County Presidential Data Source Reference

County Presidential Election Returns 2000-2024 from MIT Election Data and Science Lab (MEDSL). Use when analyzing county-level presidential vote shares, party trends, turnout, or geographic voting patterns. Key join column county_fips enables linking to census, education (CCD/SAIPE), and demographic datasets. Requires Harvard Dataverse API key (HARVARD_DATAVERSE_API_KEY env var). Categorical variables use uppercase strings, not Portal integer codes. Critical caveat: naive mode='TOTAL' filtering silently drops ~1,000 counties in 2020+ data — use 3-pattern reconstruction.

The authoritative source for county-level U.S. presidential election returns spanning 2000-2024. Provides candidate-level vote counts across all 50 states and DC, enabling vote share analysis, partisan trend mapping, and cross-domain geographic research via FIPS code joins.

CRITICAL: Value Encoding

This dataset uses uppercase string codes for categorical variables (party, mode, candidate, state) rather than integer codes. Empty strings (

""

) appear as undocumented values in

party

(501 rows, 2024) and

mode

(2,795 rows, 2024).

Context	party	mode	candidate
Standard values	DEMOCRAT , REPUBLICAN	TOTAL	BARACK OBAMA
Aggregate/meta values	OTHER , ""	"" , ELECTION DAY	OTHER , UNDERVOTES

See

./references/variable-definitions.md

for complete encoding tables.

Prerequisites

API Key Required: This data source requires a Harvard Dataverse API key to fetch data. Unlike education data sources (which use the Urban Institute's free, unauthenticated API), election data is hosted on Harvard Dataverse and requires authentication.

Setup instructions:

1. Create a free Harvard Dataverse account at https://dataverse.harvard.edu/
2. Log in, navigate to your account name (top-right) → API Token
3. Click "Create Token" and copy it
4. Set the environment variable before launching Claude Code:

   export HARVARD_DATAVERSE_API_KEY="your_token_here"
   

   For Docker users: run this inside the container after

   docker compose exec daaf-docker bash

   but before

   claude

   . To make it persistent across sessions, add it to

   ~/.bashrc

   .

If the key is missing, any fetch script will fail with a

KeyError: 'HARVARD_DATAVERSE_API_KEY'

. The orchestrator should check for this variable's existence before dispatching Stage 5 fetch tasks that use this data source.

What is the MEDSL County Presidential Returns Dataset?

• Producer: MIT Election Data and Science Lab (MEDSL)
• Coverage: County-level presidential election returns, 50 states + DC
• Frequency: Every 4 years (presidential election cycle)
• Available years: 2000, 2004, 2008, 2012, 2016, 2020, 2024
• Primary identifier:

  county_fips

  (5-digit FIPS code, stored as integer)
• Record unit: One row per county-year-candidate-party-mode combination
• Total records: 94,151 rows x 12 columns (~8.4 MB). Note: rows per year vary dramatically — 2020/2024 have ~2x rows due to mode breakdowns (~22K vs ~9.5K for earlier years)
• Source: Harvard Dataverse (DOI: 10.7910/DVN/VOQCHQ)

Reference File Structure

variable-definitions.md

coded-values.md

columns.md

quality-notes.md

mode-reconstruction.md

interpretations.md

Decision Trees

What analysis do I need?

Analyzing presidential election data?
├─ County-level vote shares → Use 3-pattern mode reconstruction (./references/mode-reconstruction.md)
│   └─ Longitudinal (cross-year) → MUST reconstruct TOTAL for 2020+ (naive filter drops ~1,000 counties)
│   └─ Single year (pre-2020) → Safe to filter mode='TOTAL'
│   └─ Single year (2020/2024) → Reconstruct unless analyzing a known TOTAL-only state
├─ Party trends → Group by year + party, use party column (not candidate name)
│   └─ Third parties → See ./references/coded-values.md (GREEN/LIBERTARIAN vary by year)
├─ Turnout analysis → Use totalvotes column (dedup per county-year before summing!)
├─ Joining with other data → Use county_fips as join key (zero-pad to 5 chars first!)
│   └─ Census/ACS data → Join on county_fips (standard 5-digit string)
│   └─ Education data (CCD/SAIPE) → Join on county_fips
│   └─ Null FIPS? → See ./references/quality-notes.md (CT, ME, RI)
└─ Voting method analysis → 2020 and 2024 only, see mode breakdown

Is this a data quality issue?

