Methodology: The New Business Map

Methodology v1.1, 16 October 2026. Applies to report v1.1. Changes are versioned; corrections are logged at the foot of this page.

1. Research question

Across US counties and metropolitan areas, where has new-business formation grown fastest since 2019, and how does that growth relate to changes in local real household income and self-employment? The study is descriptive and associational. No causal claim is made or implied.

2. Data sources, vintages and terms

This product uses the Census Bureau Data API but is not endorsed or certified by the Census Bureau.

Vintage choice: ACS 2015–19 and 2020–24 are the two most recent non-overlapping five-year periods. Comparing overlapping five-year estimates is statistically invalid and the Census Bureau advises against it.

3. Definitions

• Applications per 1,000 adults (county c, year t) = 1,000 × BA_ct / adults_ct, adults = resident population aged 18+ on 1 July of year t.
• Change in applications per adult = (rate_2025 / rate_2019 − 1) × 100.
• Applications growth minus adult growth (metro ranking) = % change in applications 2019→2025 − % change in adults 2019→2025, in percentage points.
• Real median household income, 2015–19, in 2024 dollars = B19013 (2015–19 vintage, 2019 $) × 313.689 / 255.657 = × 1.22699. Its MOE is scaled by the same factor. The 2020–24 estimate is already in 2024 dollars.
• Real income change = income_2020–24 − real income_2015–19; percentage change relative to the earlier period.
• Self-employed share = (B24080_005 + 010 + 015 + 020) / B24080_001 × 100. Change in percentage points.
• Quadrant: applications per adult up / real income down; up / up; down / down; down / up. (The data files label the first quadrant "necessity signal" for continuity with v1.0; the report does not use that term because the data argue against a necessity-driven boom.) Point-estimate quadrants and a significance-tested version (income change not significant → "not significant") are both published.
• Metro = metropolitan statistical area (2023 delineations). Applications and adults are summed over member counties; income and self-employment are taken directly from ACS metro-level estimates (medians cannot be summed).
• State = sum over counties, including Connecticut (legacy counties for 2019, planning regions for 2025).

4. Margins of error and significance testing

ACS estimates are published with 90 % margins of error. Following the Census Bureau's Understanding and Using ACS Data (chapter 8):

• Standard error SE = MOE / 1.645.
• Sum or difference of independent estimates: MOE = √(MOE₁² + MOE₂² + …) (root-sum-of-squares), used for the four self-employment cells and for the 18+ age cells.
• Proportion p = X/Y with X ⊂ Y: MOE_p = √(MOE_X² − p²·MOE_Y²) / Y; when the radicand is negative the ratio formula √(MOE_X² + p²·MOE_Y²) / Y is used instead.
• Significance of a change between two periods: Z = |X₁ − X₂| / √(SE₁² + SE₂²). A change is significant at 90 % when Z ≥ 1.645; results at Z ≥ 1.96 (95 %) are also reported. The scaled MOE of the deflated 2015–19 income is used for SE₁.
• Sentinel values in the API (−666666666 not available; −222222222; −555555555 = controlled estimate, treated as zero sampling error) are handled explicitly; counties with a missing income estimate in either period are excluded.
• Multiple testing. Each county (metro) receives one test of its income change, so at 90 % confidence about 5 % of ~800 counties would show a "significant decline" by chance alone (≈ 40; ≈ 20 at 95 %). Counts of significant changes are therefore reported only after Benjamini–Hochberg false-discovery-rate (FDR) control across all counties (separately, all metros) in the ranking universe, using two-sided p-values from the Z statistic above. The primary bar is q ≤ 0.10, with q ≤ 0.05 also reported. Unadjusted flags remain in the county table for individual look-ups; q-values are included in the downloadable county table.
• Measurement error in income change. In smaller counties the ACS income change is mostly sampling noise (estimated reliability of the county income-change measure across the universe: 0.68), which biases any correlation with application growth towards zero. The income relationship is therefore reported for all counties and, separately, for counties where the change is measured precisely (MOE of the change < 3 % of 2015–19 income), for counties with ≥ 250,000 adults, and adult-weighted, with bootstrap 95 % confidence intervals (2,000 resamples, fixed seed).
• Business application counts and population estimates carry no sampling error. County application counts carry differentially-private geometric noise (privacy-loss budget ε = 0.75 for 2019 onward), with variance 2e^(−ε)/(1−e^(−ε))² ≈ 3.4, i.e. a standard deviation of about 1.8 applications per county-year — immaterial above the population threshold. Negative noisy counts are published as zero.

