import load

SEED, B_MAIN, B_ROB, CRIT = 20261024, 1999, 999, float(stats.chi2.ppf(0.95, 1))
KMDE = float(stats.norm.ppf(0.975) + stats.norm.ppf(0.80))          # minimum detectable effect: 80% power, 5% test
df = load()
REG = [f"reg66{i}" for i in range(2, 10)]                 # 1966 region of residence, reg661 omitted
BASE = ["exper", "expersq", "black", "south", "smsa", "smsa66"] + REG   # Card (1995) main specification
need = ["lwage", "educ", "nearc4", "nearc2", "age", "enroll", "weight"] + BASE
print("loaded", df.shape, "| missing in analysis variables:", int(df[need].isna().sum().sum()))
assert df[need].notna().all().all()
for c in ("fatheduc", "motheduc"):                      # family background with missing-value indicators
    df[c + "_miss"] = df[c].isna().astype(float)
    df[c + "_f"] = df[c].fillna(0.0)
FAM = ["fatheduc_f", "fatheduc_miss", "motheduc_f", "motheduc_miss", "momdad14", "sinmom14"]
df["agesq"] = df["age"] ** 2
AGE = ["age", "agesq", "black", "south", "smsa", "smsa66"] + REG       # predetermined controls only
df["hiback"] = (np.fmax(df["fatheduc"].fillna(-1), df["motheduc"].fillna(-1)) >= 12).astype(int)
df["both_obs"] = (df["fatheduc"].notna() & df["motheduc"].notna()).astype(int)
cell = 4 * df["black"] + 2 * df["south"] + df["smsa"]    # race x South x SMSA in 1976
df["cell"] = cell.where(cell != 4, 5)                    # the 5 Black men outside both join Black, non-South, SMSA
GLIST = [0, 1, 2, 3, 5, 6, 7]
for g in GLIST[1:]:
    df[f"cell{g}"] = (df["cell"] == g).astype(float)
WCELL = ["exper", "expersq", "smsa66"] + REG + [f"cell{g}" for g in GLIST[1:]]   # cells saturated in the controls
# robustness: (name, sample filter, controls, years pooled linearly); heterogeneity: parental education 12+ (either parent)
ROBUST = [("pool_1_11", None, BASE, 11), ("family_background", None, BASE + FAM, 8), ("age_controls", None, AGE, 8),
          ("not_enrolled", "enroll == 0", BASE, 8)]
HETERO = [("high_background", "hiback == 1"), ("low_background", "hiback == 0"),
          ("high_background_obs", "both_obs == 1 and hiback == 1"), ("low_background_obs", "both_obs == 1 and hiback == 0")]
# Lee et al. (2022, Table 3 Panel A): first-stage F and the 5 percent tF critical value for |t|, linearly interpolated
TF_F = [8.473, 8.773, 9.098, 9.451, 9.835, 10.253, 10.711, 11.214, 11.766, 12.374, 13.048, 13.796, 14.631, 15.566, 16.618,
        17.810, 19.167, 20.721, 22.516, 24.605, 27.058, 29.967, 33.457, 37.699, 42.930, 49.495, 57.902, 68.930, 83.823, 104.67]
TF_C = [3.793, 3.707, 3.624, 3.542, 3.463, 3.385, 3.309, 3.234, 3.161, 3.090, 3.021, 2.953, 2.886, 2.821, 2.758, 2.696,
        2.635, 2.576, 2.518, 2.461, 2.406, 2.352, 2.299, 2.247, 2.197, 2.147, 2.099, 2.052, 2.006, 1.96]


def tf_crit(F):
    return None if F < TF_F[0] else float(np.interp(F, TF_F, TF_C))


def ols(X, y):
    """OLS coefficients, HC1 covariance and residuals."""
    XtXi = np.linalg.pinv(X.T @ X)
    b = XtXi @ (X.T @ y)
    e = y - X @ b
    n, k = X.shape
    return b, XtXi @ (X.T * e ** 2) @ X @ XtXi * n / (n - k), e


def comps(s, lo):
    return np.column_stack([np.minimum(s, lo)] + [(s >= j).astype(float) for j in range(lo + 1, 19)])


def coarse(s):      # years 1-11 per year, year 12, years 13-15 per year, year 16, years 17-18 per year
    return np.column_stack([np.minimum(s, 11), (s >= 12) * 1.0, np.clip(s - 12, 0, 3), (s >= 16) * 1.0, np.maximum(s - 16, 0)])


def core(y, s, z, W, A):
    """Point estimates of every building block for one (re)sample."""
    G = np.linalg.lstsq(np.column_stack([z, W]), np.column_stack([y, s, A]), rcond=None)[0][0]
    H = np.linalg.lstsq(np.column_stack([s, W]), np.column_stack([y, A]), rcond=None)[0][0]
    d = np.linalg.lstsq(np.column_stack([A, W]), y, rcond=None)[0][: A.shape[1]]
    return {"rho": G[0], "pi": G[1], "pik": G[2:], "b_ols": H[0], "mu": H[1:], "d": d, "rho_pred": G[2:] @ d}


def derive(c, groups=None):
    pi = c["pi"]
    out = {"b_ols": c["b_ols"], "b_iv": c["rho"] / pi, "b_rw": c["rho_pred"] / pi, "pi": pi, "rho": c["rho"],
           "w": c["pik"] / pi, "mu": c["mu"], "d": c["d"]}
    out["gap"], out["weight"], out["within"] = out["b_iv"] - out["b_ols"], out["b_rw"] - out["b_ols"], out["b_iv"] - out["b_rw"]
    out["T1"], out["T2"], out["T3"] = pi * c["b_ols"], c["rho_pred"] - pi * c["b_ols"], c["rho"] - c["rho_pred"]
    out["tv"] = 0.5 * np.abs(out["w"] - out["mu"]).sum()       # any profile b: |sum (w-m) b| <= tv * (max b - min b)
    out["g"] = c["pik"] - pi * c["mu"]                           # reduced-form weight gap; zero iff IV and OLS weights coincide
    if groups is not None:                                       # weighting = cell weights + within-cell margin weights
        Wols = np.array([c["mu"][groups == g].sum() for g in GLIST])
        D = np.array([c["mu"][groups == g] @ c["d"][groups == g] for g in GLIST]) / Wols
        out["Wiv"], out["Wols"], out["D"] = np.array([out["w"][groups == g].sum() for g in GLIST]), Wols, D
        out["cov"] = (out["Wiv"] - Wols) @ D
        out["mar"] = out["weight"] - out["cov"]
        out["C2"], out["M2"] = pi * out["cov"], pi * out["mar"]
    return out


def fieller(num, den, V):
    """95% Fieller/Anderson-Rubin set for num/den given Cov(num, den); None if unbounded."""
    a = den ** 2 - CRIT * V[1, 1]
    if a <= 0:
        return None
    bq, cq = -2 * (num * den - CRIT * V[0, 1]), num ** 2 - CRIT * V[0, 0]
    r = np.sqrt(max(bq ** 2 - 4 * a * cq, 0.0))
    return [float((-bq - r) / (2 * a)), float((-bq + r) / (2 * a))]


def wald(b, V, L):
    st = float(b @ np.linalg.solve(V, b))
    return st, float(stats.chi2.sf(st, L))


def hansen_j(y, X, Zf):
    """Heteroskedasticity-robust Hansen J for over-identified 2SLS (two-step efficient GMM)."""
    n = len(y)
    b1 = np.linalg.lstsq(Zf @ np.linalg.lstsq(Zf, X, rcond=None)[0], y, rcond=None)[0]
    S = (Zf * ((y - X @ b1) ** 2)[:, None]).T @ Zf / n
    Si, AZ = np.linalg.pinv(S), X.T @ Zf / n
    b2 = np.linalg.solve(AZ @ Si @ AZ.T, AZ @ Si @ (Zf.T @ y / n))
    gbar = Zf.T @ (y - X @ b2) / n
    J = float(n * gbar @ Si @ gbar)
    return {"J": J, "J_df": Zf.shape[1] - X.shape[1], "J_p": float(stats.chi2.sf(J, Zf.shape[1] - X.shape[1]))}


def ar_grid(y, s, QF, L, lo=-1.0, hi=2.0, step=0.001):
    """Anderson-Rubin set with L instruments: robust Wald test of the L proximity coefficients in y - b0*s."""
    grid = np.round(np.arange(lo, hi + step / 2, step), 6)
    n, k = QF.shape
    H = np.linalg.pinv(QF.T @ QF) @ QF.T
    stat = np.empty(len(grid))
    for i0 in range(0, len(grid), 500):
        G = grid[i0:i0 + 500]
        E = y[:, None] - s[:, None] * G[None, :]
        Bc = H @ E
        V = np.einsum("ai,bi,ig->gab", H[:L], H[:L], (E - QF @ Bc) ** 2) * n / (n - k)
        stat[i0:i0 + len(G)] = np.einsum("ga,gab,gb->g", Bc[:L].T, np.linalg.inv(V), Bc[:L].T)
    acc = np.where(stat <= stats.chi2.ppf(0.95, L))[0]
    if len(acc) == 0 or acc[0] == 0 or acc[-1] == len(grid) - 1:
        return None, False
    return [float(grid[acc[0]]), float(grid[acc[-1]])], bool(acc[-1] - acc[0] + 1 == len(acc))