Unexpected values?
├─ county_fips is null → CT/ME/RI in pre-2020 years (52 rows)
├─ county_fips > 72999 → Kansas City MO (FIPS 2938000, non-standard)
├─ county_fips join failures → Zero-pad to 5 chars! (AR codes = 4 digits as int)
├─ AR FIPS 5135 has two counties → Source data error: St. Francis under Sharp County
│   └─ See ./references/quality-notes.md #arkansas-fips-contamination (BLOCKER)
├─ CT counties missing from shapefile → 2022+ TIGER uses planning regions, not counties
│   └─ See ./references/quality-notes.md #connecticut-fips-geography-mismatch
├─ ~1,000 counties missing after mode filter → Use 3-pattern reconstruction, not naive filter
│   └─ See ./references/mode-reconstruction.md
├─ candidate is not a person → UNDERVOTES/OVERVOTES/SPOILED/TOTAL VOTES CAST
│   └─ Filter these OUT for candidate-level analysis
├─ party is empty string → 501 rows in 2024, undocumented
├─ mode is empty string → 2,795 rows in 2024 (may be totals OR breakdowns per state)
├─ candidatevotes is null → 37 rows (NM 2024, mode breakdown)
├─ sum(candidatevotes) > totalvotes → 49 county-years (minor rounding)
├─ Duplicate rows → 83 exact duplicates exist
└─ Alaska 2004 → District-level data, not county; FIPS = 2001-2099
    └─ See ./references/quality-notes.md #alaska-2004

Quick Reference: Election Variables

Party Values by Year

""

Mode Values by Year

TOTAL

TOTAL

TOTAL

""

For complete mode values see

./references/coded-values.md

Key Identifiers

county_fips

6037

state_po

CA

state

state

CALIFORNIA

WARNING: FIPS Zero-Padding Required for Joins

county_fips

is stored as Int64. When converting to string for joins with Census, SAIPE, CCD, or other datasets, zero-pad to 5 characters or Arkansas and other small-FIPS states will produce 4-digit codes that fail to match.

df = df.with_columns(pl.col("county_fips").cast(pl.Utf8).str.zfill(5).alias("county_fips_str"))

Missing Data Codes

null

county_fips

null

candidatevotes

0

totalvotes

0

candidatevotes

""

party

""

mode

Non-Candidate Entries in candidate Column

OTHER

TOTAL VOTES CAST

UNDERVOTES

OVERVOTES

SPOILED

Filter these out for candidate-level vote share analysis.

Best Practice: Party-Based Identification

Always use

party

column for identifying party affiliation, never

candidate

name. Candidate names are inconsistent across years (e.g., "DONALD TRUMP" in 2016 vs "DONALD J TRUMP" in 2020/2024). The

party

column (

DEMOCRAT

,

REPUBLICAN

) is stable across all years.

# CORRECT — stable across years
dem = df.filter(pl.col("party") == "DEMOCRAT")

# WRONG — misses 2016 or 2020/2024 depending on which name you use
trump = df.filter(pl.col("candidate") == "DONALD TRUMP")

Data Access

Dataset Paths

10.7910/DVN/VOQCHQ

County Presidential Returns 2000-2024.md

sources-president.tab

Codebooks

County Presidential Returns 2000-2024.md

Codebook is a Markdown file bundled in the Harvard Dataverse deposit. For human reference. The QA methodology paper is at: https://www.nature.com/articles/s41597-022-01745-0

Truth Hierarchy: When interpreting variable values, apply this priority:

1. Actual data file (what you observe in the TSV) -- this IS the truth
2. Live codebook (Markdown file in Dataverse) -- authoritative documentation, may lag
3. This skill documentation -- convenient summary, may drift from codebook

If this documentation contradicts the codebook, trust the codebook. If the codebook contradicts observed data, trust the data and investigate.

