Summary
The paper compares intergenerational persistence in China and the United States across four estimand families: income elasticity (IGE), rank-rank slope, movement out of farming by parental farm origin, and schooling coefficient/correlation. No Chinese microdata are used. Every China number is a published estimate (mainly Fan, Yi & Zhang on CFPS 2010–16; Emran & Sun on rural CHIP; the World Bank GDIM). The US side combines Opportunity Atlas public tables, Chetty et al. (2014), new GSS 1972–2024 estimates and GDIM. Each China–US pair receives a comparability grade (A–D) that fixes how it is reported, and four hypotheses (H1–H4) with a 95%-CI decision rule were registered in an internal, hashed plan, with nine logged deviations.
The headline result (H1) is that the China rank-rank slope (0.494 for the 1981–88 cohort, 0.443 for the 1970–80 cohort) exceeds the US child-individual-income slope (0.287) by 0.207 and 0.156. On identical GDIM schooling measures, China shows more persistence for the 1980s cohort and less for earlier cohorts (H2). Within China, persistence is higher for daughters and for rural-hukou children, while US metro and nonmetro county slopes differ by only 0.019 (H3). IGE orderings depend on the framework (H4). The paper presents itself as a structured, provenance-labelled synthesis, not as new Chinese estimation.
Strengths
Exact transcription of published estimates
Tables 1–4, 6, 8; Sec. 4.1, 5.3
Every published value I traced back to its source matches to the printed precision. That covers 33 of 33 Fan, Yi & Zhang estimates and SEs (Table 2 Panels A–E, Table 5), all 56 GDIM education numbers plus MU050/BHQ4 bounds and the GDIM income IGEs, all 8 Emran & Sun farm-exit shares, all 7 Chetty et al. (2014) Table I values, and the Gong et al./Mazumder robustness values (checks 02, 04, 09, 16, 17). For a synthesis paper, this is the property that matters most.
US public-data estimates reproduce
Table 1; Table 5; Table 6; Table 8 R1; Sec. 4.5, 5.1
From the Opportunity Atlas files with the Sec. 5.1 WLS method, I reproduced all 9 national rank slopes within 0.0015 (kir 0.2866 vs 0.287, kfr 0.3663 vs 0.366, sex and age-24/26/29 variants). The metro/nonmetro and Census-region county means reproduce within 0.001 with exactly the printed county counts (217/1054/1420/446) (checks 19, 20). The US side of H1 and H3 is therefore solid.
Provenance labels and comparability grades on every number
Sec. 4.6; Tables 1–8 (source columns); Table 7
Each row carries a P/A/M provenance code, a source key, the cohort and the age window. Each China–US pair has a grade (A–D) that decides whether it is reported as a tested difference or shown side by side only. Table 7 carries the grade into the hypothesis outcomes. With these labels a reader can trace every number, and my provenance checks (02, 04, 09, 16, 17) relied on them directly.
Internally consistent arithmetic
Tables 1, 2, 5, 7; Eq. (5); Sec. 5.3
Seventy-seven estimate/SE/CI rows are consistent with est ± 1.96·SE (tolerance 0.0025). The H1, H2 and H3-China differences and Eq. (5) intervals reproduce from the printed inputs to rounding (e.g. H1 0.207 with CI [0.1893, 0.2247] vs printed [0.190, 0.225]). The GSS decade n's sum exactly to 32,052 (checks 01, 03, 05, 18).
Associational framing and candid limitations
Sec. 7.3; Sec. 3.2; Sec. 10
A scan of 46 causal or mechanistic phrases found no causal claim about parental resources. Sec. 7.3 states that none of the estimates 'identifies the causal effect' and lists imputation, age windows, co-residence and multiplicity as limitations (check 14). The deviation log is also unusually frank and records its own errors (e.g. the factor-100 county-slope divisor in DEV-08).
Lifecycle robustness supports the direction of H1
Sec. 6.8; Sec. 7.1; Table 8 R1, Fig. 6
Using the US age profile (Table 8 R1), aligning ages widens rather than narrows the 1981–88 gap (0.272/0.234/0.215 at US ages 24/26/29 vs 0.207), as the paper says. H1 would only fail if the China slope fell by more than about 38% (1981–88) or 32% (1970–80) (check 15). This is a useful margin, although it does not address the income-construction issue below.
Weaknesses
Headline H1 pair is not on the 'same income concept'
major · Sec. 4.2; Sec. 4.6; Sec. 7.1 (p. 14); Abstract; Table 1; Table 7 (H1)
Sec. 4.2 says China child income is 'imputed for family members who do not report it'. In Fan et al. (Sec. 3.3), however, income is computed for the full sample as a Heckman-corrected fitted value from a regression on education, age, gender, hukou and region (R² 0.235 for children, 0.250 for parents). The China rank slope 0.494 is therefore a rank association between predicted incomes, while the US 0.287 uses observed tax-record income. On the same sample, this construction moves Fan's 1981–88 IGE from 0.176 (raw) to 0.435 (computed), and no raw-income rank slope is published (checks 15, 15b).
The paper also omits two further differences. Fan et al. drop incomes below 3,000 yuan, which the OI tables do not, and they rank children within the CFPS analysis sample rather than within a national cohort. Grade B (listing only the age caveat) and the Sec. 7.1 statement that China 'exceeds the US tax-record slope on the same income concept' are therefore not supported. Because the abstract's lead number depends on this pair, the issue affects the main claim.
Fix: describe the income construction accurately, downgrade the pair to grade C or attach the caveat to the abstract, and add a bound or sensitivity analysis for how prediction-based incomes shift rank slopes. One option is to apply the same prediction step to the US GSS or PSID data.
