Measuring Political Attitudes and Attributes with Drift Diffusion Models

APSA 2026


Andrew Trexler
UW-Madison

Christopher D. Johnston
Duke University

Latent Attributes


  • Preferences
  • Values
  • Ideology
  • Personality traits
  • Knowledge
  • Multi-item scale
    • Additive index
    • Factor analysis
    • Latent class analysis
    • Structural equation modeling

Response Latency Data


  • Easily collected with modern survey software
  • Potentially provides valuable information about the target latent trait
    • Easy decisions (i.e., strong preferences) should be quicker
    • Difficult decisions (i.e. weak or close preferences) should take longer

Drift Diffusion Model


  • First introduced by Ratcliff (1978), used often in psych/marketing
  • Useful for modeling decision processes between 2 alternatives
  • Incorporates response time into estimation

Reproduced from Henrich et al. 2023

Data


  • Original Survey
    • Recruited 1,648 U.S. adults via Lucid Marketplace
    • Fielded 2024 June 1-7
    • Analysis sample n = 1,531 respondents
    • 15 political knowledge items displayed in random order
  • 2024 ANES Pre-election Survey
    • 4 knowledge items

2024 Lucid Survey

Knowledge Items


Item Correct (%) Item Correct (%)
State Governor Party 75.9 Identify Senate Pres 38.0
Bill of Rights 70.6 Filibuster Chamber 37.2
State Leg Control 61.2 Senate Term Length 34.4
SCOTUS Term length 60.7 Electoral College Tie 25.7
Veto Override Margin 52.0 Balance of SCOTUS 20.8
Identify House Spkr 47.2 Identify UK PM 19.1
Freedom of Religion 44.5 Identify Fed Chair 18.6
Identify Chief Justice 40.1 Catch Question 6.0*

Distribution of Response Times

Estimated Latent Knowledge

Estimated Latent Knowledge

Predicting Future Correct Answers

Predicting Future Correct Answers

Predicting Future Correct Answers

Difference in R2

2024 ANES

Knowledge Items

Item Correct (%)
Control of U.S. House 65.2
Control of U.S. Senate 62.0
Senate Term Length 42.8
Foreign Aid Spending 28.7
Catch Question 1.9*

Results

Results

Computational Costs


  • Wiener diffusion models are finicky
  • Estimation takes hours
  • Potentially sensitive to specified priors
  • Struggles to handle outliers (not uncommon!)

Future Directions (?)


  • Many other potential quantities of interest
  • Values, preferences, ideology, beliefs
  • Applications for conjoint experiments
  • Refining priors to improve estimation
  • Extensions to more than 2 alternatives

Thanks!


  • Timing data includes substantive information
  • Easy to collect with web surveys
  • May enable improved estimation with fewer survey items


Contact:

Andrew Trexler
UW-Madison
atrexler.com

Chris Johnston
Duke University
sites.duke.edu/chrisjohnston