Do County Characteristics Explain Home Values?
01 / OVERVIEW OF THE STUDY

Do County Characteristics
Explain Home Values?

Real estate is often boiled down to location, location, and location. What factors of a location are important to home values?

Inside the study

Hover, focus or tap a shape to inspect it.
02 / SELECTED MODEL & RESULTS

The selected Extra Trees model

Three model families face the same held-out 2024 data. Each learns from earlier, non-overlapping periods. A simple median baseline sets the bar.

79.8%

of 2024 cross-county value variation
captured by the selected model¹

$37.9k

mean absolute prediction error
in constant 2024 dollars

3,139

counties in the held-out
2024 evaluation

¹ R² = 0.798 for the selected Extra Trees model, which uses county characteristics as predictors, rather than home values from previous periods.

How was the model chosen?

Within 2014, three county-disjoint folds rank predictors using permutation importance. Six configurations compete on the same 57 training-derived subsets: 342 comparisons on 2019 validation data. The smallest predictor set within 1% of the best validation RMSLE wins, with ties resolved by error, candidate and subset.

The winner is Extra Trees with 30 predictors. It is refitted on 2014 and 2019 before retrospective evaluation on 2024. RMSLE measures error on a logarithmic scale; lower values indicate better proportional accuracy.

How far off are the estimates?

LOWER IS BETTER
Selected using 2019 validation RMSLE
THE TAKEAWAY

County characteristics carry a substantial signal.

The selected Extra Trees model reduces proportional-scale error by 62.6% relative to the training-median baseline.

It still misses important local variation. Predictions tend to understate values at the expensive end of the market.

Download model comparison

Joint model and predictor selection retained 30 of 57 inputs. All six configurations across Elastic Net, Extra Trees and gradient boosting received the same 57 unique training-derived subsets (342 comparisons). The selected design's 2019 RMSLE was 0.17958, versus 0.18097 for the best full-feature reference. The smallest set within 1% of the global minimum validation RMSLE was selected. Compare feature sets · Read the predictor audit.

03 / FEATURE IMPORTANCE

Which features matter most?

Household income, education and other income measures lead the selected model’s ranking.

Relative permutation importance

How to read this chart

Shuffle one predictor across 2024 counties while keeping the fitted model fixed. Repeat ten times and measure the mean increase in RMSLE. The bars scale that increase relative to the highest-ranked predictor (100).

Hover or focus a row for the raw increase and permutation standard deviation. These scores describe predictive reliance. Correlated predictors can substitute for each other; the bars are neither shares of explained variance nor causal effects.

The 2014 rankings proposed subsets; these 2024 diagnostics explain the frozen selected model.

What happens when a whole group is removed?

Relative increase in 2024 RMSLE after refitting without a group, holding the selected specification fixed.

04 / SECONDARY ANALYSES

More questions.

Does the selection procedure generalize to unseen counties?

Five county-disjoint outer folds repeat selection on development data. Pooled out-of-fold R² is 0.8340, RMSLE is 0.1792 and MAE is $25,254. Extra Trees wins all five folds. Neighboring counties may still share information through spatial similarity.

Do climate measures add predictive information?

On the matched 1,594-county evaluation sample, climate inputs reduce RMSLE from 0.22733 to 0.22522 (about 0.9%). This small improvement applies to counties with complete climate coverage.

Do affordability measures improve the model?

On the same 3,139 evaluation counties, adding affordability inputs raises RMSLE from 0.21454 to 0.22092 (about 3.0%). The frozen specification performs worse with this bundle.

How do estimates respond to a feature?

Local response curves vary one predictor near observed county values and average model predictions within supported ranges. They describe the fitted model’s associations; changing a characteristic in the real world can have different consequences.

Explore response curves in the notebook ↗
A DIFFERENT QUESTION

What changes within a county?

The regression estimates associations between changes in the selected characteristics and home values, controlling for persistent county differences and shared period effects. The table shows selected coefficients and their uncertainty.

3,122balanced counties
0.2536within-county R²

The regression uses the same 30 selected inputs across 3 periods. Within R² measures the remaining log-value variation after county and period effects; it differs from held-out predictive R².

See selected estimated associations

County and period fixed effects using the 30 selected inputs; exploratory 95% county-clustered intervals conditional on predictive selection and pooled-median imputation. Percentage predictors are measured in percentage points. Income uses log1p dollars. These are not causal effects.

Secondary analyses use frozen specifications and previously inspected 2024 outcomes. Interpret differences as descriptive sensitivity evidence.

05 / COUNTY ATLAS

Mapping the model's performance

Explore where estimates match reality, and where they fall short.

2020–2024 ACS • 2024 boundaries
Loading county data…
ALASKA & HAWAII SHOWN AS INSETS

Select a county on the map or use the selector. Gray indicates no eligible target value. Boundaries are simplified; DC and small counties are easiest to find by name.

Source: Census ACS five-year estimates. Outcome: median value of owner-occupied housing, not sale price or the value of an individual home.

06 / LIMITATIONS

Read the results in context.

These are observational county-level associations using characteristics and home values from the same ACS period. They describe aggregate survey estimates rather than individual appraisals or advance forecasts. The 2024 outcomes were previously inspected, so the evaluation is retrospective.

Correlated predictors, neighboring counties, incomplete climate coverage and geographic changes constrain interpretation. ACS margins of error, spatially clustered uncertainty and uncertainty from the full model search are outside the reported intervals.