Pre-registration cannot be verified and gives little protection here
major · Title; Sec. 3.1–3.2 (DEV-03, DEV-04, DEV-08, DEV-09); Sec. 9
The plan is an internal record, not a public registry entry, and neither the plan file nor the deviation log was supplied, so the SHA-256 and the 16:57:20Z freeze time cannot be checked (check 13). Every China input and the US comparator 0.287 were public before the freeze, and H1/H3 restate orderings already visible in Fan et al. The protection that pre-registration usually provides against analytic flexibility is therefore weak.
The log also contains internal tensions:
- DEV-08 (17:13) says the registered divisor 'gives values a factor of 100 below the national slope', so a county slope had already been computed, yet DEV-04 says its rule was set 'before any county aggregate was computed'.
- DEV-09 corrects a 'clerical error' in the DEV-01–07 timestamp without giving the corrected time.
- DEV-03, DEV-04 and DEV-09 set inference rules after the freeze (US SE bound, the US-H3 equivalence band, the primary China H3 measure, the DoD SE). This conflicts in spirit with 'None changes ... the decision rule' (Sec. 3.2, p. 3).
The title's 'Pre-Registered' should be qualified. Fix: release the plan and log with independently verifiable timestamps (e.g. OSF), and give the corrected times.
Promised replication package not provided
major · Sec. 9; Tables 2, 3, 6 (provenance M); Table 1 (IGE 0.348, 0.452)
Sec. 9 promises a package with the plan, deviation log, transcribed inputs and a single script that runs in under a minute, but no code or URL/DOI accompanies the submission. All provenance-'M' numbers (GSS occupational and schooling estimates, Tables 2, 3, 6) and the US IGEs 0.348/0.452, which are computed from log cell means and cannot be rebuilt from the public percentile file (check 19), could not be verified. Fix: deposit the package with a persistent identifier.
Abstract states some results more strongly than the tables
minor · Abstract; Sec. 8; Table 2 (1940s, 1970s rows); Table 3; Table 5
(i) 'Lower for cohorts born before 1970': under the paper's own rule, the 1940s BETA difference −0.058 [−0.120, 0.004] is inconclusive. The abstract also skips the 1970s, where BETA is higher in China (+0.064) but COR is lower (−0.044), so there is no clean switch at 1970 (check 12). (ii) The farm-exit comparison is grade C. It sets China 2002 co-resident rural children, which excludes those who migrated to non-farm work, against pooled US 1994–2024 occupation codes, and has no China SE. The abstract nevertheless states it without qualification. (iii) 'US metro and nonmetro slopes differ by 0.019' refers to child-weighted means of within-county slopes, which omit the between-county component (national mean 0.252 vs national slope 0.287, check 20).
Fix: add these qualifiers to the abstract and conclusion.
H4 'consistent' verdict depends on a 0.001 rounding margin
minor · Table 1; Table 7 (H4); Sec. 6.1; DEV-05; Abstract
For the 1981–88 cohort, the GDIM difference 0.097 is smaller than the headline difference 0.098 by 0.001. Under ±0.0005 rounding of the inputs the comparison can go either way (range [−0.003, +0.001]), and the margin is about one seventh of the headline difference's SE (0.0071) (check 07). Sec. 6.1 says the split 'carries no inferential weight', but Table 7 still records 'consistent'. The headline IGE pair also compares a China individual-income IGE with a US family-income IGE. Fix: report H4 as undetermined at published precision, and do not describe the IGE ordering as a 'reversal' in the abstract without noting that the pair is not harmonised.
'Same ordering in every pair' is too strong
minor · Sec. 7.2 (p. 15); Tables 5, 6
Sec. 7.2 says the rank-rank slope 'gives the same ordering in every pair available here'. China is higher in all 36 national and sex pairs, but 8 of 72 matching subgroup pairs reverse (check 11). Examples: urban hukou 1970–80 0.225 < US metro county mean 0.249, and coastal 1970–80 0.252 < US Census-region values of 0.256–0.351. Fix: restrict the statement to the national and sex pairs.
Labelling errors in transcribed and derived rows
minor · Table 8 R1–R2; Sec. 6.8; Fig. 6b; Table 3; Table 6 note; Sec. 7.3
The values are correct, but several labels are wrong:
- Table 8 R2 'China IGE, trimmed incomes, 1981–1988 = 0.287 (0.011)' is Fan Table 2 Panel D ('Use Full Sample to Predict Lifetime Income'). The word 'trim' does not occur in Fan et al.
- R1 'children aged 24 or older' (0.499) leaves out Fan's accompanying mothers ≤ 62 restriction. That restriction is why Fig. 6b shows an age-24+ point for the 1970–80 cohort.
- Table 3 labels the prestige variant '(R3)', but R3 is the five-province check.
- The Table 6 note and Sec. 7.3 say Fan et al. report 22 provinces, whereas their Table D1 has 23 and the paper's own belt counts (8+8+7) and NBS counts (7+6+7+3) also sum to 23 (checks 02b, 08).
US BHQ4 also lies outside its own tie-break bounds
minor · Table 2 (MU050/BHQ4 rows and note); Sec. 6.2
The Table 2 note flags only MU050, but the US BHQ4 point 0.119 lies outside its own bounds [0.122, 0.133], both in the table and in the GDIM file (0.1193 vs [0.1216, 0.1333]) (check 10). Sec. 6.2's 'similar' reading still holds if the bounds are compared. Fix: extend the caveat, or compare bounds rather than point values.
Secondary literature cited imprecisely
minor · Sec. 2.1; Sec. 2.2; Sec. 6.8; Table 8 R2; ref. [3], [13], [25]
Sec. 2.2 gives Gong et al.'s single-year father–son IGE range as 0.18–0.32, but 0.183 is a father–daughter estimate; the father–son range is 0.241–0.321. Secs. 2.1 and 6.8 describe Mazumder's averaging as over father earnings, when the source uses parental family income. His 1-year and 15-year estimates also come from different samples (N 1358 vs 533), so 'raises from 0.41 to 0.68' is not a within-sample contrast (check 17). The DOI for [3] does not resolve.
