Gloudemans: Can CAMA Estimate Land Value Reliably Enough for LVT? (2000 & 2002 Lincoln Institute Working Papers)

Overview

Robert J. Gloudemans — the mass-appraisal consultant and IAAO veteran who later co-authored Fundamentals of Mass Appraisal (2011) — wrote two Lincoln Institute of Land Policy working papers directly testing the assessability premise behind a land value tax: can computer-assisted mass appraisal (CAMA) reliably decompose a property's sale price into separate land and building values, including in the built-up neighborhoods where almost nothing sells unimproved? The first, Implementing a Land Value Tax in Urban Residential Communities (2000, Lincoln Institute Product Code WP00RG1), was the primary empirical study; the second, An Empirical Analysis of the Incidence of Location on Land and Building Values (2002, WP02RG1), reruns and extends the same three datasets to test a narrower, related question — how location affects land versus building value — and along the way restates and leans on the 2000 paper's central conclusion.[1][2] Both were fetched and read in full for this page (not summarized secondhand); this page is the wiki's dedicated, systematically-mined treatment of both.

The Research Question

Standard mass-appraisal models (multiple regression analysis, MRA) estimate a property's total sale price from location and structure variables without separating land from building — the "constant" term and location coefficients that a linear model estimates cannot be cleanly assigned to one component or the other.[2] Gloudemans's papers ask whether a different model structure — a nonlinear, multiplicative "feedback" model that is mathematically decomposable into a land-value term and a building-value term — can estimate the land component with acceptable reliability, and specifically whether this remains true in the many urban neighborhoods that have few or no vacant-land sales to calibrate against directly. As the 2000 paper's introduction states: "In many urban areas most land has been improved and there are no or relatively few vacant land sales to help develop land values. From the perspective of a land value tax, this reality creates practical difficulties in determining land values, particularly for built-on land."[1]

Data and Method

Both papers draw on the same three North American residential sales databases, each supplied directly by the local assessment jurisdiction:[1][2]

• Ada County (Boise), Idaho — 12,821–15,005 usable sales (the two papers report slightly different counts for the same jurisdiction; see Limits below), 1997–1999.
• Jefferson County (suburban Denver), Colorado — 4,836 usable sales, 1996–1998.
• Clareview market area, Edmonton, Alberta — 4,382 usable sales, July 1996–June 1999 — one of twelve residential market areas the City of Edmonton uses for modeling, chosen because it has comparatively many vacant-lot sales even though roughly half its neighborhoods are fully built out with no remaining vacant-lot sales.[1]

Vacant-lot sales were a minority share of each database: 20.5% in Clareview, 14.1–14.6% in Ada County, and just 4.5% in Jefferson County — Jefferson County is the closest analogue to a fully built-up city where vacant sales are scarce.[1][2]

The method: for each jurisdiction, Gloudemans first calibrated a traditional additive MRA model using improved sales only (the field's normal practice, and this paper's benchmark), then calibrated a nonlinear "feedback" model combining vacant and improved sales together, of the general form V = πGQ × ((land terms) + (building terms)) — multiplicative qualitative factors (neighborhood, site amenities) applied to additive land and building components, decomposable algebraically into a land value and a building value for every parcel.[1][2] The 2002 paper additionally tested four variant specifications of how location effects should be split between land and buildings (ranging from "location affects land only" to "location affects land and buildings at a market-calibrated ratio"), to see which assumption the data actually supported.[2]

Findings: The Improved-Only Baseline

Before adding any vacant-land sales, the traditional additive models (the field's status quo) performed well by the profession's own accuracy metric, the coefficient of dispersion (COD) — a lower COD means tighter, more uniform assessment accuracy.[1] In the 2000 paper: Clareview's additive model produced a median ratio of 1.0006 and a COD of 5.87 (adjusted R² = .874); the comparable improved-only benchmarks were COD 5.75 in Jefferson County and COD 8.82 in Ada County, where fewer location variables were available.[1] The 2002 paper's five improved-only model variants for all three jurisdictions produced nearly identical results regardless of specification — adjusted R² of .959 (Jefferson County), .882 (Clareview), and .909 (Ada County), with CODs clustering tightly (5.39–5.52 Jefferson, 5.80–5.82 Clareview, 8.64–8.71 Ada) — leading Gloudemans to note "the amazing similarity in model performance measures across all five models" when only improved sales are used.[2]

