A replication and review of Siwan Anderson’s “Caste as an Impediment to Trade” (AEJ: Applied Economics, 2011). Groundwater, caste, and a second look at the evidence.
The regressions reproduce a positive crop-sales association. Its size and causal explanation remain uncertain. All 54 displayed coefficient/standard-error pairs checked agree within 0.1 rupee. The estimated gap for lower-caste water buyers without pumps is ₹763 more crop sales per owned acre per year in lower-caste-dominated villages. The published inference method gives a 95% interval of ₹146–₹1,379. A correction for limited and uneven village information widens it to ₹55–₹1,471. That supports a positive association, with substantial uncertainty about its magnitude.
The outcome is money from crops sold per acre owned. Physical yields, profits and losses caused by caste barriers require additional evidence. The buyer interaction survives removing any one village, but the data do not establish that the village gap is larger for buyers than for pump owners.
The points are the village-dominance interactions for water buyers and pump owners in Table 4(2), and their difference. All use the same 1,295 households and village-clustered 95% t intervals. A significant buyer interaction and an insignificant owner interaction do not establish a difference between them: the direct test has p = 0.454.
The argument and all twelve review checks are in the review. The village balance and historical-data audit shows what the existing comparisons rule out, which 1991 SHRUG variables are available, and why the missing survey-to-Census crosswalk prevents historical checks.
| Check | Result | Evidence |
|---|---|---|
| Baseline village-dominance coefficient | ₹566.5 (SE 209.0) | Reproduction |
| Full village gap for nonowner water buyers, corrected interval | ₹763 [₹55, ₹1,471] | Precision audit |
| Water-buyer interaction | ₹850.9 (SE 275.0) | OLS tables |
| Buyer minus owner interactions | ₹462.4 (SE 614.2), p = 0.454 | Direct contrasts |
| District-adjusted village coefficient, wild bootstrap | p = 0.062 | Wild bootstrap |
| IV interaction, published inference | ₹3,519.5 (SE 1,413.7) | IV table |
| IV interaction, village pairs bootstrap | 95% percentile interval [₹-395, ₹8,982] | Full-procedure bootstrap |
| Three published sample sizes | 1,295 printed; 1,127, 1,122, and 1,127 used | Sample comparison |
What does the 45% claim establish? Anderson describes lower-caste water buyers as having 45% higher agricultural yields in lower-caste-dominated villages (p. 253). The supplied program reproduces rupee coefficients but does not show the calculation behind that percentage. The paper's 45% statement has no accompanying percentage confidence interval. We therefore report the reconstructed rupee contrast and its uncertainty. The supplied files do not establish how precisely the 45% figure is estimated.
The published regression standard errors already account for village clustering and reproduce under the stated method. The buyer interaction alone is ₹851 (SE ₹275); the actual village gap for nonowner buyers also includes the village coefficient. That full contrast is ₹763 (SE ₹310), using both variances and their covariance. The larger uncertainty belongs alongside the substantive buyer comparison.
The correction for limited village information (CR2 with Satterthwaite inference) gives about 33 degrees of freedom for Anderson's buyer gap. The corresponding IHDS calculation has 6.8, because its relevant comparison cells are much sparser. These are contrast-specific calibrations of uncertainty, not counts of villages. Anderson has nonowner buyers in 44 upper-dominated and 42 lower-dominated villages. The original sample therefore has broader buyer support than IHDS.
A separate 9,999-draw wild village bootstrap gives a full buyer-gap interval of ₹106–₹1,497. The ordinary and small-sample-corrected intervals and both bootstrap variants support a positive association; none makes its size precise. These procedures quantify sampling uncertainty within the model. They do not resolve classification errors, unmeasured village differences or the water-trading mechanism. All precision checks, bootstrap results.
Higher sales per owned acre can reflect more land cultivated through leasing, repeated crop seasons, crop choice, prices, or selling more of the harvest. It need not mean greater physical output for the same crop on the same acreage. These are reproduction and sensitivity results from the original sample; the separate IHDS check is linked below. The magnitude comparison checks the claim against Kerala fishing markets, groundwater contracting, and canal irrigation, keeping the different outcomes and comparison groups explicit. It also audits what we know about the dominance cutoff.
Attributing the gap to caste barriers requires another step: establishing that other differences between the villages do not explain it. The buyer–seller pairs, water prices, and delivery records needed to investigate the proposed trading mechanism are absent from this dataset.
The economic question is why profitable alternatives do not emerge. Nearby sellers, new wells, or shared irrigation could compete away some of the gap. Distance, installation costs, and unreliable contracts could prevent that, but the data do not show which obstacle binds or whether caste causes it. The economics section of the review works through these possibilities and the evidence needed to distinguish them.
The baseline uses 1,295 households in 90 villages: 591 households in 48 high-caste-dominated villages and 704 in 42 lower-caste-dominated villages. Village samples range from 1 to 32 households, with a median of 13. Equal total weight per village gives a baseline difference of ₹586 per owned acre, compared with ₹567 under the original household weighting. This is a sensitivity check; it does not recover the original survey weights.
Following the checks discussed by Lal, Lockhart, Xu, and Zu, I compared IV and OLS on the same 1,127 households in 80 villages and bootstrapped the complete estimation procedure. The IV interaction is 3.85 times the OLS interaction, but their difference is imprecisely estimated. The IV percentile interval is −₹395 to ₹8,982; its studentized interval is ₹1,529 to ₹10,447. Both are wide, and they disagree about whether zero is excluded. The percentile interval is not a weak-instrument-robust confidence set. See the IV results and explanation.
Run from the repository root:
make deps
make run
make test
make lintmake deps restores renv.lock into .R/library; the other targets use that library when
present. R 4.6.0 was used for the recorded run. make run reproduces all six tables, performs the
robustness checks, and regenerates the figure and this README from the results. It downloads the final paper if needed, requires Poppler’s pdftotext, and takes several minutes. The prose review is a written interpretation of
the recorded analysis. REPRODUCING.md documents inference conventions, seeds,
dependencies, and the tests.
| Directory | Contents |
|---|---|
data/original/ |
Unmodified Stata files, original commands, readme, and license |
src/ |
R reproduction, sensitivity analyses, published comparisons, and figure |
tests/ |
Independent matrix-algebra checks, Stata benchmark, and data invariants |
output/ |
Machine-readable estimates, intervals, sample counts, and bootstrap draws |
ms/ |
Review and the twelve-check coverage matrix |
sources/ |
Provenance and locally cached paper and independent Stata log |
The original AEA replication archive is distributed under the terms in its license. The author’s final paper and independent 2022 Stata reproduction provide the comparison sources. See sources/README.md for cached-file provenance.
Research and writing opportunities arising from the review.
Village-classification sensitivity checks every single-village reassignment.
Completed IHDS replication: the main later-survey crop-value comparison is negative, with wide uncertainty; changing owned acreage to cultivated crop-season acreage changes the direction. It neither reproduces a stable positive gap nor decisively rules out the 45% benchmark. Run make ihds-replication with the local IHDS archive. Additional data and measurement limits cover NSS and Bihar land records.
Agy audit and our adjudication.