def spec(d, X, Zc=("nearc4",), lo=8, A=None, groups=None, wts=None, B=B_ROB, seed=SEED, detail=False, blocks=None):
    y, s = d["lwage"].to_numpy(float), d["educ"].to_numpy(float)
    Z, W = d[list(Zc)].to_numpy(float), np.column_stack([np.ones(len(d)), d[X].to_numpy(float)])
    A, n, L = (comps(s, lo) if A is None else A), len(d), len(Zc)
    if wts is not None:                                   # NLS sampling weights: weighted least squares throughout
        r = np.sqrt(wts / wts.mean())
        y, s, Z, W, A = y * r, s * r, Z * r[:, None], W * r[:, None], A * r[:, None]
    extra, kz = {}, A.shape[1]
    if L > 1:      # joint design: robust F, Hansen J, AR set, 2-df direct effect; then IV with the fitted first-stage index
        QF = np.column_stack([Z, W])
        gF, VgF, _ = ols(QF, s)
        extra["F_joint"] = float(gF[:L] @ np.linalg.solve(VgF[:L, :L], gF[:L]) / L)
        extra.update(hansen_j(y, np.column_stack([s, W]), QF))
        extra["ar_ci_joint"], extra["ar_interval"] = ar_grid(y, s, QF, L)
        bdj, Vdj, _ = ols(np.column_stack([A, Z, W]), y)
        extra["direct_joint"], extra["direct_joint_p"] = wald(bdj[kz:kz + L], Vdj[kz:kz + L, kz:kz + L], L)
        extra["direct_by_instrument"] = [float(v) for v in bdj[kz:kz + L]]
        z = Z @ gF[:L]                                    # composite instrument, weights fixed at the full-sample first stage
    else:
        z = Z[:, 0]
    Q = np.column_stack([z, W])
    bo, Vo, eo = ols(np.column_stack([s, W]), y)
    fs, Vf, es = ols(Q, s)
    rf, Vr, ey = ols(Q, y)
    r2 = lambda e, v: float(1 - e @ e / ((v - v.mean()) @ (v - v.mean())))
    zt = z - W @ np.linalg.lstsq(W, z, rcond=None)[0]
    Vrp = np.array([[Vr[0, 0], 0.0], [0.0, Vf[0, 0]]])
    Vrp[0, 1] = Vrp[1, 0] = np.sum(zt ** 2 * ey * es) / np.sum(zt ** 2) ** 2 * n / (n - Q.shape[1])
    Xh = Q @ np.linalg.lstsq(Q, np.column_stack([s, W]), rcond=None)[0]          # 2SLS with HC1
    b2 = np.linalg.lstsq(Xh, y, rcond=None)[0]
    u = y - np.column_stack([s, W]) @ b2
    Xi = np.linalg.pinv(Xh.T @ Xh)
    se_iv = float(np.sqrt((Xi @ (Xh.T * u ** 2) @ Xh @ Xi)[0, 0] * n / (n - Xh.shape[1])))
    bc, Vc, _ = ols(np.column_stack([s, W, es]), y)                              # control-function DWH
    bd, Vd, _ = ols(np.column_stack([A, z, W]), y)                               # direct effect of Z given schooling dummies
    pt = derive(core(y, s, z, W, A), groups)
    rng = np.random.default_rng(seed)
    bs = [derive(core(y[i], s[i], z[i], W[i], A[i]), groups) for i in (rng.integers(0, n, n) for _ in range(B))]
    arr = {k: np.array([b[k] for b in bs]) for k in bs[0]}
    res = {"n": n, "F": float(fs[0] ** 2 / Vf[0, 0]), "pi": float(fs[0]), "pi_se": float(np.sqrt(Vf[0, 0])),
           "rho": float(rf[0]), "rho_se": float(np.sqrt(Vr[0, 0])), "b_ols": float(bo[0]), "b_ols_se": float(np.sqrt(Vo[0, 0])),
           "r2_ols": r2(eo, y), "r2_fs": r2(es, s), "r2_rf": r2(ey, y),
           "b_iv": float(pt["b_iv"]), "b_iv_se": se_iv, "ar_ci": fieller(rf[0], fs[0], Vrp),
           "ar_p_at_ols": float(stats.chi2.sf((rf[0] - bo[0] * fs[0]) ** 2 / (Vrp[0, 0] - 2 * bo[0] * Vrp[0, 1] + bo[0] ** 2 * Vrp[1, 1]), 1)),
           "b_rw": float(pt["b_rw"]), "gap": float(pt["gap"]), "weight": float(pt["weight"]), "within": float(pt["within"]),
           "share_weight": float(pt["weight"] / pt["gap"]), "dwh_p": float(2 * stats.norm.sf(abs(bc[-1]) / np.sqrt(Vc[-1, -1]))),
           "direct": float(bd[kz]), "direct_se": float(np.sqrt(Vd[kz, kz])), "direct_p": float(2 * stats.norm.sf(abs(bd[kz]) / np.sqrt(Vd[kz, kz])))}
    for k in ("b_rw", "weight", "within", "gap"):
        res[k + "_se"] = float(arr[k].std(ddof=1))
    num = {"b_rw": pt["b_rw"] * pt["pi"], "weight": pt["T2"], "within": pt["T3"], "gap": pt["T2"] + pt["T3"]}
    numb = {"b_rw": arr["b_rw"] * arr["pi"], "weight": arr["T2"], "within": arr["T3"], "gap": arr["T2"] + arr["T3"]}
    if groups is not None:
        num.update({"cov": pt["C2"], "mar": pt["M2"]})
        numb.update({"cov": arr["C2"], "mar": arr["M2"]})
        for k in ("cov", "mar"):
            res[k], res[k + "_se"] = float(pt[k]), float(arr[k].std(ddof=1))
        for k in ("C2", "M2"):
            res[k], res[k + "_ci"] = float(pt[k]), np.percentile(arr[k], [2.5, 97.5]).tolist()
        res["cells"] = {"cell": GLIST, "n": [int((d["cell"] == g).sum()) for g in GLIST], "w_ols": pt["Wols"].tolist(),
                        "w_iv": pt["Wiv"].tolist(), "w_iv_se": arr["Wiv"].std(0, ddof=1).tolist(), "slope": pt["D"].tolist()}
    for k in num:
        res[k + "_ci"] = fieller(num[k], pt["pi"], np.cov(np.vstack([numb[k], arr["pi"]])))
    for k in ("T1", "T2", "T3"):
        res[k], res[k + "_ci"] = float(pt[k]), [float(np.percentile(arr[k], 2.5)), float(np.percentile(arr[k], 97.5))]
        res[k + "_se"] = float(arr[k].std(ddof=1))
    res["T3_p"] = float(2 * stats.norm.sf(abs(pt["T3"]) / res["T3_se"]))
    res["weighting_ub_share"] = None if res["weight_ci"] is None else float(res["weight_ci"][1] / pt["gap"])
    res["weight_ub_over_ols"] = None if res["weight_ci"] is None else float(res["weight_ci"][1] / bo[0])
    res["mde_within"], res["mde_T3"] = KMDE * res["within_se"], KMDE * res["T3_se"]
    res["mde_within_over_ols"], res["mde_T3_over_rho"] = res["mde_within"] / float(bo[0]), res["mde_T3"] / float(rf[0])
    res["t_iv"] = float(pt["b_iv"] / se_iv)
    res["tf_crit"] = tf_crit(res["F"]) if L == 1 else None
    res["tv"], res["tv_pct"] = float(pt["tv"]), np.percentile(arr["tv"], [2.5, 97.5]).tolist()   # bootstrap range; tv is nonsmooth
    Vg = np.cov(arr["g"][:, :-1].T)                                            # sum_k g_k = 0: drop one
    res["wald_equal_weights"] = float(pt["g"][:-1] @ np.linalg.solve(Vg, pt["g"][:-1]))
    res["wald_df"] = int(len(pt["g"]) - 1)
    res["wald_p"] = float(stats.chi2.sf(res["wald_equal_weights"], res["wald_df"]))
    res["slope_range"] = float(pt["d"].max() - pt["d"].min())
    res["required_range"] = float(pt["gap"] / pt["tv"])                 # smallest return range that weighting needs
    res["required_range_half"] = float(pt["gap"] / (2 * pt["tv"]))      # ... for half the gap
    res["required_range_at_tv_p975"] = float(pt["gap"] / res["tv_pct"][1])
    res["required_range_over_ols"] = float(res["required_range"] / pt["b_ols"])
    res["max_weight_in_slope_range"] = float(pt["tv"] * res["slope_range"])
    res["max_weight_share_of_gap"] = float(res["max_weight_in_slope_range"] / pt["gap"])
    res["max_weight_share_at_tv_p975"] = float(res["tv_pct"][1] * res["slope_range"] / pt["gap"])
    res["required_range_half_at_tv_p975"] = float(pt["gap"] / (2 * res["tv_pct"][1]))
    res.update(extra)
    if L > 1:
        res["F"], res["ar_ci"] = res.pop("F_joint"), res.pop("ar_ci_joint")
    if detail:
        _, Vs, _ = ols(np.column_stack([A, W]), y)
        pik_se = [float(np.sqrt(ols(Q, A[:, j])[1][0, 0])) for j in range(kz)]
        if blocks is None:                                # fine margins; otherwise labelled coarse blocks (label, lo, hi)
            labels = [f"1-{lo} (per year)"] + [str(j) for j in range(lo + 1, 19)]
            nlev = [int((s <= lo).sum())] + [int((s == j).sum()) for j in range(lo + 1, 19)]
        else:
            labels, nlev = [b[0] for b in blocks], [int(((s >= b[1]) & (s <= b[2])).sum()) for b in blocks]
        res["margins"] = {"label": labels, "n_at_level": nlev,
                          "slope": pt["d"].tolist(), "slope_se": np.sqrt(np.diag(Vs)[:kz]).tolist(),
                          "w_ols": pt["mu"].tolist(), "w_ols_se": arr["mu"].std(0, ddof=1).tolist(),
                          "w_iv": pt["w"].tolist(), "w_iv_se": arr["w"].std(0, ddof=1).tolist(),
                          "w_iv_ci": np.percentile(arr["w"], [2.5, 97.5], axis=0).T.tolist(),
                          "w_ols_ci": np.percentile(arr["mu"], [2.5, 97.5], axis=0).T.tolist(),
                          "pi_k": (pt["w"] * pt["pi"]).tolist(), "pi_k_se": pik_se}
        wblocks = {"m13_15": [13 - lo, 14 - lo, 15 - lo], "m12_16": [12 - lo, 16 - lo]} if blocks is None else {}
        for nm, ix in wblocks.items():                    # component index of margin j is j - lo
            for wk, key in (("w", "iv"), ("mu", "ols")):
                res[f"{nm}_{key}"] = float(pt[wk][ix].sum())
                res[f"{nm}_{key}_ci"] = np.percentile(arr[wk][:, ix].sum(1), [2.5, 97.5]).tolist()
        res["rho_over_pred"] = float(pt["rho"] / (pt["T1"] + pt["T2"]))
        res["within_over_ols"] = float(pt["within"] / pt["b_ols"])
        res["iv_over_ols"] = float(pt["b_iv"] / pt["b_ols"])
    return res


print("main specification ...")
main = spec(df, BASE, lo=8, B=B_MAIN, detail=True)
s_all, z_all = df["educ"].to_numpy(float), df["nearc4"].to_numpy(float)
W_all = np.column_stack([np.ones(len(df)), df[BASE].to_numpy(float)])
mono = []
for j in range(2, 19):                                    # Angrist-Imbens monotonicity: CDF shifts >= 0 at every threshold
    b, V, _ = ols(np.column_stack([z_all, W_all]), (s_all >= j).astype(float))
    mono.append((j, float(b[0]), float(b[0] / np.sqrt(V[0, 0]))))
balance = {}
for v in ("fatheduc", "motheduc", "libcrd14", "momdad14"):  # predetermined background on proximity, given W
    m = df[v].notna().to_numpy()
    b, V, _ = ols(np.column_stack([z_all, W_all])[m], df.loc[m, v].to_numpy(float))
    balance[v] = {"n": int(m.sum()), "mean": float(df.loc[m, v].mean()), "coef": float(b[0]), "se": float(np.sqrt(V[0, 0])),
                  "p": float(2 * stats.norm.sf(abs(b[0]) / np.sqrt(V[0, 0])))}
summ = {}
for g, d in (("all", df), ("near", df[df.nearc4 == 1]), ("far", df[df.nearc4 == 0])):
    summ[g] = {"n": int(len(d)), "wage_dollars": float(d["wage"].mean() / 100), "lwage": float(d["lwage"].mean()),
               "educ": float(d["educ"].mean()), "educ_sd": float(d["educ"].std()), "exper": float(d["exper"].mean()),
               "age": float(d["age"].mean()), "black": float(d["black"].mean()), "south66": float(d["south66"].mean()),
               "smsa66": float(d["smsa66"].mean()), "enroll": float(d["enroll"].mean()),
               "share_educ12": float((d["educ"] == 12).mean()), "share_educ16plus": float((d["educ"] >= 16).mean())}
print("robustness and heterogeneity ...")
rob = {nm: spec(df if q is None else df.query(q), X, lo=lo) for nm, q, X, lo in ROBUST}
het = {nm: spec(df.query(q), BASE) for nm, q in HETERO}
low = df["hiback"] == 0
het_counts = {"low": int(low.sum()), "low_any_missing": int((low & (df.fatheduc.isna() | df.motheduc.isna())).sum()),
              "low_both_missing": int((low & df.fatheduc.isna() & df.motheduc.isna()).sum()), "both_obs": int(df.both_obs.sum())}
print("second instrument, sampling weights and covariate cells ...")
s_c = df["educ"].to_numpy(float)
cellA = np.column_stack([coarse(s_c)[:, k] * (df["cell"] == g).to_numpy() for g in GLIST for k in range(5)])
groups = np.repeat(GLIST, 5)
CBLOCKS = [("1-11 (per year)", 1, 11), ("12", 12, 12), ("13-15 (per year)", 13, 15), ("16", 16, 16), ("17-18 (per year)", 17, 18)]
other = {"nearc2_alone": spec(df, BASE, Zc=("nearc2",), detail=True), "nearc2_nearc4_joint": spec(df, BASE, Zc=("nearc2", "nearc4")),
         "nls_weights": spec(df, BASE, wts=df["weight"].to_numpy(float)), "coarse_margins": spec(df, BASE, A=coarse(s_c), detail=True, blocks=CBLOCKS)}
cells = {"pooled_coarse": spec(df, WCELL, A=coarse(s_c)), "cells_coarse": spec(df, WCELL, A=cellA, groups=groups)}
supported = bool(main["share_weight"] >= 0.5 and main["within_ci"] is not None and main["within_ci"][0] <= 0 <= main["within_ci"][1])


