Research Radar: Using Machine Learning to Map State Capacity

This week, I want to highlight an interesting new AI-enabled research project designed to help us measure state ability to exercise authority.  This experiment with designing ways to model state power could transform how we understand and strengthen democracies. 

A team led by economist Gustav Agneman and colleages at the Norwegian University of Science and Technology, has developed an innovative machine learning method to measure something previously difficult to quantify: the actual presence and effectiveness of the state at local levels.

Mapping the "Uneven Reach of the State"

The core insight from Agneman is simple: state authority varies dramatically within countries, but our traditional measurement tools are too blunt to capture these variations accurately.

Here’s their three-step method:

1. First, they built an index of how people experience state authority using citizen survey data about law and order in their communities

2. Next, they connected this data with geographic indicators like altitude, road infrastructure, and territorial control

3. Finally, they trained a machine learning algorithm to predict state presence across regions where no survey data exists

The result? Highly detailed maps showing where state capacity is strong versus weak across Sub-Saharan Africa at unprecedented resolution that they claim are predictive:

• “We find that predicted state presence correlates positively with the number of government employees at the district level in Ghana.
• Our measure is shown to be strongly correlated with an indicator of administrative capacity developed by Lee and Zhang (2017).
• We demonstrate cross-sectional and inter-temporal validity using Afrobarometer data from countries and survey rounds not included in the machine learning model.” 

Why This Research Matters for Democracy

Measuring governance is hard to do with any accuracy. Their approach of combining survey measures with certain objective data points could transform several aspects of democracy work:

• Targeted interventions: Instead of country-level approaches, democracy practitioners could focus resources where governance gaps are largest in specific geographies.

• Beyond national averages: Democracy metrics often miss sub-national variations in participation and representation

• Measuring what matters: Centered on citizen experiences rather than official government claims

The Limitations We Should Consider

While novel, this approach raises important questions:

1. Data inequities persist: The methodology still relies on survey data for training, which remains limited in many of the regions that need measurement most. Even where data exists, the questions are limited.

2. Power dynamics unaddressed: The technical approach doesn't directly tackle the political forces that create governance gaps

3. Privacy and surveillance concerns: Granular mapping of state capacity could potentially be misused by authoritarian regimes

Looking Forward: Implications for Democratic Innovation

For those of us working on democracy innovations, the research offers several promising directions:

• Precision democracy building: Could we use similar methods to identify democratic participation deserts?

• Measuring representation gaps: How might this approach help identify communities whose needs aren't reflected in governance decisions?

• Evaluating interventions: Could these methods help us better measure the impact of democracy-strengthening initiatives?

As with any method, the value will ultimately depend on how it's applied. The most promising applications will be those that center citizen experiences and strengthen democratic accountability rather than merely making governance more efficient.

Further Reading

The full research paper is available at the Journal of Development Economics, titled "The uneven reach of the state: A novel approach to mapping local state presence." The authors have also made their data publicly available for other researchers and practitioners to build upon.