veterans merged county analysis

Overview

Source: /home/coolhand/html/datavis/data_trove/data/policy/veterans/merged_county_analysis.csv

Saturn profiled 3,144 rows across 18 columns. The stats below are deterministic and machine-readable; the prose is a language-model interpretation of those stats (opt-in, added after the fact, never sees raw rows).

STATEFP numeric foreign_key

This is the US Census STATEFP code, a 1-2 digit FIPS identifier for states stored numerically. Values range from 1 to 56 with 51 unique entries across 3144 rows, matching the count of US states plus DC, and the row count aligns with the number of US counties. The near-uniform spread (skew -0.08, kurtosis -1.10) and zero outliers are consistent with a categorical state code rather than a measured quantity.

Treatment: Cast to zero-padded string and treat as a categorical state key for joins, not a numeric feature.

COUNTYFP numeric identifier

COUNTYFP is the 3-digit FIPS county code, stored numerically across 3144 rows with no nulls and 329 unique values. The distribution is heavily right-skewed (skew 2.84, kurtosis 11.4) with a max of 840 well beyond Q3 of 133.5, flagging 176 outliers — expected behavior since FIPS codes are categorical identifiers, not measurements, and high values correspond to specific county assignments.

Treatment: Cast to zero-padded string and combine with STATEFP to form a 5-digit GEOID join key; do not treat as numeric.

COUNTYNS numeric identifier

COUNTYNS is the Census Bureau's permanent numeric ANSI/GNIS identifier for U.S. counties: all 3144 values are unique with no nulls or zeros, and the range (23901 to 2830254) matches the GNIS ID space. The distribution is broad but unremarkable (skew 0.17, kurtosis -0.80), as expected for an ID code rather than a measurement.

Treatment: Treat as a county-level key for joins; do not use as a numeric feature.

GEOIDFQ text identifier

This is the Census Bureau's fully-qualified GEOID (GEOIDFQ) for U.S. counties: every value is exactly 14 characters, single-token, all-caps, and follows the `0500000US` summary-level prefix followed by a state+county FIPS code. All 3144 rows are unique with no nulls or duplicates, matching the count of U.S. counties. Vocab size equals row count (3144), confirming it is a pure identifier with no analytical signal of its own.

Treatment: Use as a left-join key against Census geographies; do not feed into models.

GEOID numeric identifier

GEOID is the 5-digit FIPS code identifying US counties: every one of the 3,144 rows is unique with no nulls, and the range 1001 to 56045 matches the state+county FIPS convention (Alabama through Wyoming). The near-zero skew (-0.08) and flat kurtosis (-1.10) reflect roughly uniform coverage across state codes rather than any meaningful distribution. Treating these as numbers is misleading—they are categorical keys.

Treatment: Cast to zero-padded string and use as a join key to county-level geographies; do not model as numeric.

NAME text label

This column holds short place names — almost certainly US county names, given the dominance of 'Washington' (31), 'Franklin' (26), 'Jefferson' (26), 'Lincoln' (24) and 'Madison' (20), all classic county namesakes. Values are overwhelmingly single-word (one_word_rate 0.934, word_mean 1.07) and short (len_mean 7.0, len_max 30), with no nulls. The 41.5% duplicate_rate is expected here: the same county name recurs across states, so 3144 rows collapse to 1838 unique strings.

Treatment: Treat as a non-unique name; pair with a state/FIPS column before joining or grouping.

NAMELSAD text label

This is the full legal name of US county-equivalents (NAMELSAD from Census TIGER), with 'county' appearing 2999 times alongside 64 'parish' and 47 'city' entries reflecting Louisiana and independent-city conventions. Names are short (mean 14.08 chars, ~2 words) and heavily duplicated across states: 1262 duplicates (40.1%) driven by repeated names like 'Washington County' (30), 'Jefferson County' (25), and 'Franklin County' (24). Only 1882 unique values across 3144 rows, so this field alone does not identify a county.

Treatment: Use as a display label only; join on a state+FIPS code rather than this name to avoid duplicate collisions.

STUSPS categorical foreign_key

STUSPS is the USPS two-letter state abbreviation, with 51 distinct values across 3,144 rows — consistent with US states plus DC at the county grain. Distribution matches county counts: TX leads at 254 (8.08%), followed by GA (159), VA (133), and KY (120). No nulls and high entropy ratio (0.93) indicate clean, well-spread categorical coverage.

Treatment: left-join on this code to state-level reference tables, or one-hot/target-encode for modelling.