Unexplained GSS sample attrition in regional rows
minor · Table 6 (GSS rows); Sec. 4.4
The GSS region-at-16 n's (6,251 + 8,878 + 9,127 + 5,259 = 29,515) fall 2,537 respondents (7.9%) short of the pooled n = 32,052, and no missing-region or foreign-born category is reported (check 18). This may also explain why all regional coefficients (0.310–0.364) lie below the pooled 0.381. Fix: report the excluded group and its coefficient.
Plain-language summary and one mechanism statement go beyond the evidence
minor · Sec. 10 (plain-language summary, p. 16); Sec. 7.2 (p. 15)
The summary's 'for every ten places that parents move up the ladder, children move up about four to five places' reads a cross-sectional rank association as a dynamic, interventional effect. Sec. 7.2's 'the comparison reflects the different stages of the shift out of agriculture' is an unhedged mechanism claim that no analysis in the paper tests (check 14). Fix: rephrase both as associations.
Questions for the authors
Can a China rank-rank slope on observed (non-predicted) income be obtained, from Fan et al. or from CFPS by an approved user, and how large is the H1 gap on that basis? If the gap shrank by more than about 38% for the 1981–88 cohort, H1 would no longer be supported.
Please release the plan file and deviation log. When was the first county-level slope computed relative to the DEV-04 decision, and what is the corrected timestamp for DEV-01 to DEV-07?
Given that Emran & Sun's sample consists of co-resident adult children and excludes migrants, in which direction and by how much would the selection bias the China farm-origin difference G? Does the China–US ordering of G survive a plausible bound?
Why are the US IGEs computed as slopes of the log of cell mean incomes rather than taken directly from CHKS? How close is that construction to the CHKS mean-of-logs estimate of 0.344 under the same restrictions?
Who are the 2,537 GSS respondents missing from the region-at-16 rows, and what is their schooling coefficient?
Does the AEJ: Policy version [24] contain the province table used for Table 6, and does it list 22 or 23 provinces?
Verification
I ran 22 scripted checks: 14 confirmed, 6 discrepancies and 2 inconclusive. The source papers (Fan et al. IZA DP, Emran & Sun, Chetty et al. NBER WP, Gong et al., Mazumder) and the public GDIM and Opportunity Atlas files were downloaded and parsed. Every transcribed China value and every reproducible US value matched, and all H1–H3 differences and intervals recompute from the tables. The discrepancies are about descriptions, not numbers. The most important is that the China income measure behind the headline H1 pair is a model-predicted income for the full sample, not observed income with imputation for non-reporters (15b). The others are label errors (02b, 08), an unflagged out-of-bounds BHQ4 value (10), an overgeneral ordering claim (11) and the abstract's 'before 1970' (12). The pre-registration audit and the robustness of H1 to the income construction remain inconclusive because no plan file or microdata were available.
H1: China rank slope minus US kir slope = 0.207 [0.190, 0.225] (1981-88) and 0.156 [0.141, 0.172] (1970-80)ConfirmedAbstract; Sec. 6.1; Table 1; Table 7 (H1); Eq. (5)
Differences 0.2070 and 0.1560 exactly. CIs from printed inputs: [0.1893, 0.2247] vs claimed [0.190, 0.225]; [0.1402, 0.1718] vs claimed [0.141, 0.172] (tolerance 0.001; lower bounds deviate by 0.0007-0.0008, reachable with unrounded inputs: lower bound range [0.1873, 0.1912]). Printed CIs are slightly asymmetric about the estimate (0.017 below / 0.018 above), consistent with rounding of unrounded inputs. All four Table 1 China CIs equal est +/- 1.96 SE.
script: 01_h1_arithmetic.pyRecomputed differences and Eq. (5) CIs from Table 1 inputs (China SE 0.009/0.008, US SE 0.001); explored +/-0.0005 rounding of inputs; checked Table 1 China CIs = est +/- 1.96 SE
All transcribed China income estimates and SEs match Fan, Yi & Zhang (IZA DP 12804): national IGE/rank, n = 22,313, sex, hukou, coastal/inland splits, R1-R3 valuesConfirmedTables 1, 4, 5, 6 (coastal block), 8 (R1-R3); Sec. 4.2; Sec. 6.8
33 of 33 values and SEs match exactly (e.g. rank 0.494 (0.009) = Fan Table 2 Panel B col 8; 0.536 (0.012) = Panel E col 8; coastal rank 0.490 (0.019) = Table 5 Panel C). n = 22,313 matches Fan Table 2 note (Panel B). Label problems recorded separately as check 02b.
script: 02_fan_provenance.pyParsed Fan et al. IZA DP 12804 Table 2 (Panels A-E) and Table 5 (Panels A-C) from the downloaded PDF text and compared each of 33 estimate/SE pairs to 3 decimals
Labels of transcribed robustness rows: R2 "China IGE, trimmed incomes, 1981-1988 = 0.287 (0.011)"; R1 "children aged 24 or older"DiscrepancyTable 8 (R1, R2); Sec. 6.8; Sec. 5.5; Fig. 6b
0.287 (0.011) is Fan Table 2 Panel D "Use Full Sample to Predict Lifetime Income", not a trimmed-income estimate; the word "trim" does not occur in Fan et al. (0 occurrences). The 3,000/1,000,000-yuan exclusion applies to all Fan panels (Table 2 note), including the headline. R1 0.499 (0.010) is Panel C, which restricts to children >=24 AND mothers <=62 (Fan Table 2 note); the paper reports only the age restriction. Values themselves are correct.