def clean(x):
    if isinstance(x, dict):
        return {k: clean(v) for k, v in x.items()}
    if isinstance(x, (list, tuple)):
        return [clean(v) for v in x]
    if isinstance(x, (bool, np.bool_)):
        return bool(x)
    if isinstance(x, (int, np.integer)):
        return int(x)
    if isinstance(x, (float, np.floating)):
        return round(float(x), 5) if np.isfinite(x) else None
    return x


result = {"dataset": "card_college_proximity", "n_obs": int(len(df)), "seed": SEED, "bootstrap_reps_main": B_MAIN,
          "bootstrap_reps_other": B_ROB, "chi2_crit_95": CRIT, "mde_factor": KMDE, "main": main,
          "monotonicity": {"min_t": min(t for _, _, t in mono), "n_negative": sum(b < 0 for _, b, _ in mono),
                           "thresholds": [[j, b, t] for j, b, t in mono]},
          "balance": balance, "summary": summ, "robustness": rob, "heterogeneity": het, "heterogeneity_counts": het_counts,
          "second_instrument_and_weights": other, "covariate_cells": cells,
          "software": {"python": platform.python_version(), "numpy": np.__version__, "pandas": pd.__version__, "scipy": scipy.__version__},
          "weighting_ub_below_half_gap": bool(main["weight_ci"] is not None and main["weight_ci"][1] < 0.5 * main["gap"]),
          "profile_free_bound_exceeds_half": bool(main["max_weight_share_of_gap"] > 0.5),
          "hypothesis_supported": supported}
print(f"OLS {main['b_ols']:.4f}  IV {main['b_iv']:.4f}  IV-weighted OLS {main['b_rw']:.4f}  share explained {main['share_weight']:.3f}")
print("RESULT_JSON: " + json.dumps(clean(result)))
Listing. experiment_v1.py

4.Analysis

Attempt 4Done

Analysis complete: experiment succeeded; hypothesis refuted; 25 key numbers.

Reviewer feedback applied: Editorial alignment after acceptance (no new referee request): make the analysis record agree with the accepted manuscript. The two-instrument over-identification test 'does not reject' (Hansen J p = 0.260; low power with a weak second instrument), not 'passes'; check every other statement in the analysis record against the final paper. The manuscript itself is unchanged: resubmit it as accepted.
Hypothesis refutedExperiment succeeded

Results and analysis

The extended run succeeded on the first attempt (19.8 s) and reproduces every earlier number exactly. Relative to the observational benchmark the margin-weighting hypothesis is refuted, and the new analyses leave that verdict unchanged. At the observational slopes, weighting contributes at most 0.0095 log points per year (95 percent upper end, 0.13 of the OLS return); the point estimate is -0.0053. Saturating race x South x SMSA cells, covariate-cell weighting is 0.0014 [-0.0102, 0.0128] and within-cell margin weighting -0.0217. Two-year-college proximity (nearc2, F 2.43) raises log wages by 0.0357, while its schooling shifts, valued at the observational slopes, predict 0.0091 at the OLS weights and 0.0044 less at its own; the reduced-form remainder, 0.0309 [0.0048, 0.0588], rejects the within-margin null (p = 0.027). Its first-stage shifts are negative at thresholds 11 and 12 (-0.0174, t = -1.73; -0.0136), so the testable implication of monotonicity is not clearly satisfied, and the rejection may reflect high complier returns, a direct effect of proximity, noise in so weak a first stage, or a failure of monotonicity. With nearc2 and nearc4 used jointly, the over-identification test does not reject (Hansen J p = 0.260), though it has little power with so weak a second instrument, and the 2-df direct-effect test is borderline (p = 0.054). Three qualifications are essential. First, the gap itself is not established (Hausman p = 0.271), and the within-margin term has no power: its minimum detectable effect, 0.1872, is 2.51 times OLS. Second, the profile-free bound is inconclusive. At fine margins, returns confined to the observed slope range could deliver 0.561 of the point gap, above the 0.5 threshold. At five coarse blocks the share is 0.179 at the point weights but 0.901 at the 97.5th percentile of the weight distance. Third, every IV-versus-OLS weight difference is a point-estimate pattern (equal-weights Wald p = 0.621). The 2SLS t-ratio (2.43) is below the tF critical value (2.93), although the Anderson-Rubin set excludes zero.

Key numbers

MetricValueInterpretation
n_obs3010All men in the 1976 NLSYM extract; no analysis variable missing.
b_ols0.0747 (HC1 SE 0.0037); R2 0.300OLS return in Card's main specification.
first_stage_pi0.3199 (SE 0.0851); robust F 14.14Moderately strong first stage.
reduced_form_rho0.0421 (SE 0.0175)Proximity raises log wages by 4.2 log points.
b_iv0.1315 (HC1 SE 0.0541); t 2.43; tF 5% critical value 2.93; AR set [0.0282, 0.2812]2SLS is 1.76 times OLS; t is below the tF critical value but the AR set excludes zero.
dwh_p0.271IV = OLS not rejected; the gap itself is not established.
b_rw0.0694; Fieller set [0.0496, 0.0866]IV-weighted OLS lies below OLS.
weighting_term-0.0053 [-0.0223, 0.0095]; upper end 0.13 of OLS; RF units -0.0017 [-0.0061, 0.0027]Absolute bound: weighting contributes at most 0.0095 log points per year at the observational slopes.
within_margin_term0.0621 (SE 0.0668) [-0.0414, 0.2107]; MDE at 80% power 0.1872 (2.51 x OLS)Not informative: its minimum detectable effect is 2.5 times the OLS return.
remainder_rf0.0199 [-0.0139, 0.0524], p 0.229; MDE 0.0463 (1.10 x reduced form)The MDE exceeds the whole reduced form.
gap0.0568 [-0.0477, 0.2042]; theta -0.093Share-based statements divide by a gap whose set contains zero.
weights_blocksIV 13-15 0.514 [0.3359, 0.7600] vs OLS 0.451; IV 12+16 0.178 [0.0795, 0.2567] vs OLS 0.218; equal-weights Wald p 0.621Point-estimate patterns only; every per-margin IV interval contains the OLS weight.
profile_free_boundTV 0.134 (bootstrap range 0.093-0.467); max weighting within observed range 0.0319 = 0.561 of gap; range needed for half the gap 0.212 vs observed 0.238At fine margins the bound allows weighting to exceed half the point gap: the refutation holds relative to the observational benchmark only.
coarse_block_bound5 blocks: weighting -0.0049 [-0.0184, 0.0064]; TV 0.074 (range 0.050-0.373); max share 0.179 at the point TV, 0.901 at its 97.5th percentileWith block-constant returns the bound is tight at the point estimates but not robust to weight noise.
nearc2_aloneFS 0.1216 (0.0780), F 2.43; RF 0.0357 (0.0160); 2SLS 0.2932, AR set unbounded; RF split: predicted 0.0091, weighting -0.0044 [-0.0088, -0.0003], remainder 0.0309 [0.0048, 0.0588] p 0.027; direct effect 0.0312 (0.0148) p 0.035; equal weights p 0.015Two-year-college proximity barely moves schooling, yet raises wages; its within-margin null is rejected.
nearc2_nearc4_jointF 8.32; 2SLS 0.1571 (0.0526), AR set [0.0530, 0.3540]; Hansen J 1.27, p 0.260; b_RW 0.0646 [0.0440, 0.0799]; weighting -0.0101 [-0.0284, 0.0031]; within 0.0925 [0.0000, 0.2382]; 2-df direct-effect chi2 5.82, p 0.054Over-identification not rejected, though the test has little power with nearc2's F of 2.43; the joint within-margin set touches zero (bootstrap p = 0.050), and because the bootstrap holds the index weights fixed, the joint bootstrap sets are if anything too narrow.
covariate_cellsrace x South x SMSA saturated: OLS 0.0745, 2SLS 0.1323 AR [0.0278, 0.2849]; cell weighting 0.0014 [-0.0102, 0.0128]; within-cell margin weighting -0.0217 [-0.0582, 0.0000]; remainder 0.0782 [-0.0207, 0.2347]; pooled coarse weighting -0.0048 [-0.0187, 0.0067]Covariate-cell weighting explains essentially none of the gap.
nls_weightsOLS 0.0753; 2SLS 0.1578 AR [0.0527, 0.3316]; weighting -0.0082 [-0.0285, 0.0075]; within 0.0908 [-0.0118, 0.2569]Weighting by NLS sampling weights leaves the conclusion unchanged.
background_split_observedboth parents observed n 2220; high n 1405 F 11.31 weighting -0.0078 [-0.0329, 0.0118]; low n 815 F 0.42; original low group: 616 of 1431 with a parent missing, 253 with both missingThe low-background first stage vanishes once missing parental schooling is excluded.
robust_pool_1_11F 14.14; OLS 0.0747; 2SLS 0.1315; weighting -0.0019 [-0.0163, 0.0132]Weighting term negative.
robust_family_backgroundF 16.73; OLS 0.0723; 2SLS 0.1347; weighting -0.0054 [-0.0217, 0.0092]Weighting term negative.
robust_age_controlsF 11.64; OLS 0.0344; 2SLS 0.0827; weighting -0.0030 [-0.0198, 0.0115]Weighting term negative.
robust_not_enrolledF 12.44; OLS 0.0766; 2SLS 0.1316; weighting -0.0063 [-0.0261, 0.0084]Weighting term negative.
monotonicitymin t -0.51; 2 of 17 thresholds negativeTestable implication not rejected for the four-year-college instrument; thresholds 2-8 rest on 132 men.
hypothesis_supportedfalseDecision rule: theta < 0.5 and the weighting term's upper bound (0.0095) is below half the point gap.

Caveats

  • The refutation holds relative to the observational benchmark only. The profile-free bound allows weighting to deliver 0.561 of the point gap at fine margins (half the gap needs a return range of 0.212, inside the observed 0.238). At coarse blocks it allows 0.179 at the point weights but 0.901 at the 97.5th percentile of the weight distance.
  • The within-margin term is uninformative: its minimum detectable effect at 80 percent power is 0.1872 (2.51 times OLS), and 0.0463 in reduced-form units, more than the reduced form itself (0.0421).
  • The IV-OLS gap is not established (Hausman p = 0.271; Fieller set [-0.0477, 0.2042]), so theta and every share divide by an insignificant point gap; absolute bounds are the informative statements.
  • The nearc2 rejection could reflect a failure of exclusion common to proximity instruments, high within-margin returns for its compliers, noise in a first stage with F 2.43, or a failure of monotonicity. Its first-stage shifts are negative at thresholds 11 and 12 (-0.0174, t = -1.73; -0.0136), so the testable implication of monotonicity is not clearly satisfied and four of its normalized weights are negative (-0.143 and -0.112 among them); the negative diploma-year shift lowers the own-shift prediction (0.0047) against which the remainder of 0.0309 is measured. Its ratio-based quantities are unbounded.
  • The covariate dimension uses coarse margins (five blocks) inside seven race x South x SMSA cells, with the controls saturated in those cells. The cell IV weights are imprecise and two are negative (-0.061, -0.009); as Blandhol et al. (2022) show for linear covariate adjustment, the cell term is then descriptive rather than a positively weighted average of cell-specific effects. The 34-df equal-weights rejection (p = 0.012), which relies on a bootstrap covariance estimated from 999 draws, describes the estimators' weights, not where compliers are.
  • Family background enters only through the split. Among the 2,220 men with both parents' schooling observed, the low-background first stage is nil (F 0.42), so Card's family-background mechanism cannot be examined in this decomposition.
  • The 2SLS t-ratio of 2.43 is below the Lee et al. tF 5 percent critical value (2.93, interpolated from their Table 3); inference on b_IV rests on the Anderson-Rubin set.
  • The unweighted estimates reproduce Card's 0.320, 0.042 and 0.132 (working-paper version, Table 3, Panel A); weighting by NLS sampling weights raises 2SLS to 0.1578 but leaves the weighting term negative (-0.0082).
  • One cohort; years 1-8 rest on 132 men; self-reported schooling may be misclassified; no local-labor-market identifier for clustering.

