TM2 PopulationSim Process Overview

What is PopulationSim?

PopulationSim is a synthetic population generator that creates realistic household and person records for transportation modeling. The TM2 (Travel Model 2) pipeline synthesizes a population for the 9-county San Francisco Bay Area.

Overall Process Flow

PUMS Data → Geographic Crosswalk → Seed Population → Marginal Controls → PopulationSim → Synthetic Population

1. Data Inputs

• PUMS (Public Use Microdata Sample): Real anonymized household/person records from US Census
• Geographic Definitions: MAZ (Micro Analysis Zones), TAZ (Traffic Analysis Zones), PUMA (Public Use Microdata Areas)
• Marginal Controls: Target totals by geography (households by income, persons by age, etc.)

2. Geographic Framework

• Region: 9-county San Francisco Bay Area
• Counties: Sequential numbering 1-9 (SF=1, San Mateo=2, Santa Clara=3, Alameda=4, Contra Costa=5, Solano=6, Napa=7, Sonoma=8, Marin=9)
• Zones: 4735 Zones, 39587 MAZs
• PUMAs: 104 Public Use Microdata Areas covering the region

Step-by-Step Process

Step 1: PUMS Data Download

Purpose: Obtain raw household and person microdata from US Census Input: Census API or cached files Output:

• households_2023_raw.csv (~175k Bay Area households)
• persons_2023_raw.csv (~400k Bay Area persons)

What happens:

• Downloads 2023 5-year American Community Survey PUMS data
• Filters to Bay Area PUMAs only
• Adds inflation adjustment (2023 to 2010 dollars)
• Adds unique identifiers and crosswalk fields

Step 2: Geographic Crosswalk Creation

Purpose: Link geographic zones (MAZ-TAZ-PUMA-County) for population synthesis Input:

• TM2 zone definition files (blocks_mazs_tazs.csv, mazs_tazs_all_geog.csv)
• County mapping configuration Output: geo_cross_walk_tm2_updated.csv

What happens:

• Creates MAZ-to-TAZ-to-PUMA-to-County relationships
• Resolves conflicts (TAZs split across PUMAs)
• Converts FIPS county codes to sequential 1-9 system
• Validates all geographic relationships

Step 3: Seed Population Generation

Purpose: Process PUMS data into PopulationSim-ready format Input:

• Raw PUMS households/persons files
• Geographic crosswalk
• Income inflation factors Output:
• seed_households.csv (households with geographic assignments)
• seed_persons.csv (persons linked to households)

What happens:

• Assigns households to PUMAs based on their original PUMA
• Links persons to households via unique household IDs
• Applies income inflation from 2023 to 2010 dollars
• Creates PopulationSim-compatible field names and formats
• Handles Group Quarters (GQ) population specially

Step 4: Marginal Controls Generation

Purpose: Create target totals that synthetic population must match Input:

• Census API data (ACS tables)
• Geographic crosswalk
• Control specifications (controls.csv) Output:
• maz_marginals.csv (MAZ-level targets)
• taz_marginals.csv (TAZ-level targets)
• county_marginals.csv (County-level targets)

What happens:

• Downloads Census data for all control variables
• Aggregates to appropriate geographic levels
• Handles Group Quarters controls separately
• Creates age-income cross-tabulations
• Ensures control totals are consistent across geographies

Step 5: PopulationSim Synthesis

Purpose: Generate synthetic population matching marginal controls Input:

• Seed households and persons
• Marginal controls at all geographic levels
• PopulationSim configuration files Output:
• synthetic_households.csv (final synthetic households)
• synthetic_persons.csv (final synthetic persons)
• County summary tables

What happens:

• Balancing: Adjusts seed population weights to match controls
• Integerization: Converts fractional weights to whole households
• Assignment: Places households into specific MAZs
• Uses iterative proportional fitting (IPF) algorithm
• Handles multiple geographic levels simultaneously

Key Algorithms

Geographic Assignment

1. PUMA Assignment: Households stay in their original PUMA
2. County Assignment: Based on PUMA-to-county crosswalk
3. TAZ/MAZ Assignment: PopulationSim chooses based on controls

Control Balancing

1. IPF (Iterative Proportional Fitting): Adjusts weights to match marginals
2. Multi-level convergence: Ensures consistency across MAZ/TAZ/County
3. Constraint handling: Respects geographic and demographic relationships

Group Quarters Handling (Updated October 2025)

Important Approach Change: PopulationSim now uses person-level group quarters controls that align directly with Census data structure.

1. Person-level controls: GQ controls count individuals, not households
   • pers_gq_university: University GQ persons (Census P5_008N)
   • pers_gq_noninstitutional: Military + other GQ persons (Census P5_009N+P5_011N+P5_012N)
2. Selective institutional inclusion:
   • ✅ INCLUDED: University/college housing (dorms, student housing)
   • ✅ INCLUDED: Military barracks and base housing
   • ✅ INCLUDED: Other non-institutional group quarters (group homes, worker housing)
   • ❌ EXCLUDED: Nursing homes, prisons, mental health institutions
3. Two-level structure:
   • Household level: hhgqtype (0=regular, 1=university, 2=noninstitutional)
   • Person level: gq_type (0=regular, 1=university, 2=noninstitutional)
4. Direct Census alignment: Controls use Census person counts without household-level conversion
5. Travel behavior: Non-institutional GQ residents participate in regular travel patterns

Rationale: Person-level controls eliminate conversion assumptions and ensure direct data consistency with Census P5 series tables while maintaining travel modeling utility.

Data Quality Measures

Validation Checks

• Control totals: Synthetic population matches marginal targets within tolerance
• Geographic consistency: All households properly assigned to valid MAZs
• Demographic realism: Age/income distributions match Census patterns
• Household composition: Realistic household sizes and types

Key Metrics

• Convergence: IPF algorithm convergence to target tolerances
• Assignment rates: Percentage of households successfully assigned
• Control fit: How closely synthetic matches target marginals
• Geographic coverage: All MAZs receive appropriate population

Technology Stack

Core Software

• PopulationSim: Synthetic population generation engine
• ActivitySim: Framework for population synthesis and modeling
• Python: Primary programming language
• Pandas: Data manipulation and analysis

Data Sources

• US Census Bureau: PUMS microdata and ACS marginal data
• Census API: Automated data download
• TM2 Model: Geographic zone definitions
• MTC/ABAG: Regional planning agency data

Output Uses

Quality Assurance

Automated Validation

• Control total verification
• Geographic consistency checks
• Data type and range validation
• Cross-tabulation comparisons with Census

Manual Review

• County-level summary comparisons
• Age/income distribution review
• Household size and composition analysis
• Geographic distribution patterns