script: 02_fan_provenance.pySame parse of Fan et al. Table 2 headers and note
H2 GDIM arithmetic: BETA/COR differences and CIs for every cohort (Table 2); 1980s BETA diff 0.154 [0.121, 0.188], COR diff 0.048 [0.014, 0.082], DoD 0.106 [0.059, 0.154]; DEV-02 identity; BETA/COR ratio >1 China, <1 US; BETA gap about three times COR gapConfirmedTable 2; Table 7 (H2); Sec. 3.2 DEV-02; Sec. 6.2; Sec. 7.2; Eq. (4)-(5)
8 of 10 cohort differences/CIs match printed values within 0.0015. Two need unrounded inputs: 1960s COR CI computed [-0.2807, -0.2053] vs printed [-0.280, -0.207]; 1980s BETA diff 0.155 from printed inputs vs printed 0.154; both exactly reproducible with inputs inside rounding (e.g. 0.4825/0.3280). DoD 0.107 from printed estimates (0.106 from printed diffs), SE 0.0242 (printed 0.024), CI [0.0596, 0.1544] vs [0.059, 0.154]. DEV-02 identity = 0.48256 (printed 0.4825, rounds to published 0.483). BETA/COR: China 1.048 > 1, US 0.794 < 1. BETA gap / COR gap = 3.21 ("about three times" OK). Note 1940s BETA CI [-0.1194, 0.0054] includes 0 (see check 12).
script: 03_h2_gdim_arithmetic.pyRecomputed all 10 differences and Eq. (5) CIs from printed estimates/SEs; for the two values off by >0.0015 searched whether unrounded inputs within +/-0.0005 reproduce the printed numbers; recomputed DoD SE (sum of four squared SEs), identity and ratios
GDIM values equal the public World Bank files: education BETA/COR with SEs for China/US 1940s-1980s (parent=max), 1980s sex splits, MU050/BHQ4 with bounds; GDIM income father-son IGE China 0.517, US 0.614ConfirmedTable 2; Table 4 (GDIM rows); Table 1 (GDIM income rows); Sec. 4.4; Sec. 3.2 DEV-02
All 14 BETA/COR rows (56 numbers incl. SEs) match (e.g. CHN 1980 BETA 0.4825 (0.0122), COR 0.4609 (0.0103); USA 1980 BETA 0.3285 (0.0120), COR 0.4130 (0.0137)). Surveys are CFPS 2012 and PSID 2015 as stated. MU050/BHQ4 points and bounds match. DEV-02 identity holds in file (0.46088*4.22167/4.03218 = 0.48254 = BETA). GDIM income: CHN 0.5167, USA 0.613884 (paper 0.517/0.614); file has no SE column, consistent with "n.p.". Side findings: in the file the US BHQ4 point 0.119 lies OUTSIDE its bounds [0.122, 0.133] (see check 10); MU050 is outside its bounds in 4,720 of 6,700 rows (70.4%), supporting the paper note "most rows"; BHQ4 outside its bounds in 898 of 6,671 rows (13.5%).
script: 04_gdim_provenance.pyDownloaded GDIM_2023_03.csv (sha256 3c7994f6...) and IGE_Munoz_VanderWeide_June2025.dta (sha256 5c24608f...) from datacatalogfiles.worldbank.org; filtered CHN/USA, parent=max; compared every value to 3 decimals; counted rows where tie-break point lies outside its bounds
H3 China: eight within-China contrasts (rural-urban hukou, daughters-sons; rank and IGE; both cohorts) and CIs in Table 7 follow from Tables 4-5; all exclude zeroConfirmedTable 4; Table 5; Table 7 (H3 China rows); Sec. 6.4-6.5
All 8 differences, SEs and CIs match within 0.001 (e.g. rural-urban rank 1981-88: 0.075, SE 0.0163, CI [0.0431, 0.1069] vs printed 0.075 (0.016) [0.043, 0.107]). All lower bounds > 0 (smallest 0.0431). Subsamples are disjoint in Fan et al. (n 10,603/11,710; 6,647/15,666).
script: 05_h3_china.pyEq. (5) differences, SEs and CIs from printed Fan et al. subgroup estimates (themselves verified in check 02)
H3 US: nonmetro minus metro county-mean kir slope = 0.019 [0.013, 0.024], inside the +/-0.05 band; DEV-08 divisor 0.5 is correct and /50 gives a factor-100 errorConfirmedTable 5; Table 7 (H3 US); Sec. 4.5; Sec. 3.2 DEV-04, DEV-08
0.268-0.249 = 0.019; SE 0.0032; normal CI [0.0128, 0.0252] vs printed bootstrap CI [0.013, 0.024] (tolerance 0.0015; printed CI is a resampling interval). All US H3 variants lie inside (-0.05, 0.05). Synthetic: OLS slope 0.3000 = (p75-p25)/0.5 = 0.3000, while /50 gives 0.0030 (factor 100), so the DEV-08 correction is mathematically right. Whether the actual county values are right is tested in check 20.
script: 06_h3_us_dev08.pyDifference and normal-approximation CI from printed group means/SEs; band containment for all US H3 rows; synthetic linear county (true slope 0.30, child ranks on 0-1 scale) comparing OLS slope with (p75-p25)/0.5 and /50
H4: GDIM difference 0.097 is smaller (by 0.001) than headline difference 0.098 (1981-88) and larger than 0.046 (1970-80); Table 7 outcome consistent / not consistentConfirmedTable 1; Table 7 (H4); Sec. 6.1; DEV-05
Printed arithmetic is right: 0.614-0.517 = 0.097, 0.442-0.344 = 0.098, 0.390-0.344 = 0.046. But the 1981-88 verdict is not determined at printed precision: GDIM - headline ranges over [-0.0030, +0.0010]; with unrounded GDIM values (0.097184) the headline difference interval [0.0970, 0.0990] still straddles it. The margin (0.001) is about 1/7 of the headline difference SE (0.0071). The paper itself says the split "carries no inferential weight" (Sec. 6.1), but Table 7 still lists "consistent". The headline pair also mixes China individual-income IGE with US family-income IGE.