5.Paper writing

Attempt 10Done

Manuscript drafted: “Where Proximity Bites: Margin Weighting and the IV–OLS Gap in the Return to Schooling” (≈ 6198 words).

Reviewer feedback applied: Editorial alignment after acceptance (no new referee request): make the analysis record agree with the accepted manuscript. The two-instrument over-identification test 'does not reject' (Hansen J p = 0.260; low power with a weak second instrument), not 'passes'; check every other statement in the analysis record against the final paper. The manuscript itself is unchanged: resubmit it as accepted.

≈ 6198 words · AI-drafted manuscript · not peer-reviewed

Where Proximity Bites: Margin Weighting and the IV–OLS Gap in the Return to Schooling

Abstract

Card's college-proximity 2SLS return to schooling, 0.132, is 1.76 times OLS. We test whether the gap arises because IV and OLS weight schooling margins differently. In an exact decomposition for 3,010 NLSYM men, reweighting the per-margin OLS slopes by the Angrist–Imbens weights lowers the estimate to 0.069. At the observational slopes, margin weighting contributes at most 0.0095 log points per year (95 percent bound). Covariate-cell weighting contributes little, and two-year-college proximity rejects the within-margin null. The point estimate is nearly twice OLS, a difference the data do not establish (Hausman p = 0.27).

Keywords: return to schooling; instrumental variables; college proximity; average causal response; weak instruments

JEL codes: C26, I26, J24

I. Introduction

Why does the instrumental-variables (IV) estimate of the return to schooling so often exceed ordinary least squares (OLS)? Card's (1995) college-proximity design is the canonical case. In the National Longitudinal Survey of Young Men (NLSYM), using proximity to a four-year college as an instrument gives a return of 0.1315 log points per year in Card's main specification, 1.76 times the OLS estimate of 0.0747 (Table 3). Supply-side instruments typically deliver IV estimates at least as large as OLS (Card 1999, 2001). Card reads the pattern as high marginal returns among the men such instruments move, and the marginal-treatment-effect literature as selection on gains (Heckman, Urzua, and Vytlacil 2006; Carneiro, Heckman, and Vytlacil 2011). If the gap is instead an artefact of how the two estimators average, it says nothing about who gains from schooling.

To our knowledge, one such artefact has not been tested in this design. Under independence, exclusion and monotonicity, 2SLS with a multivalued treatment estimates the Angrist and Imbens (1995) average causal response: a weighted average of the effects of each additional year, each weight being the share of compliers moved across that year (Imbens and Angrist 1994 treat the binary case). OLS also averages per-year wage differences, with weights fixed by the schooling distribution (Yitzhaki 1996; Słoczyński 2022 shows how such implicit weights can mislead). Because wages jump at the high-school and college diplomas (Hungerford and Solon 1987; Jaeger and Page 1996), the two estimators can disagree purely because they average different margins, as linear estimators do in other settings (Løken, Mogstad, and Wiswall 2012).

We test this explanation on the 3,010 NLSYM men by decomposing the IV–OLS gap exactly into two terms. The margin-weighting term compares OLS with the OLS margin slopes averaged using the Angrist–Imbens weights; the within-margin term is the amount by which IV exceeds that reweighted benchmark. We then separate covariate-cell weighting from margin weighting, and repeat the exercise with a second instrument, proximity to a two-year college. Every term carries a weak-instrument-robust confidence set, and the within-margin null is also tested in reduced-form units. Our decision rule declares the weighting hypothesis supported only if weighting accounts for at least half of the gap and the within-margin term is indistinguishable from zero.

Relative to the observational slopes, the hypothesis is refuted. The IV-weighted benchmark, 0.0694, lies below OLS, and the margin-weighting term is −0.0053: at most 0.0095 log points per year at the 95 percent upper bound, 0.13 of the OLS return. The point estimates put more IV weight than OLS weight on the some-college years 13–15 (0.514 against 0.451) and less on the diploma years 12 and 16 (0.178 against 0.218), where the observational slopes are largest. The intervals of these IV shares contain the OLS values, however, and a Wald test does not reject equal weights (p = 0.621). Covariate-cell weighting contributes 0.0014 [−0.0102, 0.0128]. With two-year-college proximity as the instrument, the within-margin null is rejected (p = 0.027).

Three qualifications bound these claims. The refutation holds relative to the observational benchmark only: a causal profile within the observed range of slopes, but shaped against the diploma premia, could let weighting deliver 0.561 of the point gap. The within-margin term is not informative, because its minimum detectable effect is 2.51 times the OLS return. And the gap itself is not established: neither a Hausman test (p = 0.271) nor the Anderson–Rubin set for 2SLS, [0.0282, 0.2812], rules out equality of IV and OLS.

The paper builds on Kling (2001), who computed the Angrist–Imbens weights for Card's instrument but did not compare them with the OLS weights, and on Ishimaru (2024), who decomposes IV–OLS gaps into covariate-weight, margin-weight and marginal-effect components in other datasets. Blandhol et al. (2022) find that Card's linear covariate adjustment fails their condition for a LATE interpretation, yet a flexible adjustment moves his IV estimate by only about 8 percent. That concerns how the IV estimand adjusts for covariates; we measure how IV and OLS weight covariate cells and margins. Section II sets out the framework, Section III the data and estimating equations, and Section IV the results. Section V covers robustness and heterogeneity, and Section VI concludes.

II. Conceptual Framework

Let Yi(s)Y_i(s) be man ii's log wage with ss years of schooling, Si(z)S_i(z) his schooling when the proximity indicator equals z∈{0,1}z\in\{0,1\}, and WiW_i the controls. We maintain three assumptions. (A1) Conditional on WiW_i, proximity ZiZ_i is independent of potential wages and potential schooling. (A2) Exclusion: proximity affects wages only through schooling. (A3) Monotonicity: Si(1)≥Si(0)S_i(1)\geq S_i(0). Under these, Angrist and Imbens (1995) show that

βIV=∑jωj E[Yi(j)−Yi(j−1)∣Si(1)≥j>Si(0)],ωj=P(Si≥j∣Zi=1)−P(Si≥j∣Zi=0)E(Si∣Zi=1)−E(Si∣Zi=0).(1)\beta_{IV}=\sum_j \omega_j\,E\left[Y_i(j)-Y_i(j-1)\mid S_i(1)\geq j>S_i(0)\right],\qquad \omega_j=\frac{P(S_i\geq j\mid Z_i=1)-P(S_i\geq j\mid Z_i=0)}{E(S_i\mid Z_i=1)-E(S_i\mid Z_i=0)}. \tag{1}

The weights are non-negative and sum to one, and (A3) implies a testable restriction: proximity cannot lower P(Si≥j)P(S_i\geq j) at any jj. OLS is also an average over margins. With Δj\Delta_j the difference in mean log wages between adjacent schooling levels, βOLS=∑jμjΔj\beta_{OLS}=\sum_j\mu_j\Delta_j, where μj∝Cov(1{Si≥j},Si)\mu_j\propto \mathrm{Cov}(1\{S_i\geq j\},S_i) depends on the schooling distribution alone (Yitzhaki 1996).

With controls, the same algebra holds exactly in the sample. Split schooling into components AikA_{ik} that sum to SiS_i, let dkd_k be the slope on each in a wage equation saturated in schooling, and define the IV-weighted OLS benchmark bRW=∑kwkdkb_{RW}=\sum_k w_k d_k. Then

bIV−bOLS=∑k(wk−mk) dk⏟margin weighting+(bIV−bRW)⏟within margins,(2)b_{IV}-b_{OLS}=\underbrace{\sum_k (w_k-m_k)\,d_k}_{\text{margin weighting}}+\underbrace{(b_{IV}-b_{RW})}_{\text{within margins}}, \tag{2}

where mkm_k and wkw_k are the sample OLS and IV weights of Section III. The margin-weighting hypothesis is that the complier effect at each margin equals the observational slope, so the second term is zero and the first is the whole gap. The alternatives work through the second term: compliers with high returns at a given margin, selection on gains, IV and OLS weighting covariate cells differently, measurement error in schooling, and a failure of exclusion. Ability bias in the observational slopes also enters the second term but pushes it down, not up. Multiplying by the first-stage coefficient π\pi gives the identity in reduced-form units,

ρ=π bOLS+π (bRW−bOLS)+ρe,(3)\rho=\pi\,b_{OLS}+\pi\,(b_{RW}-b_{OLS})+\rho_e, \tag{3}

where ρ\rho is the reduced-form effect of proximity on log wages and ρe\rho_e its effect on the residual of the schooling-saturated wage equation. The within-margin null, ρe=0\rho_e=0, says that proximity raises wages by exactly what its schooling shifts predict from the observational profile. It involves no division by π\pi, so weak identification does not distort a test of it. If the margin slopes are estimated within covariate cells gg, the weighting term splits further into a cell term, ∑g(WgIV−WgOLS)Dg\sum_g (W^{IV}_g-W^{OLS}_g)D_g, where WgW_g are the cell totals of the weights and DgD_g is the OLS-weighted cell slope, and a within-cell margin term, in the spirit of Ishimaru (2024).

The decomposition benchmarks against observational slopes, which need not be causal. Because both weight vectors sum to one, any common per-margin return profile bb satisfies

∣∑k(wk−mk) bk∣≤TV⋅(max⁡kbk−min⁡kbk),TV=12∑k∣wk−mk∣.(4)\Big|\sum_k (w_k-m_k)\,b_k\Big|\leq TV\cdot\Big(\max_k b_k-\min_k b_k\Big),\qquad TV=\tfrac12\sum_k|w_k-m_k|. \tag{4}

The bound is attained by a profile that is high wherever wk>mkw_k>m_k and low wherever wk<mkw_k<m_k.

III. Data and Empirical Strategy

Data.—We use the 1976 cross-section of 3,010 men from the NLSYM analysed by Card (1995). The extract is distributed in the R package wooldridge and mirrored by the Rdatasets archive (Rdatasets n.d.), from which we retrieved the file on 24 October 2026 (SHA-256 326dad817467b20d…). The outcome is the natural log of average hourly earnings in 1976, recorded in cents; the mean hourly wage is 5.77 dollars (Table 1). Schooling is completed years, from 1 to 18. The instrument ZiZ_i indicates that the respondent grew up in a local labor market with an accredited four-year college; the second instrument indicates a two-year college. The controls WiW_i follow Card's main specification: potential experience and its square; indicators for being Black and for living in the South and in an SMSA in 1976; an indicator for living in an SMSA in 1966; and eight indicators for region of residence in 1966. No analysis variable is missing, so no observation is dropped.

Table 1 shows why the controls matter. Men who grew up near a four-year college have 13.53 years of schooling against 12.70, earn more, and were far more likely to live in an SMSA in 1966 (0.799 against 0.328). Conditional on WiW_i, proximity is unrelated to parents' education and to living with both parents at 14, but it is marginally associated with holding a library card at 14 (0.037, standard error 0.020; Table 2).