But the 2002 paper's central methodological warning is that this apparent stability is deceptive for the land/building split specifically, not the total-value estimate: because none of the improved-only model forms include a constant term, the fixed-cost portion of building value (site preparation, developer's profit, the value of "a residence in place") gets mechanically absorbed into the land-value estimate. Gloudemans reports implausibly high land shares as a result — 47–55% of total value in Jefferson County, ~50% in Clareview, versus a more textbook-typical 18–22% in Ada County (the jurisdiction with the fewest location variables and, not coincidentally, the least of this distortion) — and concludes: "The seemingly high land values obtained in two of the areas and highly different, more traditional results in the third call into question the reliability of the land and improvement values developed by feedback... Feedback models may purport to break out land and building values, but the allocations are not necessarily realistic."[2] The 2000 paper reaches the same conclusion independently: "Models that incorporate only improved sales are unlikely to be decomposable into reliable building and land values."[1]

Findings: Adding Vacant-Land Sales

Once vacant sales were folded into the same combined model — as benchmarks that anchor the land-value estimate to actual observed vacant-land transactions — both accuracy and plausibility improved. Per-jurisdiction, comparing the improved-only benchmark to the combined model's improved-sales accuracy (2000 paper, cross-jurisdiction table):[1]

The improved-sales accuracy barely moved from the improved-only benchmark — Clareview's COD rose only from 5.87 to 5.97, and Jefferson County's actually fell (improved) from 5.75 to 5.55 — supporting the paper's headline claim that "a combined model built from both vacant and improved sales need not sacrifice predictive accuracy for improved properties."[1] Land's resulting share of total value, once vacant sales anchored the split, came out lower and more consistent than the improved-only estimates: 29% in Ada County, 35% in Clareview, 43% in Jefferson County — versus the inflated 47–55%/~50%/18–22% figures the improved-only models produced.[1][2]

The 2002 paper's more detailed five-model comparison (Exhibit 2) found that model choice matters more once vacant sales are added, and that the best-fitting specification — one where the model itself calibrates what percentage of location's effect on land also carries over to buildings, rather than assuming a fixed 0%, 50%, or 100% split — consistently produced the lowest (best) vacant-land CODs: 11.61 in Jefferson County, 9.55 in Clareview, and 17.73 in Ada County, each better than the traditional feedback model's 14.60, (Clareview not directly comparable in this table), and 22.96 respectively.[2] Under that best model, adjustments to building value from location factors ran only 12% to 21% the size of the corresponding land adjustments — "versus closer to one-half in the models with improved sales only" — which is the paper's core empirical claim: location's dollar and percentage impact falls overwhelmingly on land, not buildings, and improved-only models understate that asymmetry.[2]

The "Neighborhoods Without Vacant Sales" Result

This is the finding both the 2000 predecessor and the wiki's Fundamentals of Mass Appraisal page already lean on, and it is the paper's own stated bottom line, quoted directly from the 2002 conclusions (numbered point 5, its final conclusion):

"Modern mass appraisal methods are capable of producing reasonable estimates of the value of land as if vacant even in neighborhoods with no or few vacant land sales, provided there are other neighborhoods in the model with adequate vacant land sales to provide reality checks."[2]

The mechanism is straightforward and important to state precisely: the combined model is calibrated jurisdiction-wide, using every vacant sale available across all neighborhoods in the dataset (232 in Jefferson County, 900 in Clareview, 2,184 in Ada County) to fix the model's land-value parameters (base lot values, location multipliers, the size-adjustment exponent). Those jurisdiction-level parameters then apply to every neighborhood in the model, including the ones that individually contributed zero vacant sales — Clareview itself is a case in point: "about half of neighborhoods in the area are virtually fully developed with no remaining vacant lot sales," yet the combined model still produced land estimates for all of them.[1] The claim is not that any single land-scarce neighborhood is self-sufficient for calibration; it is that enough vacant sales somewhere in the broader model discipline the land-value function that then extrapolates to land-scarce neighborhoods. The 2000 paper's own summary states this caveat explicitly: "a combined model affords the possibility of providing market data in built-up areas with few vacant lots, while still using those that are available to ensure that models will, on average, neither under- nor over-estimate vacant land values."[1]

