Datasets:
design stringclasses 4
values | series stringlengths 18 40 | n int64 50 16k | T_pre float64 2 8 ⌀ | phi float64 0 0.8 ⌀ | rho_hat float64 0 2.1 | cov_at_pow50 float64 0 0.85 ⌀ | cov_at_pow80 float64 0 0.72 ⌀ | min_cov_undetected float64 0 0.92 | max_b float64 2.03 6.81 | size_at_0 float64 0.01 0.06 | cov_at_0 float64 0.94 0.96 | visible bool 2
classes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
Difference-in-differences | pre-trend test, Ashenfelter dip | 50 | 4 | 0.5 | 0.865622 | 0.382923 | 0.124015 | 0.4725 | 6.06306 | 0.0505 | 0.946 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 2 | 0 | 0.784308 | 0.284276 | 0.067393 | 0.289 | 6.002812 | 0.053 | 0.9555 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 2 | 0.5 | 1.014095 | 0.506134 | 0.221308 | 0.68 | 6.015377 | 0.0445 | 0.946 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 2 | 0.8 | 1.205042 | 0.626012 | 0.359596 | 0.6835 | 6.015895 | 0.058 | 0.945 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 4 | 0 | 0.795364 | 0.284454 | 0.081389 | 0.291 | 6.045765 | 0.0535 | 0.9445 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 4 | 0.5 | 0.902546 | 0.384321 | 0.132541 | 0.457 | 6.01267 | 0.0545 | 0.949 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 4 | 0.8 | 1.010348 | 0.485389 | 0.222056 | 0.49 | 5.979749 | 0.058 | 0.9445 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 8 | 0 | 0.726903 | 0.221622 | 0.051604 | 0.2905 | 5.972905 | 0.0445 | 0.951 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 8 | 0.5 | 0.793078 | 0.289266 | 0.081119 | 0.3205 | 5.994543 | 0.05 | 0.948 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 200 | 8 | 0.8 | 0.891573 | 0.365597 | 0.134558 | 0.479 | 6.001675 | 0.0415 | 0.952 | true |
Difference-in-differences | pre-trend test, Ashenfelter dip | 1,000 | 4 | 0.5 | 0.906701 | 0.40994 | 0.136183 | 0.5055 | 5.973886 | 0.0535 | 0.946 | true |
Difference-in-differences | pre-trend test, coincident shock | 50 | 4 | 0.5 | 0 | null | null | 0 | 6.045757 | 0.0445 | 0.9625 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 2 | 0 | 0 | null | null | 0 | 6.045124 | 0.052 | 0.944 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 2 | 0.5 | 0 | null | null | 0 | 5.982204 | 0.0535 | 0.9505 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 2 | 0.8 | 0 | null | null | 0 | 6.006216 | 0.048 | 0.9455 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 4 | 0 | 0 | null | null | 0 | 5.996461 | 0.041 | 0.951 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 4 | 0.5 | 0 | null | null | 0 | 5.966587 | 0.0455 | 0.944 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 4 | 0.8 | 0 | null | null | 0 | 6.015918 | 0.0435 | 0.9515 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 8 | 0 | 0 | null | null | 0.0005 | 5.991993 | 0.0515 | 0.9525 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 8 | 0.5 | 0 | null | null | 0 | 6.012465 | 0.044 | 0.95 | false |
Difference-in-differences | pre-trend test, coincident shock | 200 | 8 | 0.8 | 0 | null | null | 0 | 6.039072 | 0.0485 | 0.9535 | false |
Difference-in-differences | pre-trend test, coincident shock | 1,000 | 4 | 0.5 | 0 | null | null | 0 | 6.024727 | 0.047 | 0.956 | false |