Table 1—Summary Statistics by College Proximity

VariableAllNear a four-year collegeNot near
Observations3,0102,053957
Hourly wage (dollars)5.776.065.16
Log hourly wage (log cents)6.2626.3116.155
Years of schooling13.2613.5312.70
Exactly 12 years (share)0.3300.3200.350
16 or more years (share)0.2710.2930.225
Potential experience (years)8.868.689.23
Black0.2340.2120.280
South in 19660.4140.3290.598
SMSA in 19660.6500.7990.328
Enrolled in school in 19760.0920.1020.072

Notes: Means for the 3,010 men in Card's (1995) NLSYM extract, observed in 1976. "Near" means the respondent grew up in a local labor market with an accredited four-year college. Wages are converted from cents to dollars.

Table 2—Balance of Predetermined Background on College Proximity

Background variableMeanProximity coefficient (SE)p-valueObservations
Father's education (years)10.0030.202 (0.165)0.2212,320
Mother's education (years)10.348−0.086 (0.135)0.5212,657
Library card at 140.6740.037* (0.020)0.0562,997
Lived with both parents at 140.789−0.022 (0.018)0.2153,010

Notes: Each row regresses the background variable on the four-year-college indicator and the Table 3 controls, for the men with that variable observed. HC1 standard errors in parentheses. *, **, *** denote significance at 10, 5, 1 percent.

Estimating equations.—The benchmarks are

yi=βSi+Wi′γ+εi,Si=πZi+Wi′ϕ+vi,yi=ρZi+Wi′λ+ui,(5)y_i=\beta S_i+W_i'\gamma+\varepsilon_i,\qquad S_i=\pi Z_i+W_i'\phi+v_i,\qquad y_i=\rho Z_i+W_i'\lambda+u_i, \tag{5}

each estimated by OLS, with the just-identified 2SLS estimate bIV=ρ/πb_{IV}=\rho/\pi. For the decomposition we write Si=min⁡(Si,8)+∑j=9181{Si≥j}S_i=\min(S_i,8)+\sum_{j=9}^{18}1\{S_i\geq j\}. Only 132 men have eight years or fewer, so years 1–8 enter as one linear piece and each later year is its own component. The schooling-saturated wage equation

yi=∑kdkAik+Wi′γ+ei(6)y_i=\sum_k d_k A_{ik}+W_i'\gamma+e_i \tag{6}

gives the margin slopes dkd_k. The OLS weight mkm_k is the coefficient on SiS_i in a regression of AikA_{ik} on SiS_i and WiW_i. The IV weight is wk=πk/πw_k=\pi_k/\pi, with πk\pi_k the coefficient on ZiZ_i in a regression of AikA_{ik} on ZiZ_i and WiW_i. Both sets of weights sum to one, so equations (2) and (3) are sample identities. The direct-effect test adds ZiZ_i to equation (6). For the covariate dimension we replace the race, South and SMSA indicators with dummies for seven cells formed from their combinations; the few Black men living outside both the South and an SMSA are merged into the Black non-South cell. We then estimate margin slopes within each cell on five coarse blocks: years 1–11 and 13–15 per year, the twelfth and sixteenth years, and years 17–18 per year. With both instruments, 2SLS is IV on the fitted first-stage index.

Identification.—The IV weights locate compliers only under (A1)–(A3): conditional on experience, race, urban residence and region, proximity must be as good as randomly assigned and must affect wages only through schooling. One instrument cannot test exclusion; the second instrument and the balance results speak to it. With linear rather than saturated controls, the 2SLS estimand need not be a positively weighted average of covariate-specific effects (Blandhol et al. 2022).

Inference.—All estimates are unweighted, as Card's are: our specification reproduces his first stage, reduced form and IV estimate (0.320, 0.042 and 0.132; Card 1995, working-paper version, Table 3, Panel A, columns 1, 3 and 5). Unweighted estimation suits the question because the object is Card's estimate and the decomposition is an identity of the unweighted estimators; Table 8 reports the NLS-weighted version. HC1 standard errors and the individual-level pairs bootstrap (1,999 replications, seed 20261024; 999 in Tables 6–8) ignore the survey's clustered design, and the extract has no local-labor-market identifier to cluster on. For bIVb_{IV} we report the Anderson and Rubin (1949) set, inverted from the robust joint covariance of ρ\rho and π\pi (on a grid, with two degrees of freedom, for the joint design), alongside the tF critical value of Lee et al. (2022). For every other ratio in equation (2), a numerator over π\pi, we invert the analogous Fieller-type test with the bootstrap covariance; reduced-form terms carry percentile intervals (Andrews, Stock, and Sun 2019 review these tools). The equal-weights test is a Wald test on the bootstrap covariance of the reduced-form weight gaps πk−πmk\pi_k-\pi m_k. Over-identification is tested with a robust Hansen J. Minimum detectable effects use 2.80 standard errors (80 percent power at 5 percent). The replication package is one deterministic Python script (Python 3.12.13, NumPy 2.5.2, pandas 3.0.5, SciPy 1.18.1) that reproduces every table in about 20 seconds; the data file's SHA-256 is verified when it is provisioned, not by the script.

IV. Results

A. Benchmarks.—OLS gives 0.0747 (0.0037) (Table 3). Proximity raises schooling by 0.3199 years (0.0851) and log wages by 0.0421, so 2SLS gives 0.1315 (0.0541): 13 log points per year of schooling against 7.5 for OLS, on a mean hourly wage of 5.77 dollars. The 2SLS t-ratio, 2.43, is below the tF 5 percent critical value at this first-stage strength, 2.93, so we attach no conventional stars to it. The Anderson–Rubin set, [0.0282, 0.2812], excludes zero but contains the OLS estimate, and the Hausman test does not reject equality (p = 0.271). Under the maintained exclusion and monotonicity assumptions, the data establish that the return is positive, not that IV exceeds OLS.

Table 3—OLS, First-Stage, Reduced-Form and 2SLS Estimates

(1) OLS(2) First stage(3) Reduced form(4) 2SLS
Dependent variableLog wageYears of schoolingLog wageLog wage
Years of schooling0.0747*** (0.0037)0.1315 (0.0541)
Four-year college nearby0.3199*** (0.0851)0.0421** (0.0175)
Robust first-stage F14.14
t-ratio; tF 5 percent critical value2.43; 2.93
Anderson–Rubin 95 percent set[0.0282, 0.2812]
2SLS / OLS1.76
Hausman test of 2SLS = OLS, p-value0.271
R2R^20.3000.4770.195
Observations3,0103,0103,0103,010

Notes: Controls: potential experience and its square; indicators for Black, South and SMSA in 1976, SMSA in 1966; eight 1966 region indicators. HC1 standard errors in parentheses, unclustered (no local-labor-market identifier). Column 4 carries no stars because its t-ratio is below the 5 percent tF critical value of Lee et al. (2022, Table 3, linearly interpolated at the observed F). The Anderson–Rubin set is inverted from the robust joint covariance of reduced form and first stage; the Hausman test is a robust control-function t-test. *, **, *** denote significance at 10, 5, 1 percent.

B. Where proximity bites.—Table 4 gives the margin slopes and both weight vectors, with intervals, by year of schooling. The diploma years dominate the observational profile: the wage slope is 0.1572 at year 12 and 0.2035 at year 16, against 0.0819, 0.0361 and −0.0218 for years 13–15. The IV weight is largest at year 13 (0.199), consistent with Kling's (2001) finding that proximity bites hardest at the transition into college. The IV-versus-OLS differences are point-estimate patterns only: every per-margin IV interval contains the OLS weight, and the equal-weights test does not reject (p = 0.621). The testable implication of monotonicity is not rejected. The estimated effect is negative at 2 of 17 thresholds, with a smallest t-statistic of −0.51, although thresholds 2–8 rest on only 132 men.

Table 4—Margin Slopes and the OLS and IV Weights, by Year of Schooling

Margin (years)Men at levelWage slope dkd_kOLS weight mkm_k [95 percent interval]IV weight wkw_k [95 percent interval]Two-year-college shift πk(2)\pi_k^{(2)}
1–8 (per year)1320.0840*** (0.0221)0.049 [0.034, 0.065]0.008 [−0.138, 0.113]−0.0105 (0.0134)
981−0.0343 (0.0599)0.032 [0.027, 0.036]0.044 [−0.011, 0.099]−0.0016 (0.0070)
101250.0149 (0.0531)0.046 [0.041, 0.051]0.076 [0.021, 0.151]0.0068 (0.0085)
111590.1128*** (0.0422)0.061 [0.056, 0.066]0.064 [−0.004, 0.136]−0.0174* (0.0101)
129920.1572*** (0.0327)0.077 [0.072, 0.083]0.083 [0.002, 0.175]−0.0136 (0.0122)
132810.0819*** (0.0236)0.145 [0.140, 0.151]0.199 [0.116, 0.335]0.0311* (0.0164)
142630.0361 (0.0316)0.154 [0.149, 0.160]0.178 [0.101, 0.277]0.0326** (0.0159)
15160−0.0218 (0.0381)0.151 [0.146, 0.157]0.138 [0.056, 0.212]0.0400*** (0.0154)
164590.2035*** (0.0341)0.141 [0.135, 0.146]0.095 [−0.004, 0.160]0.0211 (0.0150)
171510.0065 (0.0396)0.086 [0.079, 0.092]0.052 [−0.050, 0.119]0.0175 (0.0124)
182070.1479*** (0.0453)0.058 [0.052, 0.065]0.064 [0.003, 0.134]0.0156 (0.0101)
Weight on years 13–150.451 [0.436, 0.466]0.514 [0.336, 0.760]
Weight on years 12 and 160.218 [0.212, 0.224]0.178 [0.080, 0.257]
Test of equal IV and OLS weights, χ2(10)\chi^2(10) and p8.08; 0.62122.11; 0.015
Monotonicity: smallest t on 1{S≥j}1\{S \geq j\}, jj = 2–18; negative shifts−0.51; 2 of 17

Notes: N = 3,010; controls as in Table 3. "Men at level" counts men with exactly that schooling (eight years or fewer in the first row). Slopes come from equation (6), with HC1 standard errors and stars (10, 5, 1 percent). Weight intervals are pairs-bootstrap percentiles (1,999 replications); weights carry no stars, since no weight is tested against zero, and each weight column sums to one. The last column gives the two-year-college first-stage shifts in reduced-form units (HC1 standard errors). Four shifts, and hence four normalized weights, are negative, most clearly at thresholds 11 and 12 (shifts −0.0174, t = −1.73, and −0.0136; weights −0.143 and −0.112), so the testable implication of monotonicity is not clearly satisfied for that instrument. The negative shift at the diploma year, where the slope is large, lowers the instrument's own-shift prediction of its reduced form, 0.0047 (0.0091 − 0.0044 in Table 7), against which its remainder of 0.0309 is measured. Its normalized weights are also uninformative because its first stage is weak (F = 2.43). The equal-weights test is a Wald test of πk=πmk\pi_k=\pi m_k for all kk on the bootstrap covariance of these reduced-form weight gaps. The monotonicity row gives the smallest HC1 t on proximity across regressions of 1{S≥j}1\{S\geq j\} on proximity and the controls; thresholds 2–8 rest on the 132 men with eight years or fewer.

C. The decomposition.—At the observational slopes, margin weighting cannot produce a gap of the estimated size (Table 5). The IV-weighted benchmark is 0.0694, with a Fieller set of [0.0496, 0.0866]. The margin-weighting term is −0.0053, and its 95 percent set, [−0.0223, 0.0095], caps any contribution at 0.0095 log points per year, 0.13 of the OLS return. In reduced-form units the cap is 0.0027, against a reduced form of 0.0421. By our decision rule the hypothesis is refuted relative to this benchmark. The share-based statements, a share explained of −0.093 and an upper end of 0.168 of the gap, divide by a point gap whose own set contains zero; they are fragile and we do not lean on them.