Distributional Findings: Who Gains and Loses

The 2000 paper goes further than the 2002 follow-up by running an explicit tax-shift analysis — projecting who would pay more or less under a hypothetical land-only tax, holding total revenue constant. In Clareview, land value was estimated at $181.5 million of $513.5 million total sales value (35%), meaning a land-only tax base would need roughly a 286% rate increase (1/.35) to raise the same revenue as an all-property base.[1] Owners with land-to-total ratios above that 35% average would see higher bills; those below it, lower bills — and the paper found the land/total ratio was strongly correlated with living area (r = .627), year built (r = .713), building quality (r = .636), and overall property value (R² = .632, i.e. r ≈ .795) — meaning smaller, older, lower-quality, and lower-value homes systematically carry higher land-to-total ratios and would see their relative tax burden rise, while larger, newer, higher-quality, higher-value homes would see it fall.[1] Gloudemans states the conclusion plainly in the abstract: "implementation of a land value tax would be less advantageous to lower-value than higher-value residential properties."[1] Across all three jurisdictions, the estimated land share of total value ranged from 29% (Ada County) to 43% (Jefferson County).[1]

This is a genuine, evidence-based distributional caution the wiki should carry alongside the assessability finding — it is not a critique of whether land can be measured, but of who benefits once it is, under a simple flat-rate land-only design (the paper does not model exemptions, circuit-breakers, or other mitigation the wiki's transition-related pages discuss elsewhere).

Limits and Honest Caveats

• Both papers are single-author, Lincoln Institute-funded working papers, not peer-reviewed journal articles. Both carry the identical disclaimer on their cover page: "The findings and conclusions of this paper are not subject to detailed review and do not necessarily reflect the official views and policies of the Lincoln Institute of Land Policy."[1][2] Neither has been located in a peer-reviewed journal (the 2002 paper's own introductory materials note only that a summary of the 2000 results was separately published in the Journal of Property Tax and Assessment Administration, Spring 2001 — a practitioner trade journal, not evaluated further for this page).[3] The Lincoln Institute's funding is stated plainly here because the Institute has a long-standing institutional interest in land value taxation; that does not make the findings wrong, but it is provenance a reader should have.
• The papers' own fellowship-name citations are inconsistent with each other. The 2000 paper's own abstract states the research was "conducted under a John C. Lincoln Fellowship for the Lincoln Institute of Land Policy."[1] The 2002 paper, citing the same 2000 study, instead calls it "a David C. Lincoln Institute Fellowship in Land Value Taxation."[2] This wiki cannot resolve which name is correct from the primary texts alone (other Lincoln Institute fellowships elsewhere on the wiki are named for David C. Lincoln — see Fundamentals of Mass Appraisal); both direct quotations are reproduced above rather than silently reconciled.
• Residential-only, three jurisdictions, data from the late 1990s. All three databases are single-family residential sales from 1996–1999, calibrated with SPSS nonlinear regression circa 2000 — a quarter-century before this page was written and well before modern gradient-boosted or machine-learning CAMA methods (see Mass Appraisal Methods for the Cook County ML example). The papers say nothing about commercial, industrial, or agricultural land, and nothing about jurisdictions outside these three mid-sized North American markets.
• The two papers report inconsistent sample counts for the same jurisdiction. The 2000 paper's cross-jurisdiction table lists Ada County at 15,005 combined sales, 14.1% vacant; the 2002 paper's footnote lists Ada County at 12,821 sales from 1997–1999, with 2,184 vacant sales (14.6%) reported separately in its later text.[1][2] Neither paper explains the discrepancy; it is presented here as found, not reconciled.
• The headline "no or few vacant sales" finding depends on a jurisdiction-wide calibration, not a neighborhood-by-neighborhood one. As detailed above, the claim is that a broader model with adequate vacant sales somewhere can extrapolate to vacant-sale-poor neighborhoods — not that any individual data-poor area is independently verifiable. A jurisdiction with too few vacant sales anywhere is outside what these papers tested.
• Vacant-land CODs remain substantially higher than improved-property CODs throughout — consistent with, not contradicting, the IAAO's own published tolerance. Every vacant-land COD reported (9.55–22.96 across all models and jurisdictions) is markedly worse than the corresponding improved-sales COD (5.39–8.99). The IAAO's Standard on Ratio Studies, read separately and in full for this wiki, publishes a wider acceptable COD band specifically for vacant land (5.0–25.0) than for improved residential property (5.0–10.0/15.0) — Ada County's worst-case traditional-model vacant COD of 22.96 sits near the top of that IAAO band, while its best-fitting model's 17.73 sits comfortably inside it; all Jefferson County and Clareview vacant CODs are comfortably within the IAAO band under every model tested.[4] The pattern in Gloudemans's data is consistent with the IAAO's institutional experience that vacant-land ratio studies simply disperse more, for reasons plausibly related to thinner samples rather than land being intrinsically harder to value — see IAAO Standards for the fuller discussion of that open question.
• The "traditional feedback model" — closest to standard assessor practice — was not the best-performing specification for vacant land in two of three jurisdictions. Getting the paper's best results required the more sophisticated, market-calibrated Model 5 rather than the simpler assumptions ordinary assessment practice would default to; this is a modeling-sophistication burden the papers do not quantify in cost or staff-time terms.
• No independent replication was located. Searches for later work directly re-testing or replicating these specific findings did not surface a citing literature that revisits the same three databases; the papers stand as a single research program by one author across two related working papers, not a body of corroborating studies.