Difference-in-differences | pre-trend test, linear trend | 50 | 4 | 0.5 | 0.72425 | 0.244814 | 0.050018 | 0.302 | 5.994866 | 0.048 | 0.9495 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 2 | 0 | 0.317035 | null | null | 0 | 5.991463 | 0.0495 | 0.9415 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 2 | 0.5 | 0.409153 | 0.006615 | null | 0.0215 | 6.025185 | 0.0475 | 0.9565 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 2 | 0.8 | 0.489615 | 0.028751 | 0.000263 | 0.152 | 6.04161 | 0.054 | 0.9495 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 4 | 0 | 0.823478 | 0.282817 | 0.094516 | 0.4935 | 5.985198 | 0.047 | 0.953 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 4 | 0.5 | 0.769066 | 0.267704 | 0.066 | 0.3055 | 6.032939 | 0.047 | 0.9485 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 4 | 0.8 | 0.759469 | 0.244666 | 0.069853 | 0.276 | 5.968629 | 0.0485 | 0.9435 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 8 | 0 | 2.101515 | 0.849662 | 0.716051 | 0.919 | 6.023116 | 0.045 | 0.9485 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 8 | 0.5 | 1.510561 | 0.726713 | 0.551866 | 0.8265 | 6.035268 | 0.047 | 0.955 | true |
Difference-in-differences | pre-trend test, linear trend | 200 | 8 | 0.8 | 1.208059 | 0.600307 | 0.383981 | 0.6685 | 6.003167 | 0.0545 | 0.9575 | true |
Difference-in-differences | pre-trend test, linear trend | 1,000 | 4 | 0.5 | 0.764474 | 0.266975 | 0.064969 | 0.295 | 5.966084 | 0.049 | 0.954 | true |
Instrumental variables | J test, one invalid instrument | 500 | null | null | 0.79175 | 0.297906 | 0.094679 | 0.3935 | 6.513179 | 0.0505 | 0.9505 | true |
Instrumental variables | J test, one invalid instrument | 2,000 | null | null | 0.79955 | 0.308935 | 0.09203 | 0.3895 | 6.705172 | 0.0575 | 0.9525 | true |
Instrumental variables | J test, one invalid instrument | 10,000 | null | null | 0.793812 | 0.303844 | 0.090695 | 0.3925 | 6.811817 | 0.056 | 0.9395 | true |
Instrumental variables | J test, proportional violation | 500 | null | null | 0 | null | null | 0 | 6.775496 | 0.056 | 0.945 | false |
Instrumental variables | J test, proportional violation | 2,000 | null | null | 0 | null | null | 0 | 6.795342 | 0.0465 | 0.9505 | false |
Instrumental variables | J test, proportional violation | 10,000 | null | null | 0 | null | null | 0 | 6.752118 | 0.0545 | 0.947 | false |
Instrumental variables | first-stage F < 10 | 500 | null | null | 0 | null | null | 0.001 | 5.297173 | 0.009 | 0.9505 | false |
Instrumental variables | first-stage F < 10 | 2,000 | null | null | 0 | null | null | 0 | 5.33959 | 0.0115 | 0.9575 | false |
Instrumental variables | first-stage F < 10 | 10,000 | null | null | 0 | null | null | 0 | 5.331685 | 0.011 | 0.9525 | false |
Instrumental variables | zero-first-stage group, $\phi=0$ | 500 | null | null | 0 | null | null | 0.001 | 5.297173 | 0.058 | 0.9505 | false |
Instrumental variables | zero-first-stage group, $\phi=0$ | 2,000 | null | null | 0 | null | null | 0 | 5.33959 | 0.048 | 0.9575 | false |
Instrumental variables | zero-first-stage group, $\phi=0$ | 10,000 | null | null | 0 | null | null | 0 | 5.331685 | 0.046 | 0.9525 | false |
Instrumental variables | zero-first-stage group, $\phi=0.5$ | 500 | null | null | 0.185804 | null | null | 0.001 | 5.297173 | 0.058 | 0.9505 | true |
Instrumental variables | zero-first-stage group, $\phi=0.5$ | 2,000 | null | null | 0.174053 | null | null | 0 | 5.33959 | 0.048 | 0.9575 | true |
Instrumental variables | zero-first-stage group, $\phi=0.5$ | 10,000 | null | null | 0.174957 | null | null | 0 | 5.331685 | 0.046 | 0.9525 | true |
Instrumental variables | zero-first-stage group, $\phi=1$ | 500 | null | null | 0.346583 | 0.001 | null | 0.001 | 5.297173 | 0.058 | 0.9505 | true |