The within-margin term, 0.0621, is not informative. Its set is [−0.0414, 0.2107], and its minimum detectable effect at 80 percent power is 0.1872, 2.51 times the OLS return. In reduced-form units the remainder is 0.0199 (p = 0.229), and its minimum detectable value, 0.0463, exceeds the whole reduced form. Adding proximity to the schooling-saturated wage equation gives 0.0200 (0.0165; p = 0.226).

Table 5—Decomposition of the IV–OLS Gap

EstimateStandard error95 percent set
Panel A. Coefficient units (log points per year of schooling)
OLS, bOLSb_{OLS}0.07470.0037
IV-weighted OLS, bRWb_{RW}0.06940.0088[0.0496, 0.0866]
2SLS, bIVb_{IV}0.13150.0541[0.0282, 0.2812]
Margin-weighting term, bRW−bOLSb_{RW} - b_{OLS}−0.00530.0079[−0.0223, 0.0095]
Upper end of its set relative to bOLSb_{OLS}0.13
Within-margin term, bIV−bRWb_{IV} - b_{RW}0.06210.0668[−0.0414, 0.2107]
Minimum detectable within-margin term (80 percent power); relative to bOLSb_{OLS}0.18722.51
Gap, bIV−bOLSb_{IV} - b_{OLS}0.05680.0669[−0.0477, 0.2042]
Share of the point gap: weighting term, θ\theta; upper end of its set−0.0930.168
Panel B. Reduced-form units (log points of the proximity wage effect)
Predicted by OLS, πbOLS\pi b_{OLS}0.02390.0065[0.0114, 0.0368]
Margin-weighting increment−0.00170.0022[−0.0061, 0.0027]
Within-margin remainder, ρe\rho_e (bootstrap p = 0.229)0.01990.0165[−0.0139, 0.0524]
Minimum detectable remainder (80 percent power); relative to ρ\rho0.04631.10
Reduced form, ρ\rho (sum of the first three rows)0.04210.0175
Proximity added to the schooling-dummy wage equation (p = 0.226)0.02000.0165
Panel C. Profile-free bound on the weighting termFine marginsFive coarse blocks
Total-variation distance between IV and OLS weights [bootstrap range]0.134 [0.093, 0.467]0.074 [0.050, 0.373]
Observed range of the margin slopes0.2380.137
Largest weighting term with returns inside that range0.03190.0102
Its share of the point gap; at the top of the distance range0.561; 1.9520.179; 0.901
Return range needed for half the point gap; at the top of the distance range0.212; 0.0610.382; 0.076

Notes: N = 3,010; controls as in Table 3. Standard errors for bOLSb_{OLS}, bIVb_{IV} and the direct-effect coefficient are HC1; the others are pairs-bootstrap (1,999 replications). The bIVb_{IV} set is Anderson–Rubin; the other Panel A sets are Fieller-type, from the bootstrap covariance of each numerator with the first stage. Panel B intervals are bootstrap percentiles. The weighting and within-margin terms sum to the gap exactly, and the first three Panel B rows sum to ρ\rho. Minimum detectable effects are 2.80 standard errors. The shares in Panel A divide by the point gap, whose set contains zero. Panel C applies equation (4). The total-variation distance is a nonsmooth function of noisy weights, so its bootstrap range (2.5th–97.5th percentiles) is not a confidence interval; entries "at the top of the distance range" use its 97.5th percentile. The five coarse blocks are years 1–11 and 13–15 per year, years 12 and 16, and years 17–18 per year.

D. A profile-free bound.—Could a causal profile different from the observational one rescue the weighting account? At fine margins the two weight vectors are 0.134 apart in total variation (Panel C). By equation (4), returns confined to the observed range of slopes (0.238) could produce a weighting term of 0.0319, or 0.561 of the point gap, above the decision rule's threshold. Half the gap needs a return range of only 0.212, inside the observed range. Such a profile, high in years 9–14 and 18 where the IV weight exceeds the OLS weight and low in years 1–8 and 15–17, would reverse the diploma premium at year 16, but nothing in these data rules it out. With five coarse blocks the weights are estimated more precisely. The distance falls to 0.074, and the largest weighting term within the blocks' observed range (0.137) is 0.0102, 0.179 of the point gap. Yet the distance's bootstrap range reaches 0.373, at which the share would be 0.901. The refutation therefore holds relative to the observational benchmark only.

E. The covariate dimension.—Covariate weighting does not rescue the account either (Table 6). With race, South and SMSA saturated in the controls, OLS and 2SLS barely move (0.0745 and 0.1323). Estimating the margin slopes within the seven cells splits the weighting term in two. The covariate-cell term is 0.0014 [−0.0102, 0.0128]; the within-cell margin term is −0.0217 [−0.0582, 0.0000]. The IV weights favour the non-Black, non-South SMSA cell (0.597 against an OLS weight of 0.404) and the Black, South SMSA cell (0.238 against 0.076). These cell weights are imprecise. The first cell has one of the lowest cell slopes (0.0678) and the second the highest (0.0941), so the two shifts nearly offset. Two cell weights are negative (−0.061 and −0.009). As Blandhol et al. (2022) show for linear covariate adjustment, the cell term is then descriptive rather than a positively weighted average of cell-specific effects, and the χ2(34)\chi^2(34) rejection of equal weights across cells and margins (p = 0.012) describes the estimators' weights, not where compliers are. The remainder rises to 0.0782 [−0.0207, 0.2347] and is no more informative than before. These cells do not capture family background, the dimension Kling (2001) emphasizes; Section V returns to it.

Table 6—The Covariate Dimension: Margin Slopes within Race × South × SMSA Cells

EstimateStandard error95 percent set
Panel A. Decomposition with the cells saturated in the controls
OLS, bOLSb_{OLS}0.07450.0037
2SLS, bIVb_{IV} (robust F = 13.83)0.13230.0548[0.0278, 0.2849]
Gap, bIV−bOLSb_{IV} - b_{OLS}0.05790.0678[−0.0459, 0.2043]
Pooled coarse margins: margin-weighting term−0.00480.0066[−0.0187, 0.0067]
Pooled coarse margins: within-margin term0.06270.0682[−0.0407, 0.2111]
Cells × coarse margins: total weighting term−0.02030.0161[−0.0591, 0.0035]
of which covariate-cell weighting0.00140.0058[−0.0102, 0.0128]
of which within-cell margin weighting−0.02170.0150[−0.0582, 0.0000]
Cells × coarse margins: remainder, bIV−bRWb_{IV} - b_{RW}0.07820.0707[−0.0207, 0.2347]
Covariate-cell weighting in reduced-form units0.0005[−0.0027, 0.0036]
Within-cell margin weighting in reduced-form units−0.0069[−0.0136, 0.0001]
Equal IV and OLS weights: pooled coarse, χ2(4)\chi^2(4) p; cells × coarse, χ2(34)\chi^2(34) p0.4100.012
Panel B. Cell weightsMenOLS weightIV weight (SE); cell slope
Non-Black, non-South, non-SMSA3870.1090.010 (0.160); 0.0610
Non-Black, non-South, SMSA1,2020.4040.597 (0.220); 0.0678
Non-Black, South, non-SMSA2540.109−0.061 (0.176); 0.0819
Non-Black, South, SMSA4640.1630.168 (0.165); 0.0796
Black, non-South2060.0600.057 (0.101); 0.0837
Black, South, non-SMSA2180.079−0.009 (0.149); 0.0806
Black, South, SMSA2790.0760.238 (0.164); 0.0941

Notes: N = 3,010. Controls: potential experience and its square, SMSA in 1966, eight 1966 region indicators, and dummies for seven race × South × SMSA cells (the Black non-South cell includes the few Black men living outside both the South and an SMSA). Margin slopes use the five coarse blocks of Table 5, pooled or within cells. Cell weights are the cell totals of the OLS and IV weights; the cell slope is the OLS-weighted average of the cell's margin slopes. Pairs bootstrap with 999 replications; sets as in Table 5; reduced-form rows give percentile intervals.

F. A second instrument.—The weighting hypothesis predicts that an instrument with different margin weights yields a correspondingly different IV estimate, equal to its own IV-weighted OLS. Two-year-college proximity is such an instrument (Table 7). Its first stage is weak (0.1216, F = 2.43). Its estimated shifts are concentrated in years 13–15 but negative at thresholds 11 and 12 (−0.0174, t = −1.73; −0.0136; Table 4), so the testable implication of monotonicity is not clearly satisfied and several normalized weights are negative. Its weights differ from the OLS weights (equal-weights p = 0.015). Its ratio estimates are uninformative, so we read it in reduced-form units. It raises log wages by 0.0357, while its schooling shifts, valued at the observational slopes, predict 0.0091 at the OLS weights and 0.0044 less at its own, a significantly negative weighting increment, [−0.0088, −0.0003]. Measured against that own-shift prediction, 0.0047, which the negative diploma-year shift lowers, the remainder of 0.0309 [0.0048, 0.0588] rejects the within-margin null (p = 0.027), as does the direct-effect test (p = 0.035). The prediction fails. The failure may reflect high returns for its compliers, a direct effect of proximity on wages that would also threaten the four-year instrument, noise in so weak a first stage, or a failure of monotonicity at the diploma year. With both instruments, the over-identification test does not reject (Hansen J = 1.27, p = 0.260), though it has little power with so weak a second instrument, and 2SLS is 0.1571, with an Anderson–Rubin set of [0.053, 0.354]. The weighting term is −0.0101 [−0.0284, 0.0031]. The within-margin set, [0.0000, 0.2382], touches zero (bootstrap p = 0.050), and the joint test that neither instrument predicts wages given the schooling dummies is borderline (χ2(2)\chi^2(2) = 5.82, p = 0.054). With the index weights held fixed in the bootstrap, the joint bootstrap sets are if anything too narrow.

Table 7—Two-Year-College Proximity as a Second Instrument

Four-year college (baseline)Two-year college aloneBoth jointly
First stage (SE)0.3199*** (0.0851)0.1216 (0.0780)
Robust first-stage F14.142.438.32
Reduced form (SE)0.0421** (0.0175)0.0357** (0.0160)
2SLS (SE)0.1315 (0.0541)0.2932 (0.1863)0.1571 (0.0526)
Anderson–Rubin 95 percent set[0.0282, 0.2812][unbounded][0.053, 0.354]
Hansen J (p)1.27 (0.260)
IV-weighted OLS, bRWb_{RW} [95 percent set]0.0694 [0.0496, 0.0866]0.0389 [unbounded]0.0646 [0.0440, 0.0799]
Margin-weighting term [95 percent set]−0.0053 [−0.0223, 0.0095]−0.0358 [unbounded]−0.0101 [−0.0284, 0.0031]
Within-margin term [95 percent set]0.0621 [−0.0414, 0.2107]0.2543 [unbounded]0.0925 [0.0000, 0.2382]
Reduced form predicted by OLS, πbOLS\pi b_{OLS}0.02390.0091
Margin-weighting increment [95 percent interval]−0.0017 [−0.0061, 0.0027]−0.0044 [−0.0088, −0.0003]
Within-margin remainder [95 percent interval]; p0.0199 [−0.0139, 0.0524]; 0.2290.0309 [0.0048, 0.0588]; 0.027
Instrument added to the schooling-dummy wage equation (SE); p0.0200 (0.0165); 0.2260.0312 (0.0148); 0.035
Both instruments added: coefficients; χ2(2)\chi^2(2), p0.0313, 0.0203; 5.82, 0.054

Notes: N = 3,010; controls as in Table 3. "Two-year college" indicates growing up near a two-year college. The joint design is 2SLS with both indicators, equivalently IV on the fitted first-stage index. Its F is the robust joint first-stage F, its Anderson–Rubin set inverts a robust two-degree-of-freedom test on a grid of step 0.001, and its bootstrap (999 replications) holds the index weights at their full-sample values, so its sets are if anything too narrow; the lower end of its within-margin set, 0.0000, corresponds to a bootstrap p of 0.050. Hansen J is the robust two-step statistic with one degree of freedom. "Unbounded": the first stage is not significant at 5 percent. Reduced-form rows are in log points of the instrument's wage effect, with percentile intervals. HC1 standard errors; *, **, *** denote significance at 10, 5, 1 percent.

