Algorithmic Decision Making & Census Data

Week 2: MUSA 5080

Dr. Elizabeth Delmelle

2025-09-15

Today’s Agenda

What We’ll Cover

Part 1: Algorithmic Decision Making

• What are algorithms in public policy?
• When algorithmic decision making goes wrong
• Current policy responses

Part 2: Active Learning

• Small group scenarios: designing ethical algorithms
• Discussion and reflection

Part 3: Census Data Foundations

• Understanding census data for policy analysis
• Geography and data availability

Part 4: Hands-On Census Data with R

• Live demonstration of key functions
• Practice exercises

Part 1: Algorithmic Decision Making

Opening Question

Discuss with your table (1 minutes):

What is an algorithm?

Think beyond just computer code - how do you make decisions in daily life?

What Is An Algorithm?

Definition: A set of rules or instructions for solving a problem or completing a task

Examples:

• Recipe for cooking

• Directions to get somewhere
  

• Decision tree for hiring

• Computer program that processes data to make predictions

Algorithmic Decision Making in Government

Systems used to assist or replace human decision-makers

Based on predictions from models that process historical data containing:

• Inputs (“features”, “predictors”, “independent variables”, “x”)
• Outputs (“labels”, “outcome”, “dependent variable”, “y”)

Real-World Examples

Criminal Justice Recidivism risk scores for bail and sentencing decisions

Housing & Finance
Mortgage lending and tenant screening algorithms

Healthcare Patient care prioritization and resource allocation

Clarifying Key Terms

Data Science → Computer science/engineering focus on algorithms and methods

Data Analytics → Application of data science methods to other disciplines

Machine Learning → Algorithms for classification & prediction that learn from data

AI → Algorithms that adjust and improve across iterations (neural networks, etc.)

Public Sector Context

Long history of government data collection:

• Civic registration systems
  
• Census data
• Administrative records
• Operations research (post-WWII)

What’s new?

• More data (official and “accidental”)
• Focus on prediction rather than explanation
• Harder to interpret and explain

Why Government Uses Algorithms

Governments have limited budgets and need to serve everyone

Algorithmic decision making is especially appealing because it promises:

• Efficiency - process more cases faster
• Consistency - same rules applied to everyone
  
• Objectivity - removes human bias
• Cost savings - fewer staff needed

But does it deliver on these promises?

When Algorithms Go Wrong

Remember: Data Analytics Is Subjective

Every step involves human choices:

• Data cleaning decisions
• Data coding or classification
  
• Data collection - use of imperfect proxies
• How you interpret results
• What variables you put in the model

These choices embed human values and biases

Healthcare Algorithm Bias

The Problem:

Algorithm used to identify high-risk patients for additional care systematically discriminated against Black patients

What Went Wrong:

• Algorithm used healthcare costs as a proxy for need
• Black patients typically incur lower costs due to systemic inequities in access
  
• Result: Black patients under-prioritized despite equivalent levels of illness

Scale: Used by hospitals and insurers for over 200 million people annually

Source: Obermeyer et al. (2019), Science

Criminal Justice Algorithm Bias

COMPAS Recidivism Prediction:

The Problem:

• Algorithm 2x as likely to falsely flag Black defendants as high risk
• White defendants often rated low risk even when they do reoffend

Why This Happens:

• Historical arrest data reflects biased policing patterns

• Socioeconomic proxies correlate with race

• “Objective” data contains subjective human decisions

Source: ProPublica investigation

Dutch Welfare Fraud Detection

The Problem:

• “Black box” system operated in secrecy
• Impossible for individuals to understand or challenge decisions
• Disproportionately targeted vulnerable populations

Court Ruling:

• Breached privacy rights under European Convention on Human Rights
• Highlighted unfair profiling and discrimination
• System eventually shut down

Part 2: Active Learning Exercise

Small Group Challenge (10 Minutes! We keep a tight ship around here.)

At your table, pick one scenario and answer three prompts.

Prompts (plain English, no tech):

1. Proxy: What would you use to stand in for what you want?
2. Blind spot: What data gap or historical bias could skew results?
3. Harm + Guardrail: Who could be harmed, and one simple safeguard?

Pick one scenario

1. Emergency response prioritization (natural disasters)
   
2. School enrollment assignment
   
3. Automated traffic enforcement (red-light cameras)
   
4. Housing assistance allocation
   
5. Predictive policing

Example (so you see the level)

Scenario: Emergency response

• Proxy: 911 call volume → stand-in for “need”
  
• Blind spot: Under-calling where trust/connectivity is low
  
• Harm + Guardrail: Wealthier areas over-prioritized → add a vulnerability boost (age/disability) and a minimum-service floor per zone

Discuss at your table (8 minutes)

Answer these out loud and on one device or notepad:

• Proxy → “We’d use ____ as a stand-in for ____.”
• Blind spot → “This could miss/undercount ____ because ____.”
• Harm + Guardrail → “Group ____ could be hurt by ____. We’d add ____ (one safeguard).”

