County characteristics and US home values

The study asks how county characteristics relate to median owner-occupied home values and whether those relationships generalize across periods and counties. It estimates observational associations, not individual-home prices or causal effects. Website framework, map geometry and hosting remain a separate design step; this pipeline produces the evidence and static chart data for that site.

Run and reproduce

python -m src.cli.run_county_characteristics
python -m unittest discover -s tests -v

Optional arguments are --database PATH, --output-dir PATH, and --quick. Quick mode uses fewer candidates and one annual forecast validation year; its outputs go into a quick/ subdirectory. Full mode uses six characteristics candidates and all five annual forecast validation years. Neither mode changes the source DuckDB marts. Model artifacts are trusted local pickle files; load only files you generated yourself.

Results are in data/modeling/county_characteristics/. Start with REPORT.md and its figures. This generated directory is ignored by Git under the existing repository rules. Source, tests and methodology are versionable; copy selected exports into the eventual website project when its design is chosen.

Dataset and feature contract

  • study.duckdb, table county_period_panel, and county_period_panel.parquet contain the same analytical panel with unique (county_fips, year) keys. FIPS is a five-character string. year is the ACS period end, period_start is four years earlier. The periods end in 2014, 2019 and 2024.
  • The outcome is target_median_owner_occupied_home_value_2024_usd and its natural log log_value. Data preparation excludes missing/nonpositive targets, territories, negative Census sentinel estimates and impossible percentages. Negative climate temperatures remain valid. Models give equal weight to each county-period row; complete longitudinal counties contribute three rows each.
  • Primary predictors are economic, social, housing structure and hazard characteristics. Geography identifiers, price history, geographic price summaries and value-derived ratios never enter the primary design. Percentages remain in percentage points; dollar estimates and counts use log1p.
  • The registry must define a feature in all three periods. A feature with no observed values in 2014 is excluded using training data alone. In these marts, computer/broadband measures fail this observed-availability check despite registry labels existing. Other missing predictors are median-imputed inside each training fit, with missingness indicators; Elastic Net also standardizes its predictors inside each fit.
  • Affordability and mortgage measures are a separate sensitivity analysis. Complementary renter/owner and mortgage/no-mortgage percentages are not both included. Percentage families may still correlate; coefficients and importance are interpreted jointly rather than as independent causal effects.
  • Hazard event totals cover the same five-year ACS window. A complete five-year history is required for each sum. The existing event marts assume absent events are zero; reporting/mapping gaps may remain. Climate is a five-year mean of annual measures, requiring five observed annual values, and is tested only on covered counties. Base and augmented climate models use identical training and test rows.

feature_manifest.csv gives each source, transformed predictor, group, role, unit and inclusion/exclusion reason. feature_definitions_by_year.csv preserves the actual ACS variable mappings. source_key_audit.csv, coverage.csv, missingness.csv and exclusions.csv document the sample and joins. Duplicate source or output keys cause the run to fail.

Geographic comparability

State labels are derived from official FIPS prefixes, including newer county equivalents missing from the static county reference. Missing metro labels are Unknown, not assumed nonmetro. Metro status is a static descriptive label used only for error breakdowns.

Cross sections retain all eligible counties in each period. The longitudinal analysis requires all three periods, complete values for its six specified predictors, and excludes Connecticut and known affected Alaska, South Dakota and Virginia codes. This conservatively excludes name/code-only changes too; no approximate crosswalk is invented. geography_audit.csv reports eligibility. This is a substantial-change screen, not proof that all retained boundaries are identical. Minor boundary changes and ACS sampling margins of error are not modeled. The forecast sensitivity also requires continuous 2010–2024 histories.

Experiments and selection

  1. Characteristics-only prediction. Fit on 2014, select parameters within each family on 2019 RMSLE, and select the overall model against the 2014 training-median baseline using the same validation criterion. Freeze that decision in selection.json. Refit family winners on 2014 + 2019 and evaluate 2024 once per specified experiment. Predictors and targets are contemporaneous: this measures transferable associations, not an advance forecast.
  2. Geographic generalization. Five county-disjoint folds use only 2014 and
    1. Each fold repeats parameter selection on its training counties (2014 to 2019), refits on both periods of those counties, and scores the excluded counties. Thus a county cannot influence preprocessing or tuning for its own geographic validation predictions. Adjacent counties may still be correlated; this is not a spatially buffered evaluation.
  3. Within-county associations. A balanced three-period regression removes county and period means before fitting log value on log1p household income, bachelor’s attainment, vacancy, detached-unit share, commute time and household size. County-clustered covariance accounts for absorbed effects in its degrees of freedom; confidence intervals use a t distribution with counties minus one degrees of freedom. No time-invariant geography coefficients are estimated.
  4. Retrospective annual forecasting. Compare previous value with a learned price-history-only model and history plus characteristics. Independently tune each learned design on expanding 2018–2022 validation folds, fit through 2022, and evaluate 2023–2024 using each row’s observed prior-year features. This is a sequence of one-release-ahead predictions with fixed fitted parameters, not a two-step recursive forecast. It uses equal comparison rows. Adjacent ACS windows overlap and release/CPI vintages are not reconstructed.

The candidate grid reuses the existing Elastic Net, Extra Trees and histogram gradient boosting implementations. selection.json records parameters; run_summary.json records software versions, seed, sample, limitations and a panel fingerprint. The selected ML artifact retains its development fit. If the median baseline wins, the manifest says so and the ML artifact is explicitly labeled best_ml, not claimed to be the selected predictor.

Interpretation and website exports

MAE is in 2024 dollars; RMSLE measures proportional-scale error; R² uses dollar values. There is no R²-based model selection. error_breakdowns.csv reports the selected model’s performance by state, metro status and observed value decile.

feature_group_experiments.csv removes each primary group using the fixed validation-selected ML specification on the same rows, plus affordability and matched climate sensitivities. These are diagnostics, not another selection round. group_permutation.csv jointly permutes all columns in a group (five repetitions) for each model family; its spread is permutation variability, not a confidence interval. Correlated groups can substitute for each other.

standardized_coefficients.csv contains development-fit Elastic Net coefficients; coefficient_stability.csv contains descriptive fits for each period using that fixed specification. associations_by_period.csv contains pairwise Spearman correlations. These descriptive 2024 fits never change the selected predictor.

local_response_curves.csv supplies accumulated local log effects for income, education and vacancy: training-derived bins lie within the 5th–95th percentiles, and test bins require at least 20 rows. Empty bins are omitted, not interpolated. Binwise effects compare predictions at each bin’s edges only for counties inside that bin; these remain model associations, not interventions.

website_data/ contains JSON for the comparison tables and charts. counties.json has FIPS, labels, periods, observed values, six interpretable characteristics, 2024 predictions and residuals; missing values are JSON null. Join to matching-period geometry for a county map. summary.json includes provenance and limitations. The report follows the presentation’s question, data journey, evidence, interpretation, forecasting and limitations sequence.

Source guidance