pith. sign in

arxiv: 2606.30816 · v1 · pith:ZNJ7POX7new · submitted 2026-06-29 · 📊 stat.AP

Spatial Dependence in the Self-Response: Spatial Dependence, Modeling, and Operational Consequences

Emanuel Ben-David This is my paper

Pith reviewed 2026-07-01 01:29 UTC · model grok-4.3

classification 📊 stat.AP
keywords spatial dependencecensus non-responselow response scorespatial error modelspatial Durbin modelqueen contiguityspatial cross-validation
0
0 comments X

The pith

Error-type spatial models capture residual dependence in census non-response better than lag processes when tested out of sample.

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

The Census Bureau's Low Response Score predicts mail non-return rates from tract covariates using ordinary least squares. After this fit, substantial spatial autocorrelation remains in the residuals. Model comparisons across the spatial autoregressive family show that the dependence is mainly in the errors rather than a global endogenous lag. Spatial Durbin models achieve the lowest in-sample AIC, but spatial-block cross-validation reverses the ranking and favors the spatial error models. This indicates that local neighborhood spillovers can be included without assuming a contagion mechanism across all tracts.

Core claim

With 71,076 tracts and the twenty-five Erdman-Bates predictors under queen-contiguity weights, OLS produces a Moran's I of 0.399. Formal diagnostics attribute the leftover dependence to error processes. The spatial Durbin model minimizes AIC, yet spatial error models (SEM and SDEM) achieve the best performance under spatial-block validation, while the SDM performs worst out of sample. The SDEM supplies an interpretable middle ground by treating neighborhood effects as local spillovers.

What carries the argument

Comparison of the spatial autoregressive model family (OLS, SAR, SEM, SDM, SDEM) under queen-contiguity weights, ranked by in-sample AIC and by random versus spatial-block cross-validation.

If this is right

  • Low Response Score models should incorporate spatial error terms rather than lag terms to improve generalization.
  • Local neighborhood demographic effects can be represented as spillovers without requiring a global response-contagion process.
  • Official planning instruments of this type should be assessed with spatial-block validation in addition to in-sample fit measures.
  • The SDEM specification supplies both statistical performance and direct interpretation of neighboring tract influences.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar spatial error specifications may improve other geographically clustered survey planning scores that currently rely on plain OLS.
  • If tract boundaries or weights definitions change, the relative performance of error versus lag models could shift and should be rechecked.
  • The finding that in-sample AIC and spatial validation disagree suggests that many spatial applications would benefit from explicit out-of-sample spatial testing.

Load-bearing premise

Queen-contiguity weights defined on 2010 tract boundaries capture the relevant spatial dependence and the chosen predictors plus spatial terms account for all systematic variation in non-response rates.

What would settle it

Finding that the spatial Durbin model outperforms the spatial error models under repeated spatial-block validation on held-out tracts would falsify the claim that error-type dependence is primary.

Figures

Figures reproduced from arXiv: 2606.30816 by Emanuel Ben-David.