V. Robustness and Heterogeneity

Table 8 repeats the decomposition across specifications, and the weighting term is negative in every one with a bounded set. Pooling years 1–11 gives −0.0019 and five coarse blocks −0.0049. Family-background controls give −0.0054, age in place of experience −0.0030, and excluding men enrolled in 1976 −0.0063. The NLS sampling weights give −0.0082, while they raise 2SLS to 0.1578. The age specification matters because potential experience is computed from schooling, so treating it as exogenous conditions on a variable the instrument moves. Age makes every control predetermined and turns the estimand into the return at a given age, net of forgone experience (OLS 0.0344, 2SLS 0.0827).

Table 8—Robustness and Heterogeneity of the Decomposition

SpecificationNFOLS (SE)2SLS [AR set]bRWb_{RW} [95 percent set]Weighting term [95 percent set]Within-margin term [95 percent set]Equal weights, p
Baseline3,01014.140.0747 (0.0037)0.1315 [0.0282, 0.2812]0.0694 [0.0496, 0.0866]−0.0053 [−0.0223, 0.0095]0.0621 [−0.0414, 0.2107]0.621
Years 1–11 pooled3,01014.140.0747 (0.0037)0.1315 [0.0282, 0.2812]0.0728 [0.0560, 0.0902]−0.0019 [−0.0163, 0.0132]0.0587 [−0.0447, 0.2024]0.643
Five coarse margin blocks3,01014.140.0747 (0.0037)0.1315 [0.0282, 0.2812]0.0698 [0.0534, 0.0838]−0.0049 [−0.0184, 0.0064]0.0617 [−0.0412, 0.2078]0.441
Family-background controls3,01016.730.0723 (0.0038)0.1347 [0.0358, 0.2768]0.0669 [0.0480, 0.0838]−0.0054 [−0.0217, 0.0092]0.0678 [−0.0308, 0.2054]0.556
Age in place of experience3,01011.640.0344 (0.0028)0.0827 [−0.0030, 0.2163]0.0314 [0.0130, 0.0474]−0.0030 [−0.0198, 0.0115]0.0514 [−0.0328, 0.1918]0.619
Excluding men enrolled in 19762,73212.440.0766 (0.0038)0.1316 [0.0205, 0.2963]0.0703 [0.0479, 0.0880]−0.0063 [−0.0261, 0.0084]0.0613 [−0.0516, 0.2427]0.301
NLS sampling weights3,01012.960.0753 (0.0040)0.1578 [0.0527, 0.3316]0.0670 [0.0437, 0.0847]−0.0082 [−0.0285, 0.0075]0.0908 [−0.0118, 0.2569]0.278
Two-year college as sole instrument3,0102.430.0747 (0.0037)0.2932 [unbounded]0.0389 [unbounded]−0.0358 [unbounded]0.2543 [unbounded]0.015
Both colleges as joint instruments3,0108.320.0747 (0.0037)0.1571 [0.053, 0.354]0.0646 [0.0440, 0.0799]−0.0101 [−0.0284, 0.0031]0.0925 [0.0000, 0.2382]0.092
A parent with 12+ years (missing counted as below)1,57912.380.0785 (0.0053)0.1029 [−0.0185, 0.2443]0.0650 [0.0282, 0.0879]−0.0135 [−0.0486, 0.0075]0.0379 [−0.0713, 0.1823]0.212
No parent with 12+ years (missing counted as below)1,4313.700.0677 (0.0052)0.1903 [unbounded]0.0632 [unbounded]−0.0045 [unbounded]0.1270 [unbounded]0.048
Both parents observed: a parent with 12+ years1,40511.310.0753 (0.0057)0.0809 [−0.0563, 0.2249]0.0675 [0.0389, 0.0901]−0.0078 [−0.0329, 0.0118]0.0134 [−0.1162, 0.1590]0.389
Both parents observed: neither with 12+ years8150.420.0757 (0.0067)0.1707 [unbounded]0.1050 [unbounded]0.0293 [unbounded]0.0657 [unbounded]0.054
No parent with 12+ years: men with a parent's schooling missing; with both missing616; 253
Two-year college, reduced-form units−0.0044 [−0.0088, −0.0003]0.0309 [0.0048, 0.0588]
No parent with 12+ years, reduced-form units−0.0010 [−0.0088, 0.0083]0.0281 [−0.0110, 0.0721]
Both observed, neither with 12+, reduced-form units0.0028 [−0.0126, 0.0188]0.0063 [−0.0538, 0.0633]

Notes: Each row repeats Tables 3–5 (999 bootstrap replications; baseline 1,999); controls as in Table 3 unless stated. Five coarse blocks as in Table 5. Family background adds parents' education (zero-filled, with missing-value indicators) and indicators for living with both parents and with a single mother at 14. Age replaces potential experience and its square with age and its square. NLS sampling weights: weighted least squares throughout. "A parent with 12+ years": either parent's observed schooling is 12 or more. "Unbounded": the first stage is not significant at 5 percent. The last three rows give the weak-first-stage designs' terms in reduced-form units, with percentile intervals.

The family-background split needs care. Of the 1,431 men classified as having no parent with 12 or more years, 616 have at least one parent's schooling missing and 253 have both missing. Among the 2,220 men with both parents observed, the first stage is strong enough for inference only for the 1,405 with a high-school-educated parent (F = 11.31). Their weighting term is −0.0078 [−0.0329, 0.0118]. For the other 815 the first stage vanishes (F = 0.42). In reduced-form units their weighting increment is 0.0028 [−0.0126, 0.0188], so the split cannot test Card's (1995) finding that proximity matters most for sons of poorly educated parents. Three of the fifteen equal-weights tests we report reject at 5 percent: the two-year-college instrument (p = 0.015), the 34-df cell-level test (p = 0.012) and the low-background group with missing parental schooling counted as low (p = 0.048), whose first stage has F = 3.70. In none of the three does the weighting term move toward the gap.

Three threats remain. First, selection may bias the observational slopes differently at different margins, for instance more strongly at the diploma years, and Section IV.D shows that the data cannot exclude a causal profile of this kind. Second, exclusion may fail. The library-card imbalance and the two-year-college remainder are consistent with that, and any failure loads onto the within-margin term. Third, self-reported schooling may be misclassified across adjacent years, blurring the per-margin slopes. We did not saturate the covariates beyond the seven cells, and could not bring family background into the cell decomposition with a usable first stage.

VI. Conclusion

The point estimate is nearly twice OLS, a difference the data do not establish. Measured against the observational slopes, margin weighting is not the reason. At the observational slopes, reweighting OLS to the instrument's margins lowers it slightly, and the margin-weighting term is capped at 0.0095 log points per year. Covariate-cell weighting adds 0.0014. An instrument that shifts schooling at different margins, two-year-college proximity, fails the prediction the weighting account makes. The refutation is relative to the observational benchmark. A causal profile within the observed range but shaped against the diploma premia could let weighting deliver more than half the point gap. The within-margin term, where the alternatives would show up, has too little power to discriminate among them.

Two lessons generalize. Decomposing an IV–OLS gap is informative only with inference that respects a moderate first stage. Absolute bounds on each term are more robust than shares of a gap that may be zero. And the margin-weighting term need not be positive. Here it is slightly negative for the four-year instrument and significantly negative, in reduced-form units, for the two-year one. The two-year-college remainder also warns that proximity may affect wages directly. Separating that from high returns for compliers would require an instrument whose exclusion is not in doubt.

References

Anderson, T. W., and Herman Rubin. 1949. "Estimation of the Parameters of a Single Equation in a Complete System of Stochastic Equations." Annals of Mathematical Statistics 20 (1): 46–63. https://doi.org/10.1214/aoms/1177730090.

Andrews, Isaiah, James H. Stock, and Liyang Sun. 2019. "Weak Instruments in Instrumental Variables Regression: Theory and Practice." Annual Review of Economics 11: 727–53. https://doi.org/10.1146/annurev-economics-080218-025643.

Angrist, Joshua D., and Guido W. Imbens. 1995. "Two-Stage Least Squares Estimation of Average Causal Effects in Models with Variable Treatment Intensity." Journal of the American Statistical Association 90 (430): 431–42. https://doi.org/10.1080/01621459.1995.10476535.

Blandhol, Christine, John Bonney, Magne Mogstad, and Alexander Torgovitsky. 2022. "When Is TSLS Actually LATE?" NBER Working Paper 29709, revised January 2025. https://doi.org/10.3386/w29709.

Card, David. 1995. "Using Geographic Variation in College Proximity to Estimate the Return to Schooling." In Aspects of Labour Market Behaviour: Essays in Honour of John Vanderkamp, edited by L. N. Christofides, E. K. Grant, and R. Swidinsky, 201–22. Toronto: University of Toronto Press. Also NBER Working Paper 4483. https://doi.org/10.3386/w4483.

Card, David. 1999. "The Causal Effect of Education on Earnings." In Handbook of Labor Economics, Vol. 3A, 1801–63. Amsterdam: Elsevier. https://doi.org/10.1016/S1573-4463(99)03011-4.

Card, David. 2001. "Estimating the Return to Schooling: Progress on Some Persistent Econometric Problems." Econometrica 69 (5): 1127–60. https://doi.org/10.1111/1468-0262.00237.

Carneiro, Pedro, James J. Heckman, and Edward J. Vytlacil. 2011. "Estimating Marginal Returns to Education." American Economic Review 101 (6): 2754–81. https://doi.org/10.1257/aer.101.6.2754.

Heckman, James J., Sergio Urzua, and Edward Vytlacil. 2006. "Understanding Instrumental Variables in Models with Essential Heterogeneity." Review of Economics and Statistics 88 (3): 389–432. https://doi.org/10.1162/rest.88.3.389.

Hungerford, Thomas, and Gary Solon. 1987. "Sheepskin Effects in the Returns to Education." Review of Economics and Statistics 69 (1): 175–77. https://doi.org/10.2307/1937919.

Imbens, Guido W., and Joshua D. Angrist. 1994. "Identification and Estimation of Local Average Treatment Effects." Econometrica 62 (2): 467–75. https://doi.org/10.2307/2951620.

Ishimaru, Shoya. 2024. "Empirical Decomposition of the IV–OLS Gap with Heterogeneous and Nonlinear Effects." Review of Economics and Statistics 106 (2): 505–20. https://doi.org/10.1162/rest_a_01169.

Jaeger, David A., and Marianne E. Page. 1996. "Degrees Matter: New Evidence on Sheepskin Effects in the Returns to Education." Review of Economics and Statistics 78 (4): 733–40. https://doi.org/10.2307/2109960.

Kling, Jeffrey R. 2001. "Interpreting Instrumental Variables Estimates of the Returns to Schooling." Journal of Business and Economic Statistics 19 (3): 358–64. https://doi.org/10.1198/073500101681020006.

Lee, David S., Justin McCrary, Marcelo J. Moreira, and Jack Porter. 2022. "Valid t-Ratio Inference for IV." American Economic Review 112 (10): 3260–90. https://doi.org/10.1257/aer.20211063.

Løken, Katrine V., Magne Mogstad, and Matthew Wiswall. 2012. "What Linear Estimators Miss: The Effects of Family Income on Child Outcomes." American Economic Journal: Applied Economics 4 (2): 1–35. https://doi.org/10.1257/app.4.2.1.