Example Fetch

# Fetch from Harvard Dataverse API
import os, requests, polars as pl, io

api_key = os.environ["HARVARD_DATAVERSE_API_KEY"]
# Get file ID from dataset metadata first, then download
# File: countypres_2000-2024.tab (TSV format)
file_url = "https://dataverse.harvard.edu/api/access/datafile/{file_id}"
r = requests.get(file_url, params={"key": api_key, "format": "original"})
df = pl.read_csv(io.BytesIO(r.content), separator='\t')

# Filter to California, 2020, TOTAL mode only
ca_2020 = df.filter(
    (pl.col("state_po") == "CA") &
    (pl.col("year") == 2020) &
    (pl.col("mode") == "TOTAL")
)

Filtering

# Common filter patterns for county presidential data

# 1. Cross-year analysis: WARNING — naive filter drops ~1,000 counties in 2020+!
#    Use 3-pattern mode reconstruction instead. See ./references/mode-reconstruction.md
#    The single-line filter below is ONLY safe for single-year analysis on a state
#    known to have TOTAL rows (e.g., 2000-2016 data, or a confirmed TOTAL-only state).
longitudinal = df.filter(pl.col("mode") == "TOTAL")  # UNSAFE for 2020+ multi-state!

# 2. Remove non-candidate rows (UNDERVOTES, OVERVOTES, etc.)
candidates_only = df.filter(
    ~pl.col("candidate").is_in(["TOTAL VOTES CAST", "UNDERVOTES", "OVERVOTES", "SPOILED"])
)

# 3. Major party analysis — RECOMMENDED: use party column, not candidate name
#    Candidate names change across years (e.g., "DONALD TRUMP" vs "DONALD J TRUMP")
two_party = df.filter(pl.col("party").is_in(["DEMOCRAT", "REPUBLICAN"]))

# 4. Exclude Alaska 2004 anomaly
clean = df.filter(~((pl.col("state_po") == "AK") & (pl.col("year") == 2004)))

# 5. Exclude rows with null county_fips (for join operations)
joinable = df.filter(pl.col("county_fips").is_not_null())

Common Pitfalls

TOTAL

mode == 'TOTAL'

UNDERVOTES

OVERVOTES

SPOILED

TOTAL VOTES CAST

county_fips

state_po

county_name

""

pl.col("county_fips").cast(pl.Utf8).str.zfill(5)

party

candidate

totalvotes

Critical: Mode Column Behavior Change

WARNING: Naive

mode == "TOTAL"

filtering drops ~1,000 counties in 2020+ data. Multiple states report ONLY mode breakdowns (no TOTAL rows). A simple filter silently removes all their counties. Use 3-pattern mode reconstruction instead. See

./references/mode-reconstruction.md

for the full code pattern and validation.

The

mode

• 2000-2016: All rows have

  mode = 'TOTAL'

  (aggregate county totals only)
• 2020: 11 states report by voting method alongside TOTAL; 10 states have ONLY breakdowns (AR, AZ, GA, IA, KY, MD, NC, OK, SC, VA)
• 2024: Additional breakdown states; some have empty string

  mode

  (which may represent totals OR breakdowns depending on the state)

For any cross-year or multi-county analysis, use 3-pattern mode reconstruction:

1. Pattern 1: TOTAL present → keep TOTAL, drop breakdowns
2. Pattern 2: Only breakdowns, no TOTAL → sum

   candidatevotes

   across modes
3. Pattern 3: Empty-string mode = totals → reclassify after per-state verification

Empty-string detection must be per-state — NC 2024 empty-string rows are breakdowns (multiple per county-candidate), while other states' empty-string rows are totals (one per county-candidate). See

./references/mode-reconstruction.md

Related Data Sources

county_fips

education-data-source-saipe

county_fips

education-data-source-ccd

county_fips

Note: This is the first election domain dataset in DAAF. Cross-domain joins with education data are possible via

county_fips

Topic Index

./references/columns.md

./references/columns.md

./references/coded-values.md

./references/coded-values.md

./references/coded-values.md

./references/coded-values.md

./references/variable-definitions.md

./references/quality-notes.md

./references/quality-notes.md

./references/quality-notes.md

./references/quality-notes.md

./references/quality-notes.md

./references/quality-notes.md

./references/quality-notes.md

./references/mode-reconstruction.md

./references/mode-reconstruction.md

./references/mode-reconstruction.md

./references/mode-reconstruction.md

./references/columns.md

./references/quality-notes.md

./references/quality-notes.md

./references/coded-values.md

./references/variable-definitions.md

./references/interpretations.md

./scripts/