script: 07_h4_knife_edge.pyRecomputed differences; propagated +/-0.0005 rounding on each 3-decimal input; repeated with unrounded GDIM income file values (check 04); compared the 0.001 margin with the sampling SE of the headline difference
China belt means/medians/ranges (Table 6) and NBS four-region means (Table 8 R5) follow from Fan et al. province IGEs; 'Fan et al. (2021) report 22 provinces'DiscrepancyTable 6 (China belts, note); Table 8 R5; Sec. 6.6; Sec. 7.3
All 6 belt rows and 4 NBS rows reproduce: e.g. East 1981-88 mean 0.3822 (printed 0.382), median 0.3725 (printed 0.372), range 0.067-0.636 exact; NBS Northeast 0.3157 (printed 0.316). Values confirmed. But Fan Table D1 is titled 'IGE Estimates by Cohort in 23 Provinces/Municipalities' and 23 provinces are parsed; the paper's own belt counts sum to 8+8+7 = 23 and NBS counts to 7+6+7+3 = 23, and the 8 listed absent provinces imply 31-8 = 23. The Table 6 note and Sec. 7.3 ('covers 22 provinces') state 22. Minor factual error. Also: Table 6/8 cite [24] (AEJ 2021) while Table D1 is in the IZA DP online appendix.
script: 08_regional_belts.pyParsed Fan, Yi & Zhang (IZA DP 12804) appendix Table D1 (23 provinces, early/late cohort IGEs) and recomputed means, medians, ranges and counts under standard three-belt and NBS four-region assignments; tolerance 0.0006 on means/medians, exact on ranges and counts.
Occupational mobility (Table 3): China shares and farm-origin differences G = 0.575, 0.617, 0.283, 0.298 transcribed from Emran & Sun (2015); US differences 0.049/0.065/0.027, father-only 0.078/0.127/0.021, fishing-forestry 0.049 and their CIs; subgroup n'sConfirmedTable 3; Eq. (3); Sec. 6.3
All 8 China shares match Emran & Sun Table 5 Panel B exactly (e.g. sons 1988 neither 16.4%, both 73.9%; daughters 2002 43.0% / 72.8%); G = 0.575, 0.617, 0.283, 0.298 reproduce exactly. Sample definition (ages 18-60) matches. All 7 US differences reproduce; normal CIs within 0.002 of printed bootstrap CIs (e.g. sons [0.0421, 0.0879] vs [0.042, 0.086]). n: 377+379 = 756, 1174+1274 = 2448. SEI/prestige CIs = est +/- 1.96 SE. Note: Emran & Sun define non-farm as >=3 months of non-farm work in the survey year, the US uses occupation codes; with US shares of 0.94-0.99 the US G is bounded near the ceiling, which the grade C label only partly conveys. US GSS microdata values themselves were not re-estimated.
script: 09_occupation.pyDownloaded Emran & Sun, World Bank PRWP 7316 (PDF sha256 20d70148e980b49c...), parsed Table 5 Panel B, recomputed G; recomputed US differences and normal-approximation CIs from printed shares/SEs (tolerance 0.0011 on differences, 0.003 on CI ends since printed CIs are bootstrap); summed subgroup n's.
Only MU050 lies outside its own tie-breaking bounds; BHQ4 is similar (0.125 China, 0.119 US) with overlapping boundsDiscrepancyTable 2 (MU050/BHQ4 rows and note); Sec. 6.2
Values and bounds match the file to 3 dp. China MU050 36.4852 outside [38.2429, 40.2975] and US MU050 35.5089 outside [37.7817, 38.0018], as the paper says. But US BHQ4 = 0.1193 is also OUTSIDE its own bounds [0.1216, 0.1333] (by 0.0023), visible in Table 2 itself (0.119 vs [0.122, 0.133]). The note flags only MU050, and Sec. 6.2 calls BHQ4 'similar' using 0.119. The bounds do overlap, so the qualitative 'similar' reading survives if bounds are used instead of the point value. Minor, but the caveat should extend to US BHQ4.
script: 10_bhq4_bounds.pyRead the China/USA 1980s parent=max rows of the GDIM 2023 CSV and tested each random-tie-break value against its own lb/ub.
'The rank-rank slope ... gives the same ordering in every pair available here' (China above US)DiscrepancySec. 7.2; Tables 1, 4, 5, 6, 8
China is above the US in all 36 national and sex pairs (e.g. lowest China 0.390 sons 1970-80 vs highest US sons 0.366 kfr). But 8 of 72 subgroup pairs reverse: urban hukou 1970-80 0.225 (Table 5) < US metro county-mean kir 0.249 (Table 5), and coastal rank slope 1970-80 0.252 (Table 6) < US Census-region values 0.260/0.269/0.262 (kir) and 0.326/0.350/0.351/0.256 (CHKS family). The paper notes these subgroup concepts differ (grade C, county means omit between-county variance), so the reversals are not like-for-like, but the sentence 'every pair available here' is too strong; it holds for the national and sex pairs only.
script: 11_rank_ordering.pyEnumerated all 72 China-vs-US rank-rank slope pairs on matching breakdowns (national, daughters, sons, urban/metro, rural/nonmetro, regions) from printed table values and counted orderings.
Abstract/Conclusion statements match the tabulated results, in particular 'persistence is higher in China for the 1980s birth cohort (difference 0.154) and lower for cohorts born before 1970' on identical GDIM schooling measuresDiscrepancyAbstract; Sec. 7.1; Sec. 8; Table 2; Table 7; decision rule Sec. 3.1/5.4
1980s: BETA +0.154 [0.121, 0.188] and COR +0.048 [0.014, 0.082], both above zero, consistent. Before 1970: COR differences are below zero for all three decades, and BETA for 1950s (-0.160 [-0.205, -0.116]) and 1960s (-0.120 [-0.157, -0.082]); but the 1940s BETA difference -0.058 [-0.120, 0.004] includes 0, i.e. 'inconclusive' under the paper's own rule, so 'lower for cohorts born before 1970' holds for the coefficient only for 1950s-1960s. The abstract also skips the 1970s, where BETA is higher in China (+0.064 [0.025, 0.102]) while COR is lower (-0.044 [-0.081, -0.006]), so the reversal point is not cleanly 1970. All other abstract numbers (0.494/0.287/0.207 [0.190, 0.225], within-China orderings, US 0.019) match the tables. The farm-origin claim (China G 0.283/0.298 vs US 0.049) has no China SE and cannot be tested under the rule. Minor wording overstatement, not a main-claim error.
script: 12_abstract_vs_results.pyApplied the registered decision rule (95% CI entirely on one side) to every Table 2 cohort difference for BETA and COR; cross-checked each other abstract number against the tables.