Figure 1
Figure 1. Figure 1: Observed 2010 mail non-return rate by census tract (the modeling outcome, shown on the rate scale). Darker tracts return their forms at high rates; brighter tracts are harder to count. The rate clusters into coherent regions—the rural South, the southern border, the desert Southwest, and urban cores—rather than varying tract by tract. This spatial clustering, left unmodeled by an own-tract OLS specificatio… view at source ↗
Figure 2
Figure 2. Figure 2: The specification search in one plane: in-sample fit (∆AIC, log scale; lower is better) versus residual spatial autocorrelation (Moran’s I; zero is fully absorbed). The desirable region is the lower-left. OLS and SLX fit poorly and leave large residual autocorrelation; the pure lag model (SLM) fits better but still leaves I = 0.05; the error and Durbin models (SEM, SDM, SDEM, SARAR) cluster in the lower-le… view at source ↗
Figure 3
Figure 3. Figure 3: Direct (own-tract) and indirect (neighbor) impacts of each predictor on logit(non-return), per one standard deviation, for SDEM (blue, with 95% intervals) and SDM (red). Predictors are ordered by SDEM total effect. The direct effects coincide across the two models; the indirect effects are systematically larger under SDM because its endogenous lag propagates each covariate through the global multiplier (I … view at source ↗
Figure 4
Figure 4. Figure 4: The spatial model’s adjustment to the non-spatial prediction (SDEM minus OLS predicted non￾return rate, percentage points). Green: SDEM predicts a higher non-return rate than OLS because the neighborhood is harder to count than own-tract covariates imply; purple: the reverse. The adjustments form coherent regions—raised across Texas and the south-central states, lowered across the upper Midwest and Northea… view at source ↗
Figure 5
Figure 5. Figure 5: SDEM fit across U.S. tracts: (a) observed non-return rate, (b) SDEM predicted rate (same color scale), and (c) residual (observed minus predicted, percentage points; red = under-prediction, blue = over￾prediction). Predicted closely reproduces observed. The residual map shows small, spatially unstructured errors—the salt-and-pepper texture corresponding to the near-zero residual Moran’s I—confirming that t… view at source ↗
Figure 6
Figure 6. Figure 6: SDEM fit for the District of Columbia: (a) observed, (b) predicted, (c) residual (percentage points). DC’s east–west non-response gradient—low in the northwest, high east of the Anacostia—is repro￾duced by the model; residuals are modest and show no remaining systematic gradient. Finally, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The SDEM neighbor-demographic contribution W X ˆθ (log-odds) for the District of Columbia: the portion of each tract’s predicted non-return that derives from its neighbors’ characteristics. The contribution is positive and largest across the eastern and southeastern wards and near zero in the central northwest, showing that the spillover is spatially structured and reinforces the city’s east–west non-respo… view at source ↗
Figure 8
Figure 8. Figure 8: Spatial-block cross-validation folds: each color is one of five contiguous regions, withheld in its entirety in turn. Because a held-out tract’s neighbors are withheld with it, no local information leaks from training to test —a conservative estimate of skill on unsampled geography, in contrast to random folds, where a held-out tract is surrounded by training tracts. 11 [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 9
Figure 9. Figure 9: Out-of-sample RMSE (percentage points) for each model under random (interspersed) and spatial-block (region held out) cross-validation; the line length is the neighbor-reliance gap. OLS has a negligible gap (no spatial term to lose). The spatial models predict much better under random folds than under spatial blocks, a gap near 1 pp, indicating that much of their apparent skill is local interpolation rathe… view at source ↗
Figure 10
Figure 10. Figure 10: Residual variance by tract-size quintile (median occupied housing units), for OLS and SDEM residuals, with the delta-method prediction 1/[n p(1 − p)]. Variance falls about twofold from the smallest to the largest tracts, tracking the delta-method curve. SDEM lies below OLS (lower overall variance) but parallel to it (the same size gradient), showing that the spatial error model does not remove the heteros… view at source ↗
Figure 11
Figure 11. Figure 11: Prediction reliability: the absolute SDEM residual (percentage points) by tract; brighter tracts have larger errors. Error concentrates in the rural South and border, the sparsely populated interior West (large-area, small- population tracts), and scattered urban pockets, while the densely settled East and Midwest are accurate. The clustered error variance (Moran’s I = 0.10) is the spatial footprint of th… view at source ↗
read the original abstract