Choose ONE guardrail type:

• Prioritize vulnerable groups
  
• Cap disparities across areas (simple rule)
  
• Human review + appeals for edge cases
  
• Replace a bad proxy (collect the right thing)
  
• Publish criteria & run a periodic bias check

Lightning shares (2–3 tables)

In ≤20 seconds, say:

• Scenario, one proxy, one harm, one guardrail

Class quick poll: Would that guardrail help?

• 👍 Green light
• 👎 Red light

Part 3: Census Data Foundations

Why Census Data Matters

Census data is the foundation for:

• Understanding community demographics

• Allocating government resources
  

• Tracking neighborhood change

• Designing fair algorithms (like those we just discussed)

Connection: The same demographic data used in census goes into many of the algorithms we analyzed

Census vs. American Community Survey

Decennial Census (2020)

• Everyone counted every 10 years
• 9 basic questions: age, race, sex, housing
• Constitutional requirement
• Determines political representation

American Community Survey (ACS)

• 3% of households surveyed annually
• Detailed questions: income, education, employment, housing costs
• Replaced the old “long form” in 2005
• A big source of data we’ll use this semester

ACS Estimates: What You Need to Know

1-Year Estimates (areas > 65,000 people)

• Most current data, smallest sample

5-Year Estimates (all areas including census tracts)

• Most reliable data, largest sample
• What you’ll use most often

Key Point: All ACS data comes with margins of error - we’ll learn to work with uncertainty

Census Geography Hierarchy

Nation
├── Regions  
├── States
│   ├── Counties
│   │   ├── Census Tracts (1,500-8,000 people)
│   │   │   ├── Block Groups (600-3,000 people)  
│   │   │   │   └── Blocks (≈85 people, Decennial only)

Most policy analysis happens at:

• County level - state and regional planning
• Census tract level - neighborhood analysis
• Block group level - very local analysis (tempting, but big MOEs)

2020 Census Innovation: Differential Privacy

The Challenge: Modern computing can “re-identify” individuals from census data

The Solution: Add mathematical “noise” to protect privacy while preserving overall patterns

The Controversy: Some places now show populations living “underwater” or other impossible results

Why This Matters: Even “objective” data involves subjective choices about privacy vs. accuracy. Also, the errors.

Accessing Census Data in R

Traditional approach: Download CSV files from Census website

Modern approach: Use R packages to access data directly

Benefits of programmatic access:

• Always get latest data
• Reproducible workflows
  
• Automatic geographic boundaries
• Built-in error handling

We’ll use the tidycensus package starting in Lab 1

Understanding ACS Data Structure

Data organized in tables:

• B19013 - Median Household Income
• B25003 - Housing Tenure (Own/Rent)
  
• B15003 - Educational Attainment
• B08301 - Commuting to Work

Each table has multiple variables:

• B19013_001E = Median household income (estimate)
• B19013_001M = Median household income (margin of error)

You’ll learn to find the right variables for your research questions

Working with Margins of Error

Every ACS estimate comes with uncertainty

Rule of thumb:

• Large MOE relative to estimate = less reliable
• Small MOE relative to estimate = more reliable

In your analysis:

• Always report MOE alongside estimates
• Be cautious comparing estimates with overlapping error margins
• Consider using 5-year estimates for greater reliability

Two Types of Census Data

Summary Tables (what we’ll use mostly)

• Pre-calculated statistics by geography
• Median income, percent college-educated, etc.
• Good for: Mapping, comparing places

PUMS - Individual Records

• Anonymous individual/household responses
• Good for: Custom analysis, regression models
• More complex but more flexible

When New Data Comes Out

ACS 1-year estimates: Released in September (previous year’s data)

ACS 5-year estimates: Released in December

Decennial Census: Released on rolling schedule over 2-3 years

For Lab 1: We’ll use 2018-2022 ACS 5-year estimates (latest available)

Data Sources You’ll Use

TIGER/Line Files

• Geographic boundaries (shapefiles)
• Census tracts, counties, states
• Now released as shapefiles (easier to use!)