Plan frozen 2026-11-09T16:57:20Z before any analysis; nine logged deviations, none of which changes an estimand, a comparison pair or the decision rule; plan hash 80c451e8...InconclusiveSec. 3.1; Sec. 3.2 (DEV-01 to DEV-09); title ('Pre-Registered'); Sec. 9
Countable parts hold: 9 distinct deviations DEV-01..09; DEV-08 logged 17:13 (16 min after freeze) and DEV-09 17:20 (23 min). The registration itself cannot be verified: it is an internal record, not a public registry entry (paper says so), and no plan file or replication package was supplied (code_url null), so the SHA-256 cannot be recomputed and the freeze time cannot be independently dated. Internal tensions found: (i) DEV-08 states that the registered divisor 'gives values a factor of 100 below the national slope from the same file', so a county slope had been computed before 17:13, while DEV-04 claims it was 'decided before any county aggregate was computed'; the DEV-01..07 header time was itself 'corrected' post hoc in DEV-09 ('clerical error'), so the ordering rests on an edited timestamp. (ii) Several deviations set inference choices after the freeze (DEV-03 US SE = 0.001; DEV-09 makes the rank slope primary for H3-China and fixes the DoD SE), which bear on 'none changes ... the decision rule' in spirit though not in the letter. None of these changes a reported conclusion as far as can be seen (H1-H3 intervals are far from their boundaries).
script: 13_prereg_audit.pyText audit script over Sec. 3 of paper.txt: extracted freeze time, deviation ids, logged times and 'decided before' qualifiers; checked internal ordering; looked for the plan file among the supplied inputs.
The paper uses associational, not causal, language throughout ('none identifies the causal effect', Sec. 7.3)ConfirmedAbstract; Secs. 6-8; Sec. 10 plain-language summary (paper.txt lines 900-921)
No hit makes a causal claim about parental resources: 'causal' appears only in the disclaimer (L870: 'none identifies the causal effect of parental resources'), the plain-language summary ends 'they do not show why the patterns exist', and most 'because' hits explain measurement or grading choices; L828 'because the dispersion ... differs' is the Eq. (4) identity. Minor exceptions: (i) the plain-language summary (L922-924) renders the cross-sectional rank slope as 'For every ten places that parents move up the ladder, children move up about four to five places', a dynamic/interventional reading of an association (values themselves match 0.443-0.494 and 0.287); (ii) interpretive mechanism statements without supporting analysis: 'consistent with a role for women's labour supply and spouses' earnings' (Sec. 7.2, hedged) and 'the comparison reflects the different stages of the shift out of agriculture' (L835, unhedged); (iii) the title's 'Children's Outcomes' and the abstract's 'predict' are neutral. Overall the associational framing holds.
script: 14_causal_language.pyRegex scan of the paper text (references excluded) for 18 causal/mechanistic patterns; each of 46 hits read in context.
H1 (China rank slope above US) is robust to measurement choices; 'If the China cohort's younger ages were to understate its slope in the same way as in the US data, the H1 difference would widen rather than narrow'InconclusiveSec. 6.8; Sec. 7.1; Sec. 7.3; Table 8 R1-R2
Lifecycle part confirmed arithmetically: matching the US at ages 24/26/29 widens the 1981-88 difference to 0.272/0.234/0.215 (vs 0.207), and the China 24+ restriction moves 0.494 to 0.499. Break-even: H1 stays supported unless the China slope falls by more than 38.3% (1981-88, to 0.3047) or 31.6% (1970-80, to 0.3028). Fan's alternative full-sample prediction (Panel D) gives rank 0.475/0.419, so H1 survives (diffs 0.188/0.132). But the income-construction step moves Fan's late-cohort IGE from 0.176 (raw) to 0.435 (computed) on the same observed-pair sample, a 59.5% difference, and Fan publish no raw-income rank slope. Whether the rank slope is equally sensitive cannot be determined from published tables, so robustness of H1 to the income construction is untested, not shown.
script: 15_h1_sensitivity.pyRecomputed H1 under the US age profile (Table 8 R1), computed the break-even fall in the China slope for the registered rule, and compared it with the shifts that income construction produces in Fan et al. Table 2 (parsed from the IZA DP).
China income measure: 'Child income is the child's individual income, imputed for family members who do not report it ("computed" income)'; H1 pair graded B ('same estimand with a listed caveat'), Sec. 7.1 'on the same income concept'DiscrepancySec. 4.2 (paper.txt l. 238-240); Sec. 4.6 grade B; Sec. 7.1
Fan et al. 'compute income for the full sample of adult children with and without income using individual characteristics Xi, the calculated inverse Mills ratio and the estimated coefficients', i.e. a fitted value from a log-income regression on education, age polynomials, gender, hukou and region (R-squared 0.235 children, 0.250 parents) for every person, not only for non-reporters. On the same observed-pair sample, computed vs raw income changes the late-cohort IGE from 0.176 to 0.435. The China 'rank-rank slope' 0.494 is therefore a rank association between predicted incomes, while the US kir 0.287 uses observed tax-record income. The paper's description understates the difference, and 'same income concept' (Sec. 7.1) and grade B for H1 are hard to sustain; the predicted-income construction strips idiosyncratic income variation and could raise persistence measures (direction plausible, size not quantifiable from published tables).
script: 15_h1_sensitivity.pyParsed Fan, Yi & Zhang IZA DP 12804 Sec. 3.3 and Table 2 Panel A (script 15_h1_sensitivity.py).