See Also

• Objection: Land value can't be assessed accurately — the objection page this research bears directly on
• Fundamentals of Mass Appraisal (Gloudemans & Almy, 2011) — the textbook Gloudemans later co-authored; its "The Authors" section first flagged these two working papers
• Mass Appraisal Methods — the concept page on CAMA, hedonic regression, and land/building separation techniques generally
• IAAO Standards: Ratio Studies & Property Tax Policy — the COD/PRD/PRB definitions and the vacant-land tolerance band this page's CODs are measured against
• Kolbe et al.: Berlin land-value appraisal — independent land-valuation accuracy evidence from a different jurisdiction and era
• Lincoln Institute of Land Policy — the funder and publisher of both working papers

Sources

1. Robert J. Gloudemans, Implementing a Land Value Tax in Urban Residential Communities, Lincoln Institute of Land Policy Working Paper WP00RG1 (January 2000), 26 pp. Landing page with PDF (fetched and read in full, 2026-07-18) — used for the research design, the three-database description, the additive-model baseline results (COD 5.87 Clareview / 5.75 Jefferson County / 8.82 Ada County), the combined vacant+improved cross-jurisdiction COD table, the land-share and tax-shift analysis (35%/29%/43% land shares, the 286% rate-increase implication, the r=.627/.713/.636/R²=.632 correlations), the "John C. Lincoln Fellowship" funding attribution, and the "not subject to detailed review" disclaimer.
2. Robert J. Gloudemans, An Empirical Analysis of the Incidence of Location on Land and Building Values, Lincoln Institute of Land Policy Working Paper WP02RG1 (2002), 28 pp. PDF (fetched and read in full, 2026-07-18) — used for the five-model specification comparison, the improved-only distortion finding and its quoted conclusion, the Exhibit 1/Exhibit 2 accuracy tables (adjusted R², COD by model and jurisdiction), the "neighborhoods with no or few vacant land sales" quoted conclusion, the 12–21%-of-land building-impact finding, the "David C. Lincoln Institute Fellowship in Land Value Taxation" funding attribution (footnote 1, citing the 2000 predecessor), and the "not subject to detailed review" disclaimer.
3. Lincoln Institute of Land Policy publications catalog description for WP00RG1, cross-referenced via web search (2026-07-18) — used only for the incidental detail that a summary of the 2000 paper's results appeared in the Journal of Property Tax and Assessment Administration, Spring 2001; that journal article itself was not located or read for this page.
4. International Association of Assessing Officers, Standard on Ratio Studies (approved April 2013) — wiki summary, read in full — used for the vacant-land vs. improved-property acceptable COD ranges (5.0–25.0 vs. 5.0–10.0/15.0) cited in Limits and Honest Caveats.