Instrumental variables | zero-first-stage group, $\phi=1$ | 2,000 | null | null | 0.338492 | 0.000484 | null | 0 | 5.33959 | 0.048 | 0.9575 | true |
Instrumental variables | zero-first-stage group, $\phi=1$ | 10,000 | null | null | 0.332566 | 0.000011 | null | 0.0005 | 5.331685 | 0.046 | 0.9525 | true |
Regression discontinuity | covariate test, one-sided sorting | 1,000 | null | null | 0.67583 | null | null | 0.47 | 2.0335 | 0.047 | 0.952 | true |
Regression discontinuity | covariate test, one-sided sorting | 4,000 | null | null | 0.666358 | 0.147091 | null | 0.233 | 3.527956 | 0.038 | 0.942 | true |
Regression discontinuity | covariate test, one-sided sorting | 16,000 | null | null | 0.668993 | 0.169811 | null | 0.312 | 4.102021 | 0.035 | 0.954 | true |
Regression discontinuity | covariate test, two-sided sorting on $U$ | 1,000 | null | null | 0.652297 | 0.174984 | null | 0.312 | 3.681616 | 0.045 | 0.939 | true |
Regression discontinuity | covariate test, two-sided sorting on $U$ | 4,000 | null | null | 0.659739 | 0.181905 | 0.012835 | 0.226 | 4.565134 | 0.052 | 0.946 | true |
Regression discontinuity | covariate test, two-sided sorting on $U$ | 16,000 | null | null | 0.669188 | 0.191767 | 0.011755 | 0.233 | 4.776853 | 0.046 | 0.947 | true |
Regression discontinuity | covariate test, two-sided sorting on $V$ | 1,000 | null | null | 0 | null | null | 0.121 | 3.111668 | 0.05 | 0.945 | false |
Regression discontinuity | covariate test, two-sided sorting on $V$ | 4,000 | null | null | 0 | null | null | 0.013 | 4.177501 | 0.053 | 0.957 | false |
Regression discontinuity | covariate test, two-sided sorting on $V$ | 16,000 | null | null | 0 | null | null | 0.005 | 4.664646 | 0.062 | 0.936 | false |
Regression discontinuity | density test, one-sided sorting | 1,000 | null | null | 0.73961 | 0.507746 | null | 0.47 | 2.0335 | 0.053 | 0.952 | true |
Regression discontinuity | density test, one-sided sorting | 4,000 | null | null | 0.4673 | null | null | 0.067 | 3.527956 | 0.057 | 0.942 | true |
Regression discontinuity | density test, one-sided sorting | 16,000 | null | null | 0.401384 | null | null | 0.023 | 4.102021 | 0.05 | 0.954 | true |
Regression discontinuity | density test, two-sided sorting on $U$ | 1,000 | null | null | 0 | null | null | 0.046 | 3.681616 | 0.06 | 0.939 | true |
Regression discontinuity | density test, two-sided sorting on $U$ | 4,000 | null | null | 0 | null | null | 0.008 | 4.565134 | 0.056 | 0.946 | true |
Regression discontinuity | density test, two-sided sorting on $U$ | 16,000 | null | null | 0 | null | null | 0.002 | 4.776853 | 0.05 | 0.947 | true |
Regression discontinuity | density test, two-sided sorting on $V$ | 1,000 | null | null | 0 | null | null | 0.121 | 3.111668 | 0.05 | 0.945 | false |
Regression discontinuity | density test, two-sided sorting on $V$ | 4,000 | null | null | 0 | null | null | 0.013 | 4.177501 | 0.041 | 0.957 | false |
Regression discontinuity | density test, two-sided sorting on $V$ | 16,000 | null | null | 0 | null | null | 0.005 | 4.664646 | 0.055 | 0.936 | false |
Regression discontinuity | either test, one-sided sorting | 1,000 | null | null | 0.821426 | 0.508897 | null | 0.47 | 2.0335 | 0.047 | 0.952 | true |
Regression discontinuity | either test, one-sided sorting | 4,000 | null | null | 0.676663 | 0.178338 | null | 0.233 | 3.527956 | 0.054 | 0.942 | true |
Regression discontinuity | either test, one-sided sorting | 16,000 | null | null | 0.649496 | 0.147039 | null | 0.155 | 4.102021 | 0.042 | 0.954 | true |