Published US values from CHKS 2014: rank slope family income 0.341, individual income 0.287, IGE 0.344, p10-p90 IGE 0.452; by sex individual-income rank 0.257/0.317, family IGE 0.342/0.349; with CIsConfirmedTable 1; Table 4; Sec. 2.1; Sec. 6.1; Sec. 6.4
All 7 values match Table I exactly: IGE 0.344 (0.0004) -> [0.343, 0.345]; family rank 0.341 (0.0003) -> [0.340, 0.342]; individual rank 0.287 (0.0003) -> [0.286, 0.288]; female/male individual rank 0.257/0.317; female/male IGE 0.342/0.349 (SEs 0.0005/0.0006, printed as 0.001). p10-p90 IGE 0.452 is in the CHKS text; core sample is the 1980-1982 cohorts as stated.
script: 16_chks_provenance.pyDownloaded NBER WP 19843 (PDF sha256 07930805469a474f...), parsed Table I rows 1, 4, 7 (core/male/female) with SEs, recomputed est +/- 1.96 SE rounded to 3 dp.
Secondary published values: Gong et al. urban China father-son IGE 0.241 (0.033) single-year and 0.603 (0.086) permanent; ranges 0.18-0.32 and 0.45-0.74; Mazumder 0.414 (0.075) with 1 year and 0.680 (0.134) with 15 years of father incomeConfirmedTable 8 R2; Sec. 2.1; Sec. 2.2; Sec. 6.8
All four Table 8 values and SEs match exactly (Gong T2 0.241 [0.0330], T3 0.603 [0.086]; Mazumder T1 row 1 0.414 (0.075) N=1358, row 15 0.680 (0.134) N=533). Two minor labelling errors: (i) Sec. 2.2 'father-son IGEs of 0.18 to 0.32 with single-year father income' - the father-son range is 0.241-0.321; 0.183 is a father-daughter estimate (the 0.45-0.74 permanent-income range is correctly father-son, 0.445-0.735). (ii) Sec. 2.1 and 6.8 describe Mazumder's averaging as 'father earnings', but his Table 1 is 'father-son IGE in family income' (parent-generation family income). Also the 1-year and 15-year estimates come from different samples (N 1358 vs 533).
script: 17_secondary_provenance.pyDownloaded Gong, Leigh & Meng IZA DP 4811 (sha256 599bf5bb...) and Mazumder Chicago Fed WP 2015-04 (sha256 ad0b493e...); parsed Gong Tables 2-3 and Mazumder Table 1.
Internal consistency of reported CIs and sample sizesConfirmedTables 2, 3, 5, 6, 8; Sec. 5.3 decade rows (paper.txt L445-449); Table 6 region n columns; Sec. 3 (n = 32,052)
77 est/SE/CI rows screened; all consistent with est +/- 1.96 SE within 0.0025 (the single flag, L450 BHQ4, is a parse artefact of a bounds row). GSS decade n's 8369+9757+6872+4289+2765 = 32052, exactly matching n = 32,052. Pooled 0.381 lies inside the decade range 0.369-0.391 (n-weighted mean 0.3848). Minor: Table 6 GSS region-at-16 n's (6251, 8878, 9127, 5259) sum to 29515, which leaves 2537 respondents (7.9%) of the 32,052 unaccounted for. No missing-region category is reported. Atlas county n sum 3137, CHKS county n sum 2765.
script: 18_internal_consistency.pyRegex-parsed every est/SE/[lo,hi] row in paper.txt and compared CI half-width with 1.96*SE and centre with estimate (tol 0.0025); summed decade and region n's; compared pooled coefficient with weighted decade means
US national rank-rank slopes reproduced from Opportunity Atlas public tables (provenance A)ConfirmedTable 1 (kir 0.287, kfr 0.366); Sec. 6.3 / Table 4 (daughters 0.246, sons 0.325); Table 8 (ages 24/26/29: 0.222/0.260/0.279); Sec. 5.1 method
9/9 slopes reproduced within 0.0015: kir 0.2866 (paper 0.287), kfr 0.3663 (0.366), kir daughters 0.2460 (0.246), sons 0.3251 (0.325), kfr daughters 0.3671 (0.367), kfr sons 0.3656 (0.366), kir at 24/26/29 0.2220/0.2604/0.2787 (0.222/0.260/0.279). The IGEs 0.348 and 0.452 could not be reproduced because the file has no cell mean dollar incomes.
script: 19_us_national_reproduction.pyWLS of mean child rank on parent percentile over the 100 cells of the Atlas national_percentile_outcomes.csv (Census mirror), weights = child count per cell, as described in Sec. 5.1
US county-based values: metro/nonmetro (H3 US) and Census-region child-weighted means of county kir slopes (provenance A)ConfirmedTable 5 (metro 0.249 [0.244, 0.254], nonmetro 0.268 [0.266, 0.270], difference 0.019 [0.013, 0.024]); Table 6 US Census regions (0.260/0.269/0.262/0.212; n 217/1054/1420/446); Sec. 4.5; DEV-08
All values reproduced within 0.001: metro 0.2491 SE 0.0027 [0.244, 0.254]; nonmetro 0.2680 SE 0.0013 [0.266, 0.270]; difference 0.0189 [0.013, 0.025] vs paper 0.019 [0.013, 0.024]; NE 0.2595, MW 0.2690, S 0.2616, W 0.2124 with county counts exactly 217/1054/1420/446. The child-weighted mean of county slopes nationally is 0.252 against the national slope 0.287, which shows that the within-county means omit the between-county component, as the paper acknowledges (Sec. 4.5). The US side of H3 therefore reproduces.