STATE_NAME categorical feature

STATE_NAME holds US state labels across 3,144 rows with exactly 51 unique values (50 states plus likely DC) and zero nulls. The distribution mirrors county counts per state: Texas leads at 254 (8.1%), followed by Georgia (159) and Virginia (133), consistent with this being one row per US county. Entropy ratio of 0.93 indicates a fairly even spread across states given their natural county-count differences.

Treatment: use as a categorical grouping key or one-hot/target-encode for modelling.

LSAD categorical metadata

LSAD is a Census Legal/Statistical Area Description code identifying the type of geographic entity for each of 3144 rows. The distribution is extremely imbalanced: code '06' (county) accounts for 95.39% of rows, leaving only 9 distinct codes and an entropy ratio of 0.117. Minor categories like '15', '25', and 'PL' tail off quickly into single-digit counts.

Treatment: Collapse rare codes into an 'other' bucket or drop, since one value dominates.

ALAND numeric feature

ALAND looks like land-area in square meters for 3,144 unique geographic units (matching the U.S. county count), with no nulls or zeros. The distribution is extremely right-skewed (skew 26.8, kurtosis 953) — the median is 1.59B while the max reaches 377B, and 11.5% of rows flag as outliers. A handful of very large areas dominate the mean (2.91B) versus the median.

Treatment: log-transform before modelling to tame the heavy right tail.

AWATER numeric feature

AWATER is the standard Census TIGER field for water-area in square meters, here at what looks like county granularity given n=3144 unique values. The distribution is extremely right-skewed (skew 13.18, kurtosis 210.8): the median is 19.4M but the mean is 222M and the max reaches 25.99B, with 440 outliers (14.0% of rows). One row is zero, and all 3144 values are unique, so this behaves like a continuous geographic feature rather than a key.

Treatment: Apply a log1p transform before any modelling to tame the 13.2 skew and heavy outlier tail.

fips numeric identifier

This is the 5-digit US county FIPS code: every one of the 3144 rows is unique, there are no nulls, and the range 1001–56045 spans the standard state+county encoding. The distribution is essentially uniform across the code space (skew −0.08, kurtosis −1.10), as expected for an identifier rather than a measurement.

Treatment: left-join on this id; do not use as a numeric feature.

active_duty_est numeric feature

Numeric estimate of active-duty population per record, with 3144 rows and 3028 unique values suggesting one row per geographic unit (likely county-level given the row count). The distribution is severely right-skewed (skew 13.14, kurtosis 288.57): median is 11698 but mean is 53782.95 and the max reaches 5240842, with 449 outliers (14.3%). No nulls or zeros, and the IQR of 27868 is dwarfed by a std of 176262.59.

Treatment: log-transform before modelling to tame the heavy right tail.

veterans_est numeric feature

Estimated veteran counts per row (likely US counties given n=3144), ranging from 0 to 244,160 with a median of 1,547.5 but a mean of 5,419.5. The distribution is heavily right-skewed (skew 8.01, kurtosis 100.0) with 408 outliers (12.98%) reflecting a few highly populous areas dwarfing the rest. Near-zero null and zero rates, so coverage is essentially complete.

Treatment: log1p-transform before modelling to tame the heavy right skew.

total_pop numeric feature

Looks like a per-row population total across 3,144 rows (suggestive of US counties), with no nulls and 3,080 unique values. The distribution is severely right-skewed (skew 13.17, kurtosis 289.76): median is 25,784.5 but the mean is 105,310.94 and the max reaches 9,936,690, with 440 rows (14.0%) flagged as outliers. Min is 50 and zero_rate is 0, so every row carries a real count.

Treatment: log-transform before regression to tame the heavy right tail.

active_duty_per_10k numeric feature

A per-capita rate (active duty personnel per 10,000) reported across 3,144 rows with no nulls, no zeros, and every value unique. The distribution is tight around a mean of 4,693.79 and median of 4,733.28 with std 592.22, mildly left-skewed (-0.38), and 57 outliers (1.81%) span a range from 1,708.13 to 7,200.00. The 3,144 row count strongly suggests one record per US county.

Treatment: Use as-is as a continuous feature; the mild skew does not require transformation.

veterans_per_100 numeric feature

This column reports veterans per 100 residents, with 3143 unique values across 3144 rows (likely one row per US county). Values range from 0 to 18.09 with a mean of 6.19 and median of 5.98, showing a mild right skew (0.88) and 125 outliers (~3.98%) on the high end. Only one row is zero, so the distribution is effectively continuous and well-populated.

Treatment: Use as-is for modelling; optionally winsorize the upper ~4% outliers.