Regression discontinuity | either test, two-sided sorting on $U$ | 1,000 | null | null | 0.570986 | 0.071714 | null | 0.085 | 3.681616 | 0.057 | 0.939 | true |
Regression discontinuity | either test, two-sided sorting on $U$ | 4,000 | null | null | 0.587126 | 0.086217 | null | 0.226 | 4.565134 | 0.037 | 0.946 | true |
Regression discontinuity | either test, two-sided sorting on $U$ | 16,000 | null | null | 0.603943 | 0.100301 | 0.005758 | 0.233 | 4.776853 | 0.053 | 0.947 | true |
Regression discontinuity | either test, two-sided sorting on $V$ | 1,000 | null | null | 0 | null | null | 0.121 | 3.111668 | 0.054 | 0.945 | false |
Regression discontinuity | either test, two-sided sorting on $V$ | 4,000 | null | null | 0 | null | null | 0.013 | 4.177501 | 0.049 | 0.957 | false |
Regression discontinuity | either test, two-sided sorting on $V$ | 16,000 | null | null | 0 | null | null | 0.005 | 4.664646 | 0.055 | 0.936 | false |
Selection on observables | placebo outcome, $\kappa=0$ | 500 | null | null | 0 | null | null | 0 | 5.504087 | 0.052 | 0.9485 | false |
Selection on observables | placebo outcome, $\kappa=0$ | 2,000 | null | null | 0 | null | null | 0 | 5.841247 | 0.0485 | 0.9425 | false |
Selection on observables | placebo outcome, $\kappa=0$ | 10,000 | null | null | 0 | null | null | 0 | 5.929053 | 0.055 | 0.947 | false |
Selection on observables | placebo outcome, $\kappa=0.5$ | 500 | null | null | 0.625239 | 0.148711 | 0.008695 | 0.1775 | 5.504087 | 0.0555 | 0.9485 | true |
Selection on observables | placebo outcome, $\kappa=0.5$ | 2,000 | null | null | 0.629439 | 0.141307 | 0.01135 | 0.1645 | 5.841247 | 0.047 | 0.9425 | true |
Selection on observables | placebo outcome, $\kappa=0.5$ | 10,000 | null | null | 0.639041 | 0.156499 | 0.01094 | 0.164 | 5.929053 | 0.0575 | 0.947 | true |
Selection on observables | placebo outcome, $\kappa=1$ | 500 | null | null | 0.99573 | 0.482666 | 0.199253 | 0.4885 | 5.504087 | 0.051 | 0.9485 | true |
Selection on observables | placebo outcome, $\kappa=1$ | 2,000 | null | null | 1.003744 | 0.490856 | 0.208287 | 0.6985 | 5.841247 | 0.0475 | 0.9425 | true |
Selection on observables | placebo outcome, $\kappa=1$ | 10,000 | null | null | 0.990324 | 0.496717 | 0.193732 | 0.6715 | 5.929053 | 0.0525 | 0.947 | true |
Selection on observables | placebo, $\kappa=1$, high noise | 500 | null | null | 0.625239 | 0.148711 | 0.008695 | 0.1775 | 5.504087 | 0.0555 | 0.9485 | true |
Selection on observables | placebo, $\kappa=1$, high noise | 2,000 | null | null | 0.629439 | 0.141307 | 0.01135 | 0.1645 | 5.841247 | 0.047 | 0.9425 | true |
Selection on observables | placebo, $\kappa=1$, high noise | 10,000 | null | null | 0.639041 | 0.156499 | 0.01094 | 0.164 | 5.929053 | 0.0575 | 0.947 | true |
Selection on observables | placebo, $\kappa=1$, low noise | 500 | null | null | 1.268392 | 0.655719 | 0.390304 | 0.689 | 5.504087 | 0.049 | 0.9485 | true |
Selection on observables | placebo, $\kappa=1$, low noise | 2,000 | null | null | 1.267933 | 0.654267 | 0.38741 | 0.6985 | 5.841247 | 0.044 | 0.9425 | true |
Selection on observables | placebo, $\kappa=1$, low noise | 10,000 | null | null | 1.265481 | 0.64726 | 0.389663 | 0.6715 | 5.929053 | 0.049 | 0.947 | true |
The Identification Gap — results
Every simulation and empirical result behind the preprint The Identification Gap: What Diagnostic Tests Can and Cannot Reveal About Causal Research Designs (Thanatchon Chongmankhong, November 2026).