The U.S.\ Census Bureau's Low Response Score (LRS) is a central planning instrument for identifying places likely to require additional self-response outreach and nonresponse follow-up. The published LRS is intentionally interpretable: it is built from tract-level covariates using an ordinary least squares specification. That transparency, however, leaves open an important question for official statistics: how much spatial structure remains after the own-tract covariates have done their work, and what form does that structure take? Using the observed 2010 Census mail non-return rate for 71,076 U.S. census tracts and the twenty-five Erdman--Bates LRS predictors, this paper compares the full spatial autoregressive model family under queen-contiguity weights and validates the leading candidates with both random and spatial-block cross-validation. OLS leaves strong residual spatial autocorrelation ($I=0.399$). Formal diagnostics and model comparisons indicate that the remaining dependence is primarily error-type rather than a global endogenous lag process. Although the spatial Durbin model minimizes in-sample AIC, spatial-block validation reverses that ranking: the error-family models (SEM/SDEM) generalize best, while the AIC-best SDM is weakest out of sample. The SDEM provides an interpretable middle ground, absorbing residual spatial dependence while representing neighborhood demographic effects as local spillovers. Robustness checks show that these conclusions are invariant to the weights definition and are not an artifact of tract-size-driven heteroskedasticity. The results suggest that LRS-style response models should be evaluated with spatial validation, not only in-sample fit, and that local neighborhood context can be operationally meaningful without invoking a global response-contagion mechanism.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper analyzes residual spatial dependence in 2010 U.S. Census tract mail non-response rates (71,076 tracts) after OLS regression on the 25 Erdman-Bates LRS predictors. It reports strong residual Moran's I (0.399), uses LM diagnostics and the full spatial autoregressive family (SAR, SEM, SDM, SDEM, etc.) under queen-contiguity weights, and finds that spatial-block cross-validation reverses the in-sample AIC ranking: error-family models (SEM/SDEM) generalize best while the AIC-best SDM performs worst out-of-sample. The SDEM is presented as an interpretable compromise capturing local spillovers without global lag contagion. Robustness to alternative weights and heteroskedasticity is claimed.

Significance. If the central claim holds, the work is significant for official statistics practice: it shows that reliance on in-sample AIC alone can select models that fail spatial generalization, and that neighborhood demographic spillovers can be modeled operationally via error or Durbin-error specifications without invoking a global endogenous response process. The explicit use of spatial-block CV and the contrast between in-sample and out-of-sample rankings provide a concrete, falsifiable demonstration that is directly relevant to census planning tools such as the Low Response Score.

major comments (2)
  1. [Abstract] Abstract (robustness paragraph): the claim that 'these conclusions are invariant to the weights definition' is load-bearing for the error-type vs. lag-type classification, yet the abstract provides no enumeration of the alternative weight matrices examined nor any table or figure showing how the LM test statistics or the SEM/SDEM vs. SDM ranking change under those alternatives. Without this, it is impossible to assess whether queen contiguity is merely one convenient choice or whether the lag-vs-error distinction is robust to plausible changes in neighborhood definition.
  2. [Abstract] Abstract (model comparison and CV paragraph): the reversal of ranking under spatial-block CV is the key empirical result supporting the error-type conclusion, but the abstract does not report the number of blocks, the block-construction rule, or the precise out-of-sample metric (e.g., RMSE, MAE, or log-score) used to declare SEM/SDEM superior. These details are necessary to evaluate whether the spatial-block design truly isolates the dependence structure or inadvertently favors error models by construction.
minor comments (1)
  1. [Abstract] The abstract states OLS leaves I=0.399 but does not indicate whether this is the raw or residual Moran's I, nor the exact p-value or permutation test used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these focused comments on the abstract. Both points correctly identify places where additional detail would strengthen the presentation of our robustness and validation results. We address each below and will revise the abstract accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract (robustness paragraph): the claim that 'these conclusions are invariant to the weights definition' is load-bearing for the error-type vs. lag-type classification, yet the abstract provides no enumeration of the alternative weight matrices examined nor any table or figure showing how the LM test statistics or the SEM/SDEM vs. SDM ranking change under those alternatives. Without this, it is impossible to assess whether queen contiguity is merely one convenient choice or whether the lag-vs-error distinction is robust to plausible changes in neighborhood definition.

    Authors: We agree the abstract should be more explicit. The manuscript already reports results under queen contiguity as the primary specification and states that conclusions are invariant under alternatives. To make this transparent in the abstract we will add a parenthetical enumeration of the matrices tested (rook contiguity, k=4 and k=8 nearest-neighbor, and 5 km / 10 km distance-band weights) and note that both the LM diagnostics and the out-of-sample ranking of error-family versus lag-family models remain unchanged. A concise supplementary table summarizing the key LM statistics and CV rankings across weights will also be referenced. revision: yes

  2. Referee: [Abstract] Abstract (model comparison and CV paragraph): the reversal of ranking under spatial-block CV is the key empirical result supporting the error-type conclusion, but the abstract does not report the number of blocks, the block-construction rule, or the precise out-of-sample metric (e.g., RMSE, MAE, or log-score) used to declare SEM/SDEM superior. These details are necessary to evaluate whether the spatial-block design truly isolates the dependence structure or inadvertently favors error models by construction.