5. Universe and exclusion rules (all published in the county table as flags)

1. Population threshold. Counties with fewer than 50,000 adults in either 2019 or 2025 are excluded from rankings and counts (2,337 of 3,156 counties; the retained 801 hold about 83 % of the US adult population outside Connecticut).
2. Connecticut (17 rows) is excluded from county and metro tables: its eight counties were replaced by nine planning regions in 2022, so 2019 and 2025 geographies differ. Reported at state level only. Puerto Rico is not in the BFS county file.
3. Formation-service concentration screen. EIN applications are located by the address on the application; registered agents, virtual offices and out-of-state LLC formation services therefore concentrate applications in a small number of counties. A county is flagged and removed from rankings if any of: (i) its 2025 rate is ≥ 3 × the national county-sum rate (≥ 63 per 1,000 adults); (ii) its 2025 rate is ≥ 1.5 × its own state's rate and its 2019→2025 growth in rate is ≥ 2 × the national growth (≥ 107 %); (iii) it is a documented out-of-state formation hub (Sheridan, Laramie and Natrona WY; Kent, New Castle and Sussex DE; Carson City NV; Flathead MT). 40 counties are flagged, 17 of them above the population threshold; the full list is available on request. Metros containing a flagged county are excluded from metro rankings (15). The rule is applied mechanically. In v1.1 the state-ratio threshold in rule (ii) was tightened from 2× to 1.5× after review, because the fastest-growing counties under the v1.0 rule included state capitals and cities reported to host national registered-agent firms; this removed seven large counties (Erie PA, Sacramento CA, Ingham MI, Bernalillo NM, Travis TX, Houston AL, Berkshire MA). Counties just below the threshold (Sangamon IL 1.48×, Spokane WA 1.43×, Stark OH 1.40×) remain in the tables and are flagged as cautions in the narrative.
4. Base-year anomaly. A county whose 2019 count exceeds 1.4 × the larger of its 2017 and 2018 counts is excluded, because its 2019 base is not a stable pre-pandemic baseline. The 99th percentile of this ratio among large counties is 1.21; one large county (Fairbanks North Star, AK: 628 → 1,372 → 2,148 → 982) and 81 small ones trip the rule. The same rule is applied to metro totals.
5. Missing data (24 rows): Connecticut duplicates, 2019 Alaska boundary changes, discontinued FIPS codes, and one county with no published income estimate.

6. Computation and verification

• Code: a Python analysis script (Python 3.10, pandas 2.3, numpy 2.2, openpyxl 3.1). Every threshold above is a named parameter in CONFIG.
• Independent verification: A separate validation script re-reads the raw files with the standard library only and recomputes the headline counts; it agrees exactly with the analysis script. Re-running with ACS-derived adults instead of Population Estimates gives 35 (vs 32) counties with > 25 % growth and an unadjusted significant income decline. The FDR counts and the income-relationship checks are produced by a robustness script from the reconciled county table. Totals reconcile with the published national BA series to within 0.07 % (2019) and 0.001 % (2025).
• Charts on this site are drawn directly from the published tables.
• Structural checks: table cell IDs are verified against the tables' own subtotals at run time; the run aborts if the layout differs.

7. Known limitations

• Cross-sectional association across two periods; no causal inference. The positive income association is modest (r ≈ 0.2) and appears only where income change is measured precisely; across all counties it is flat.
• Applications ≠ formations; most applications never become employer businesses.
• ACS five-year estimates average over five years (mid-points 2017 and 2022); income changes are therefore smoothed and lag the 2025 application data.
• Geographic attribution of applications is imperfect (§5.3); ranking positions of individual counties near the top of growth tables should be read with that caution.
• County rates use resident adults as the denominator; commuting, seasonal residents and business addresses in central business districts raise rates in some urban cores (e.g. Manhattan) for reasons unrelated to residents' entrepreneurship.
• Population denominators combine Population Estimates Vintage 2019 (for 2019) and Vintage 2025 (for 2025); later vintages revised the post-2020 base. The effect on rates is small relative to the +53 % change but is not zero.
• A metro is excluded entirely if any member county is flagged by the formation-service screen (e.g. Philadelphia, because New Castle County DE is a documented hub); metro coverage is therefore conservative.
• Inflation adjustment uses the ratio of CPI-U annual averages (2024 / 2019 = 1.22699).
• The monthly context chart shows the Census national series for context only; it is not a finding.

8. Reproducibility, licensing and corrections

• All inputs are public U.S. Government statistical products and may be redistributed; this study redistributes none of them, only links and derived tables.
• Published derived tables: the county table (801 rows) and the metro table (355 rows), with margins of error, significance flags and false-discovery-rate q-values. Fuller tables (all 3,156 counties, state totals, the formation-service list and robustness outputs) and the analysis code are available to journalists and researchers on request. Licence for Censal's derived tables: CC BY 4.0, attribution "Censal Research, The New Business Map (2026)".
• Capability statement: this study uses published aggregate tables. Censal's methods for microdata studies (PUMS/IPUMS extracts with replicate-weight variance estimation) are documented separately.
• Corrections: errors reported to the contact address on censalresearch.com are assessed within five working days; material corrections are made in place, dated and listed below.

Correction log: - v1.1 (16 Oct 2026, before first publication): multiple-testing (FDR) control added; income relationship re-assessed for measurement error (v1.0's "no relationship" replaced by a modest positive association among precisely measured counties); formation-service rule (ii) tightened from 2× to 1.5× state rate (universe 808 → 801 counties, 362 → 355 metros); headline revised accordingly.