- 📖 Paper online: https://thanatchon36.github.io/identification-gap/paper/
- 💻 Code and replication: https://github.com/thanatchon36/identification-gap
- 🗄️ Zenodo archive: https://doi.org/10.5281/zenodo.24243854
The paper asks what a passed diagnostic test (placebo outcome, overidentification, pre-trends, density or covariate
continuity) actually certifies. For a violation that moves the estimator and the diagnostic linearly, everything depends
on the estimator's standardised bias b and the detectability ratio ρ = ν / b.
Files
Each CSV is its own config: load_dataset("thanatchon36/identification-gap", "simA_did").
| Config | Content |
|---|---|
emp_card_chr |
Card (1995): Conley–Hansen–Rossi bounds over γ_max |
emp_castle_eventstudy |
castle doctrine event-study coefficients |
emp_castle_rm |
castle doctrine relative-magnitudes bounds |
emp_lalonde |
LaLonde (NSW) estimates, placebo tests and standardised bias vs experimental benchmark |
emp_prop99_gaps |
Proposition 99 synthetic-control gaps by state and year |
emp_rdd_covariates |
House/Senate RDD covariate-continuity tests |
plasmode |
plasmode simulations on LaLonde–PSID and castle-doctrine covariates |
plasmode_info |
plasmode design metadata |
simA_did |
Experiment A: difference-in-differences, controlled violations |
simA_iv |
Experiment A: instrumental variables, controlled violations |
simA_rdd |
Experiment A: regression discontinuity, controlled violations |
simA_soo |
Experiment A: selection on observables, controlled violations |
simB_fixable |
Experiment B: five estimation problems, naive vs modern estimators |
simC_sensitivity |
Experiment C: coverage and relative width of six sensitivity analyses |
summary_series |
detectability ratio ρ̂ and coverage at 50/80% power for every simulated series |
theory_certification |
power to certify |
theory_frontier |
coverage–power frontier: coverage and diagnostic power over (ρ, b) |
theory_literature |
coverage after a pre-test, averaged over a literature of violations |
theory_mc_verification |
Monte Carlo check of the closed forms (2×10⁵ draws per cell) |
theory_pretest |
unconditional vs conditional (on passing) coverage over (r, ρ, b) |
theory_table |
coverage at 50/80/90% diagnostic power and certification power, by ρ |
JSON files in results/ hold scalar summaries (theory_key_numbers.json, emp_*.json).
Common columns in the simulation tables: R replications, bias, sd, mean_se, rmse, b (bias / sd),
coverage (of the nominal 95% CI) with coverage_mcse, power_* (rejection rate of a diagnostic), pass_*
(share passing), cov_pass_* (coverage conditional on passing).
Reproduce
git clone https://github.com/thanatchon36/identification-gap && cd identification-gap
pip install -r requirements.txt
bash run_all.sh # about 15 minutes on one core; rewrites results/ and rebuilds the paper
Raw data
The empirical inputs are public replication datasets and are not redistributed here; get them from their sources:
LaLonde/NSW (Imbens & Xu 2024, xuyiqing/lalonde), Card (1995) (Wooldridge data via Rdatasets), castle doctrine and
Lee–Moretti–Butler (Cunningham, Causal Inference: The Mixtape), Senate RDD (rdrobust), Proposition 99 (synthdid).
Citation
@misc{chongmankhong2026identification,
author = {Chongmankhong, Thanatchon},
title = {The Identification Gap: What Diagnostic Tests Can and Cannot Reveal About Causal Research Designs},
year = {2026},
note = {Preprint},
doi = {10.5281/zenodo.24243854},
url = {https://github.com/thanatchon36/identification-gap}
}
- Downloads last month
- 84