    Authors: The referee is correct that these operational details belong in the abstract. The spatial-block CV used 100 blocks formed by k-means clustering on tract centroids (ensuring no shared borders between training and test blocks), with RMSE as the out-of-sample metric. We will revise the abstract sentence to read: 'spatial-block cross-validation (100 k-means centroid blocks, RMSE) reverses that ranking: the error-family models (SEM/SDEM) generalize best...' This addition directly addresses the concern without lengthening the abstract beyond its limit. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on held-out validation

full rationale

The paper's core result—that residual dependence after the 25 Erdman-Bates covariates is primarily error-type (SEM/SDEM) rather than lag-type (SDM/SAR), and that spatial-block CV reverses the AIC ranking—is obtained by fitting standard spatial autoregressive models to the 2010 tract data under queen-contiguity weights and then evaluating predictive performance on spatially blocked hold-out sets. No equation reduces a fitted quantity to itself by construction, no parameter is estimated on a subset and then relabeled a prediction, and no uniqueness theorem or prior self-citation is invoked to force the model family. The derivation therefore remains self-contained against external benchmarks (the observed non-response rates and the explicit cross-validation partitions).

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review; free parameters are the spatial coefficients fitted by maximum likelihood; main domain assumption is that queen contiguity captures neighborhood structure; no invented entities mentioned.

free parameters (1)
  • spatial autoregressive coefficient(s)
    Fitted in SEM, SAR, SDM, SDEM specifications to capture residual dependence.
axioms (1)
  • domain assumption Queen-contiguity weights matrix correctly encodes spatial neighbors for census tracts
    Used for all reported models and robustness checks.

pith-pipeline@v0.9.1-grok · 5827 in / 1227 out tokens · 38328 ms · 2026-07-01T01:29:46.812210+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

10 extracted references

  1. [1]

    K., Florax, R., and Yoon, M

    Anselin, L., Bera, A. K., Florax, R., and Yoon, M. J. (1996). Simple diagnostic tests for spatial dependence. Regional Science and Urban Economics, 26(1), 77--104

    1996

  2. [2]

    Burridge, P. (1981). Testing for a common factor in a spatial autoregression model. Environment and Planning A, 13(7), 795--800

    1981

  3. [3]

    Census Bureau

    U.S. Census Bureau. (2026). Updating the Planning Database with the latest socioeconomic, demographic and housing data. Random Samplings Blog. https://www.census.gov/newsroom/blogs/random-samplings/2026/03/updating-the-planning-database.html

    2026

  4. [4]

    Elhorst, J. P. (2010). Applied spatial econometrics: raising the bar. Spatial Economic Analysis, 5(1), 9--28

    2010

  5. [5]

    Erdman, C., and Bates, N. (2017). The Low Response Score (LRS): A metric to locate, predict, and manage hard-to-survey populations. Public Opinion Quarterly, 81(1), 144--156

    2017

  6. [6]

    Goulard, M., Laurent, T., and Thomas-Agnan, C. (2017). About predictions in spatial autoregressive models: optimal and almost optimal strategies. Spatial Economic Analysis, 12(2--3), 304--325

    2017

  7. [7]

    H., and Prucha, I

    Kelejian, H. H., and Prucha, I. R. (2010). Specification and estimation of spatial autoregressive models with autoregressive and heteroskedastic disturbances. Journal of Econometrics, 157(1), 53--67

    2010

  8. [8]

    LeSage, J., and Pace, R. K. (2009). Introduction to Spatial Econometrics. Chapman and Hall/CRC, Boca Raton

    2009

  9. [9]

    K., and LeSage, J

    Pace, R. K., and LeSage, J. P. (2008). A spatial Hausman test. Economics Letters, 101(3), 282--284

    2008

  10. [10]

    R., et al

    Roberts, D. R., et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, 40(8), 913--929

    2017