Week 2 Notes - Algorithmic Decision Making & Census Data

Algorithmic Decision Making & Census Data

Part 1: Algorithmic Decision Making

What is an Algorithm?

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

Algorithm Decision Making in Gov

• Systems used to assist or replace human decision-makers
  • Based on predictions from models that process historical data containing:
    • Inputs: features, predictors, ind. variables, x, etc.
    • Outputs: labels, outcomes, dep. variables, y, etc.
• Read World Ex: Mortgage lending and tenant screening algorithms

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

Public Sector Context

Long history of government data collection:

• Civic registration systems
• Census data
• Administrative records
• Operations research (post-WW2)

What’s new?

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

Why Gov Uses Algorithms

• Governments have limited budgets and need to serve everyone
• Algorithmic decision making is especially appealing bc it wrongly promises:
  • Efficiency: produces cases faster
  • Consistency: same rules applied to everyone
  • Objectivity: removes human bias
  • Cost savings: fewer staff needed (labor is expensive!!)

When Algorithms Go Wrong

Data Analytics is subjective!

• Every step involves human choice
  • Data cleaning decisions
  • Data coding/classifications
  • Data collection (use of imperfect proxies)
  • Result interpretations
  • Variables chosen to be put in the model
• Human values and biases are embedded!

Bias is everywhere!

• Healthcare algorithms have systematically discriminated against Black patients
• Algorithms used healthcare costs as a proxy for need - Black patients typically incur lower costs due to systemic inequities in access - Resulted in the under-prioritization of Black patients despite equivalent levels of illness
• Criminal Justice algorithms use biased (racist) policing data

Part 2: Active Learning

Done in class

Part 3: Census Data Foundations

Why Census Data Matters

Census Data is the foundations for

• Understanding community demos
• Allocating government resources
• Tracking neighborhood change
• Designing “fair” algorithms

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

Census vs American Community Survey (ACS)

• Decennial Census
  • Everyone counted
  • 9 basic questions (age, race, sex, housing)
  • Constitutional requirement
  • Determines apportionment
• ACS
  • 3% of households analyzed annually
  • Detailed questions (income, education, employment, housing costs)
  • Replaced the old “long form census” in 2005

ACS Estimates - 1 year estimates (areas > 65,000 ppl) - Most current data, too small of sample for most use - 5 year estimate (all areas) - Most reliable data, large sample - What we use most often - Key Point: All ACS data comes with margins of error

Most policy analysis occurs at county, census tract, and block groups levels

2020 Census Innovation: Differential Privacy

• The Challenge: Modern computing can re-identify individuals from census data
• The Solutions: Add mathematical “noise” to protect privacy whilst preserving patterns
• The Controversy: Some places now show pops living underwater or in other impossible places
• Why this matters: Even “objective” data involves subjective choices abt privacy v accuracy
  • Also errors

Accessing Census Data in R

We will use the tidycensus package

Census data structure is as follows: - Data organized into tables - i.e. B19013: Median Household Income - Each table has multiple variables - B19013_001E: Median household income (estimate) - B19013_001M: Median household incomee (margin of errro)

Working with margins of error

• Every ACS estimate comes with uncertainty
  • Large MOE relative to estimate = less reliable
  • Small MOE relative to estimate = more reliable
• In analysis:
  • Always report MOE alongside estimates
  • Be cautious comparing estimates with overlapping error margins
  • Consider using 5year estimates for greater reliability

Two types of Census Data

• Summary Tables (what we’ll mostly use)
  • Precalc statistics by geo
  • Good for mapping, geo comparison
• PUMS - individual records
  • anonymous individual/HH responses
  • Good for custom analysis, regression models

Data sources

• TIGER/Line Files
  • geographic boundaries (.shp)
• Historical Data Sources
  • NHGIS: Historical census data -Neighborhood Change Database -Longitudinal Tract Database: track changes over time

Part 4: Hands on Use

get_acs

Output

• GEOID: Geo identifier
• NAME: Human-readable location name
• variable: Census variable code
• estimate: Actual value
• moe: Margin of error

Professional Tables