Rdatasets. n.d. "card: Card (1995) NLSYM Extract, from the R Package wooldridge." Data file accessed October 24, 2026. https://vincentarelbundock.github.io/Rdatasets/csv/wooldridge/card.csv.

Słoczyński, Tymon. 2022. "Interpreting OLS Estimands When Treatment Effects Are Heterogeneous: Smaller Groups Get Larger Weights." Review of Economics and Statistics 104 (3): 501–9. https://doi.org/10.1162/rest_a_00953.

Yitzhaki, Shlomo. 1996. "On Using Linear Regressions in Welfare Economics." Journal of Business and Economic Statistics 14 (4): 478–86. https://doi.org/10.1080/07350015.1996.10524677.

6.Review

Attempt 2Done

Automated review: Accept (novelty 3, rigor 4, clarity 4, reproducibility 5).

Accept
Novelty 3/5Rigor 4/5Clarity 4/5Reproducibility 5/5

Refereed by Claude Fable 5.1, a different model from the one that wrote the paper (Claude Opus 5.5).

Across three rounds the author has met every required change, including the new analyses required in round 1 (the two-year-college instrument alone and jointly, the covariate-cell split, minimum detectable effects, the both-parents-observed split, the sampling-weighted specification, the tF critical value) and the round-2 text corrections arising from them (the negative two-year-college shifts and their bearing on monotonicity and on the size of the remainder, the negative cell-level IV weights, the joint design's borderline set and fixed index weights, the checksum claim, the reference details). The executed record is unchanged since round 2, and every number in the manuscript, including those in the newly edited passages, traces to it; I had earlier reproduced the main and the added analyses independently from the data file. The paper's claims are now sized to its evidence: relative to the observational slopes, margin weighting cannot account for Card's IV-OLS gap (the weighting term is capped at 0.0095 log points per year, 0.13 of OLS; covariate-cell weighting adds 0.0014), with the explicit caveats that a differently shaped causal profile could deliver more than half the point gap, that the within-margin term has no power, and that the gap itself is not established. The two-year-college instrument's reduced-form rejection of the within-margin null is reported with the weak first stage, exclusion and monotonicity all named as possible reasons. The contribution is narrow for a general-interest journal, which the editor will weigh separately; as a referee I have no further required changes and recommend acceptance. This report is reissued unchanged in substance on a record-alignment pass: the resubmitted manuscript is byte-identical to the version accepted in round 3 and RESULT_JSON is unchanged, the pass serving only to align the pipeline's analysis record (for example, 'passes' to 'does not reject' for the over-identification test) with the accepted text.

Strengths

  • All five round-2 changes are made exactly as asked, and each is correct against the record. The Table 4 note and Section IV.F now report that two-year-college proximity's shifts are negative at thresholds 11 and 12 (-0.0174, t = -1.73; -0.0136), that four shifts and hence four normalized weights are negative (-0.143 and -0.112 among them; the record's four are -0.086, -0.013, -0.143 and -0.112), that the testable implication of monotonicity is therefore not clearly satisfied for that instrument, that the negative diploma-year shift lowers the own-shift prediction (0.0047 = 0.0091 - 0.0044) against which the 0.0309 remainder is measured, and that monotonicity failure is one of the readings of that rejection. The t-ratio of -1.73 follows from the record (-0.01739/0.01008 = -1.725); my earlier -1.72 came from the rounded table entries.
  • Section IV.E names the two negative cell-level IV weights (-0.061, -0.009), cites Blandhol et al. (2022) for why linear covariate adjustment can produce them, calls the cell term descriptive rather than a positively weighted average of cell-specific effects, and reads the 34-degree-of-freedom rejection (p = 0.012) as a statement about the estimators' weights rather than about compliers.
  • The joint design is now reported with the right qualifications in both the text and the Table 7 note: the over-identification test 'does not reject' (J = 1.27, p = 0.260) with the low-power caveat, the within-margin set [0.0000, 0.2382] is tied to a bootstrap p of 0.050 (record 0.04999), and the fixed first-stage index weights are stated to make the joint sets if anything too narrow.
  • The replication sentence now says what the executed script does: it reproduces every table, while the data file's SHA-256 is verified at provisioning rather than by the script. The two reference page ranges are correct (Hungerford and Solon 69(1): 175-77; Jaeger and Page 78(4): 733-40), and the Card attribution is now specific enough to be checked.
  • The record is unchanged (same log, script and RESULT_JSON as in round 2), so every number verified in rounds 1 and 2 still holds, and the figures in the altered passages (-0.0174, -1.73, -0.0136, -0.143, -0.112, 0.0047, 0.0091, 0.0044, 0.0309, 0.050, 1.27, 0.260, -0.061, -0.009, 0.012) all trace to it. No overstatement has crept in with the additions; the conclusion still opens with the point estimate being nearly twice OLS, a difference the data do not establish, and the refutation is stated as relative to the observational benchmark.

Weaknesses

  • The contribution remains narrow for this journal: a precisely bounded zero on one explanation of a gap that the paper itself shows is not statistically established, using a borrowed decomposition on a textbook extract. This has not changed and could not be changed by revision; it is a matter for the editor's judgement of fit rather than a defect in the manuscript.
  • The within-margin term is uninformative by design (minimum detectable effect 2.5 times OLS; in reduced-form units larger than the reduced form itself), so the paper says where the gap is not rather than what it is. The paper now states this plainly.
  • The two-year-college evidence rests on an instrument with F = 2.43 whose monotonicity implication is not clearly met and whose exclusion is in doubt; the paper now lists all of these readings, so the finding is correctly presented as a prediction failure with several possible causes rather than as evidence for any one of them.
  • Covariates are saturated only across seven race x South x SMSA cells; family background could not be brought into the cell decomposition with a usable first stage. The paper says so.

Required changes

—

Record

§Event timeline

  1. Run createdQuestion: “Why is the college-proximity IV estimate of the return to schooling nearly twice the OLS estimate? Testing whether the gap is an Angrist-Imbens margin-weighting artefact' language=en discipline=economics venue='American Economic Review” · output language: English · models: agent:claude-opus-5-5
  2. Stage started · Literature reviewAttempt 1 · agent:claude-opus-5-5
  3. awaiting_agent · Literature reviewsearch response for literature_review: literature_review-search-1.response.json
  4. Resume requested
  5. Stage started · Literature reviewAttempt 2 · agent:claude-opus-5-5
  6. Search complete · Literature review30 sources · 18,862 characters of notes
  7. awaiting_agent · Literature reviewcomplete_json response for literature_review: literature_review-complete_json-1.response.json
  8. Resume requested
  9. Stage started · Literature reviewAttempt 3 · agent:claude-opus-5-5
  10. Search complete · Literature review30 sources · 18,862 characters of notes
  11. Stage completed · Literature review
  12. Stage started · IdeationAttempt 1 · agent:claude-opus-5-5
  13. awaiting_agent · Ideationcomplete_json response for ideation: ideation-complete_json-1.response.json
  14. Resume requested
  15. Stage started · IdeationAttempt 2 · agent:claude-opus-5-5
  16. Awaiting approval · Ideation
  17. Approved · IdeationIdea 1 selected
  18. Stage started · ExperimentationAttempt 1 · agent:claude-opus-5-5
  19. Dataset provisioned · Experimentationcard_college_proximity · 3,010 observations · 34 variables · sha256 326dad817467
  20. awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
  21. Resume requested
  22. Stage started · ExperimentationAttempt 2 · agent:claude-opus-5-5
  23. Dataset provisioned · Experimentationcard_college_proximity · 3,010 observations · 34 variables · sha256 326dad817467
  24. Experiment round · ExperimentationRound 1: exit 0; result parsed (9.579s)
  25. Stage completed · Experimentation
  26. Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
  27. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  28. Resume requested
  29. Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
  30. Stage completed · Analysis
  31. Stage started · Paper writingAttempt 1 · agent:claude-opus-5-5
  32. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  33. Resume requested
  34. Stage started · Paper writingAttempt 2 · agent:claude-opus-5-5
  35. Stage completed · Paper writing
  36. Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
  37. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  38. Resume requested
  39. Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
  40. Awaiting approval · Review
  41. Sent back for revision · ReviewRe-running from Paper writing: “The referee recommends minor revision. Address each required change below using only the results the executed script already produced; do not add analyses or numbers that are not in RESULT_JSON.\n\nRequ”
  42. Stage started · Paper writingAttempt 3 · agent:claude-opus-5-5
  43. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  44. Resume requested
  45. Stage started · Paper writingAttempt 4 · agent:claude-opus-5-5
  46. Stage completed · Paper writing
  47. Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
  48. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  49. Resume requested
  50. Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
  51. Awaiting approval · Review
  52. Approved · Review
  53. Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/economics-en/runs/ad07220e3e37/REPORT.md
  54. Reopened after sign-off · ReviewReopened after sign-off; re-running from Review: “Independent review: the paper is re-refereed by an independent referee (Claude Fable 5.1) that took no part in writing it.”
  55. Stage started · ReviewAttempt 3 · agent:claude-fable-5-1
  56. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  57. Resume requested
  58. Stage started · ReviewAttempt 4 · agent:claude-fable-5-1
  59. Awaiting approval · Review
  60. Sent back for revision · ReviewRe-running from Experimentation: “Independent referee (Claude Fable 5.1): major_revision. Required changes: (1) [analysis] Run the two-year-college proximity design from the same file: (a) nearc2 as the sole instrument, with its margi”
  61. Stage started · ExperimentationAttempt 3 · agent:claude-opus-5-5
  62. Dataset provisioned · Experimentationcard_college_proximity · 3,010 observations · 34 variables · sha256 326dad817467
  63. awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
  64. Resume requested
  65. Stage started · ExperimentationAttempt 4 · agent:claude-opus-5-5
  66. Dataset provisioned · Experimentationcard_college_proximity · 3,010 observations · 34 variables · sha256 326dad817467
  67. Experiment round · ExperimentationRound 1: exit 0; result parsed (19.794s)
  68. Stage completed · Experimentation
  69. Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
  70. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  71. Resume requested
  72. Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
  73. Stage completed · Analysis
  74. Stage started · Paper writingAttempt 5 · agent:claude-opus-5-5
  75. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  76. Resume requested
  77. Stage started · Paper writingAttempt 6 · agent:claude-opus-5-5
  78. Stage completed · Paper writing
  79. Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
  80. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  81. Resume requested
  82. Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
  83. Awaiting approval · Review
  84. Sent back for revision · ReviewRe-running from Paper writing: “Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [text] In Section IV.F and the Table 4 note, state that the two-year-college shifts are negative at thresholds 11”
  85. Stage started · Paper writingAttempt 7 · agent:claude-opus-5-5
  86. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  87. Resume requested
  88. Stage started · Paper writingAttempt 8 · agent:claude-opus-5-5
  89. Stage completed · Paper writing
  90. Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
  91. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  92. Resume requested
  93. Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
  94. Awaiting approval · Review
  95. Sent back for revision · ReviewRe-running from Analysis: “Editorial alignment after acceptance (no new referee request): make the analysis record agree with the accepted manuscript. The two-instrument over-identification test 'does not reject' (Hansen J p = ”
  96. Stage started · AnalysisAttempt 3 · agent:claude-opus-5-5
  97. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  98. Resume requested
  99. Stage started · AnalysisAttempt 4 · agent:claude-opus-5-5
  100. Stage completed · Analysis
  101. Stage started · Paper writingAttempt 9 · agent:claude-opus-5-5
  102. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  103. Resume requested
  104. Stage started · Paper writingAttempt 10 · agent:claude-opus-5-5
  105. Stage completed · Paper writing
  106. Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
  107. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  108. Resume requested
  109. Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
  110. Awaiting approval · Review
  111. Approved · Review
  112. Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/economics-en/runs/ad07220e3e37/REPORT.md