Historical Data Sources:

• NHGIS (nhgis.org) - Historical census data
• Neighborhood Change Database
• Longitudinal Tract Database - Track changes over time

Part 4: Hands-On Census Data with R

Live Demo Setup

Let’s see tidycensus in action with some basic examples:

Follow along: We’ll work through examples together, then you’ll practice in Lab 1

Basic get_acs() Function

Most important function you’ll use:

Rows: 52
Columns: 5
$ GEOID    <chr> "01", "02", "04", "05", "06", "08", "09", "10", "11", "12", "…
$ NAME     <chr> "Alabama", "Alaska", "Arizona", "Arkansas", "California", "Co…
$ variable <chr> "B01003_001", "B01003_001", "B01003_001", "B01003_001", "B010…
$ estimate <dbl> 5028092, 734821, 7172282, 3018669, 39356104, 5770790, 3611317…
$ moe      <dbl> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, N…

Key parameters: geography, variables, year, survey

Understanding the Output

Every ACS result includes:

• GEOID - Geographic identifier
• NAME - Human-readable location name
  
• variable - Census variable code
• estimate - The actual value
• moe - Margin of error

This is the foundation for all your analysis

Working with Multiple Variables

# A tibble: 6 × 6
  GEOID NAME                 total_popE total_popM median_incomeE median_incomeM
  <chr> <chr>                     <dbl>      <dbl>          <dbl>          <dbl>
1 42001 Adams County, Penns…     104604         NA          78975           3334
2 42003 Allegheny County, P…    1245310         NA          72537            869
3 42005 Armstrong County, P…      65538         NA          61011           2202
4 42007 Beaver County, Penn…     167629         NA          67194           1531
5 42009 Bedford County, Pen…      47613         NA          58337           2606
6 42011 Berks County, Penns…     428483         NA          74617           1191

Note: output = "wide" gives you one row per place, multiple columns for variables

Data Cleaning Essentials

Clean up messy geographic names:

# A tibble: 67 × 2
   NAME                           county_name
   <chr>                          <chr>      
 1 Adams County, Pennsylvania     Adams      
 2 Allegheny County, Pennsylvania Allegheny  
 3 Armstrong County, Pennsylvania Armstrong  
 4 Beaver County, Pennsylvania    Beaver     
 5 Bedford County, Pennsylvania   Bedford    
 6 Berks County, Pennsylvania     Berks      
 7 Blair County, Pennsylvania     Blair      
 8 Bradford County, Pennsylvania  Bradford   
 9 Bucks County, Pennsylvania     Bucks      
10 Butler County, Pennsylvania    Butler     
# ℹ 57 more rows

Functions you’ll use: str_remove(), str_extract(), str_replace()

Calculating Data Reliability

This is crucial for policy work:

# A tibble: 2 × 2
  reliability         n
  <chr>           <int>
1 High Confidence    57
2 Moderate           10

Key functions: case_when() for categories, MOE calculations

Finding Patterns with dplyr

• Find counties with highest uncertainty
• Summarize by reliability category

Professional Tables

Making results presentation-ready:

Counties with Highest Income Data Uncertainty

County	Median Income	MOE %
Forest	46,188	9.99
Sullivan	62,910	9.25

Key points:

• Use kable() for professional formatting
• Add descriptive column names and captions
  
• Format numbers appropriately

Quick Practice

Try this with a neighbor (5 minutes):

Using the pa_reliability data we just created:

1. Filter for counties with “High Confidence” data
2. Arrange by median income (highest first)
   
3. Select county name and median income
4. Slice the top 3 counties

We’ll share answers before moving on

Policy Connection

Why this matters:

• Algorithmic fairness: Unreliable data can bias automated decisions
• Resource allocation: Know which areas need extra attention
• Equity analysis: Some communities may be systematically under-counted
• Professional credibility: Always assess your data quality

This connects directly to our algorithmic bias discussion

Connecting the Dots

From Algorithms to Analysis

Today’s key connections:

Algorithmic Decision Making → Understanding why your analysis matters for real policy decisions

Data Subjectivity → Why we emphasize transparent, reproducible methods in this class

Census Data → The foundation for most urban planning and policy analysis

R Skills → The tools to do this work professionally and ethically

Questions for Reflection

As you work with data this semester, ask:

1. What assumptions am I making in my data choices?
2. Who might be excluded from my analysis?
3. How could my findings be misused if taken out of context?
4. What would I want policymakers to understand about my methods?

These questions will make you a more thoughtful analyst and better future policymaker

Next Steps

Before Next Class

1. Complete Lab 0 if you haven’t finished
2. Post your weekly notes - reflect on today’s discussion
3. Start Lab 1 - census data exploration (begins today!)

Questions?