script: 20_us_county_reproduction.pyAtlas county_outcomes.csv (1978-83): county slope = (kir_p75 - kir_p25)/0.5; RUCC 2013 metro 1-3; state FIPS to Census regions; child-weighted means (kir_pooled_pooled_n) with county bootstrap B = 2000
References
33 references checked
- [23] Fan, Yi, Zhang. Rising intergenerational income persistence in China. IZA Discussion Paper 12804, 2019
Exists and is the source of all China income values (check 02). It is cited for something it does not quite support. Sec. 4.2 (paper.txt L239) describes child income as 'imputed for family members who do not report it', but Fan et al. Sec. 3.3 compute a fitted (Heckman-corrected) income for the FULL sample, with prediction R-squared 0.235 for children and 0.250 for parents (check 15b). Table 8 row R2 labels Fan's Table 2 Panel D (full-sample prediction, 0.287) as 'trimmed incomes'; the word 'trim' does not occur in Fan et al. R1 omits Fan's mother-age <= 62 restriction (check 02b).
- [24] Fan, Yi, Zhang. American Economic Journal: Economic Policy 13(1):202-230, 2021 (doi 10.1257/pol.20170097)
Exists (Crossref match). It is cited for the province IGEs behind the China belt and NBS-region rows (Table 6 L664-669, Table 8 L783-786, Sec. 4.5). Those values reproduce exactly from Table D1 of the IZA DP [23] online appendix (check 08); we did not examine whether the AEJ version prints the same table. That table lists 23 provinces, and the paper's own counts sum to 23, while the Table 6 note and Sec. 7.3 say '22 provinces'.
- [25] Gong, Leigh, Meng. Intergenerational income mobility in urban China. IZA DP 4811, 2010
Exists and the Table 8 values (0.241, 0.603) match (check 17). Sec. 2.2 (L105) cites it for 'father–son IGEs of 0.18 to 0.32 with single-year father income'. The 0.18 (0.183) end of that range is a father–daughter estimate; the father–son single-year range in the source is 0.241–0.321.
- [13] Mazumder. Estimating the intergenerational elasticity and rank association in the US. FRB Chicago WP 2015-04
Exists (SSRN DOI 10.2139/ssrn.2620727) and the values 0.414 and 0.680 match (check 17). Sec. 2.1 (L88), Sec. 6.8 (L748) and Table 8 (L775-776) describe the averaging as over 'father earnings'/'father income'; the source table is the father–son IGE in parental family income. The 1-year and 15-year estimates also come from different samples (N = 1358 vs 533), so the 'raises from 0.41 to 0.68' contrast is not a within-sample comparison.
- [3] Xie, Hu. An introduction to the China Family Panel Studies (CFPS). Chinese Sociological Review 47(1):3-29, 2014
The DOI given, 10.2753/CSA2162-0555470101, does not resolve: Crossref and doi.org both return 404, and Crossref title/author search finds no Chinese Sociological Review record with that title. The citation is only used to describe the CFPS (L28, L211), so this is a bibliographic defect, not a support problem; the DOI should be corrected.
- [27] Hertz et al. The inheritance of educational inequality. B.E. Journal of Economic Analysis & Policy 7(2), 2007
Exists (Crossref match, doi 10.2202/1935-1682.1775; journal issue dated 2008). The claim at L109–110 (rural China parent–child schooling correlation 0.20, 'among the lowest of 42 countries') matches the ranking table as reported in a web-search summary of the paper and the authors' VoxEU column (rural China correlation 0.20, rank 41 of 42). This was not verified by a script, so it is spot-check only and no issue is asserted beyond the year (2007 vs 2008).
Suggestions
Describe the China income measure as a model-predicted lifetime income for all respondents (Fan et al. Sec. 3.3), regrade the H1 pair, and attach the caveat to the abstract's headline number.
Add a sensitivity analysis showing how a prediction-based income construction changes rank slopes, for example by applying Fan's procedure to PSID or GSS income and comparing it with observed income.
Deposit the plan, deviation log, transcribed inputs and code with a persistent identifier and an externally verifiable timestamp, so that the hash and the claim that the plan preceded the analysis can be checked.
Qualify the title and Sec. 3.2: state that all inputs were public before registration, and list which deviations set decision or inference rules after the freeze.
Revise the abstract: '1950s–1960s' (or 'for COR, all pre-1970 cohorts') instead of 'before 1970'; give the farm-exit comparison its grade-C hedge; say 'means of within-county slopes' for the US metro/nonmetro contrast.
Report H4 as undetermined at published precision rather than 'consistent', and note that the headline IGE pair mixes individual and family income.
Correct the labels: Table 8 R2 (Panel D full-sample prediction, not trimmed incomes), the R1 mother-age restriction, '(R3)' in Table 3, and '22 provinces'. Also correct the citation details for Gong et al., Mazumder and the DOI of [3].
Extend the Table 2 tie-break caveat to US BHQ4, and report the GSS respondents without region-at-16.
Ethics and integrity
The paper lists 'IterSCI' as its sole author, and the requester states that it was produced end to end by an AI research system. Readers and venues should know that no human author is named who is accountable for the analysis or for the integrity of the pre-registration. I found no text in the paper or the requester notes that tried to instruct the reviewer. I found no sign of fabricated values: every published number I traced matches its source exactly. Data use is appropriate. CFPS microdata are restricted, and the paper correctly states that it used none.
Scope of this review
Only the PDF was supplied. No LaTeX source, code, replication package, plan file or deviation log was available, so the pre-registration hash and timestamps and all provenance-'M' estimates (GSS occupational, SEI/prestige and schooling coefficients) could not be re-estimated. The US IGEs 0.348 and 0.452 could not be reproduced because the public Atlas percentile file has no dollar means. China values were checked against the IZA DP 12804 version of Fan et al. and its online appendix, not the AEJ: Policy article. Emran & Sun were checked against the World Bank working paper. The Hertz et al. ranking claim was spot-checked against secondary summaries only.