📊 statistics

Crowd-Sourced Geographies of Income: Using Google Maps Points of Interest as High-Frequency Proxies for Sub-Municipal Income Estimation in Sao Paulo, Brazil

This paper demonstrates that crowd-sourced Google Maps Points of Interest can serve as a high-frequency, low-cost proxy for estimating household income at the sub-municipal level in São Paulo, achieving a strong predictive performance (R² of 0.65) that offers a viable alternative to costly, infrequent census data for social policy in Brazil.

Original authors: Adrienne C. Kinney, Anya Workman, Ademar Takeo Akabane, Jenna Barac, Paulo Fernando Braga Carvalho, Jeova Farias, Fernando Nascimento, Paulo Ricardo da Silva Oliveira

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Adrienne C. Kinney, Anya Workman, Ademar Takeo Akabane, Jenna Barac, Paulo Fernando Braga Carvalho, Jeova Farias, Fernando Nascimento, Paulo Ricardo da Silva Oliveira

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to draw a map of a city's wealth, but the only map you have is a photograph taken twelve years ago. That is the reality for many policymakers in Brazil and other middle-income countries. They need to know exactly which neighborhoods are struggling to send help, but the official government census—the gold standard for this data—happens only once every ten years. In the long gaps between these censuses, the city changes: new shops open, old ones close, and families move, but the official income map stays frozen in time. To solve this, scientists are turning to "crowd-sourced" data, which is information collected by millions of regular people using their phones and apps. Think of it like a massive, living diary of a city, written by everyone who walks, drives, or shops there. This paper lives at the intersection of geography, economics, and computer science, asking a simple but powerful question: Can we guess how much money a neighborhood makes just by looking at what businesses and places are listed on Google Maps right now?

The researchers in this study decided to test this idea in São Paulo, Brazil, a massive city with over 26,000 tiny neighborhood blocks called "census sectors." They treated Google Maps like a giant, digital inventory of the city. Instead of asking people how much money they make, they asked the map: "What is in this neighborhood?" They looked for specific types of places, like banks, schools, hospitals, churches, and restaurants. The theory is that the mix of these places acts like a fingerprint for wealth. A neighborhood full of high-end coffee shops and private clinics likely has different income levels than one dominated by small, informal markets or specific types of community centers.

The team built a computer model to learn the connection between these "Points of Interest" (POIs) and actual income data from the 2022 census. They fed the model the list of places found in each neighborhood and taught it to predict the income. To make sure the model was actually learning and just memorizing the answers, they used a clever trick: they hid entire strips of the city from the computer during training and only showed it the answers for those hidden strips at the very end. This is like giving a student a practice test with most of the questions, then giving them a completely new section of the test they've never seen before to see if they really understand the material.

The results were quite promising. The best model, which used a specific mathematical technique to simplify the data, was able to predict the income of these unseen neighborhoods with an accuracy score (called R²) of about 0.65. In plain English, this means the model could explain roughly two-thirds of the differences in income between neighborhoods just by looking at the list of businesses on Google Maps. The study suggests that this method is a fast, cheap, and up-to-date way to refresh income maps between official censuses. It can help governments spot neighborhoods that are changing quickly or where people might be falling through the cracks of the official system.

However, the paper is careful not to call this a perfect solution. The model wasn't great at predicting the very richest or the very poorest areas; it tended to guess that the rich were a bit poorer and the poor were a bit richer than they actually were. The authors explain that this happens because the data has limits: in very rich areas, the map might not list every single luxury shop, and in very poor areas, many businesses might be informal and not listed on Google at all. Furthermore, the study explicitly argues against the idea that this data should replace the official census entirely. Instead, they suggest it should be used as a "living signal" to update estimates between the big, slow, official surveys. The study also highlights that the data is biased toward formal businesses; if a neighborhood relies on street vendors or informal services that aren't on Google Maps, the model might miss the true economic picture.

Ultimately, this paper suggests that while we can't perfectly replace the census with a Google search, we can use the digital footprint of a city to get a much fresher, more frequent look at its economic health. It's a tool that turns the chaotic, everyday activity of a city—where people shop, eat, and work—into a readable map of wealth, offering a way to keep social policies from becoming outdated before they even hit the ground.

Technical Summary: Crowd-Sourced Geographies of Income in São Paulo

Problem Statement
Accurate, up-to-date income data at the sub-municipal scale is critical for social policy in middle-income countries like Brazil, yet traditional census mechanisms fail to provide it with the necessary frequency. In Brazil, the primary source for small-area income estimates is the Instituto Brasileiro de Geografia e Estatística (IBGE) Censo Demográfico, conducted decennially. The most recent complete census was in 2010; the subsequent census was delayed by budgetary constraints and the COVID-19 pandemic, with fieldwork only beginning in 2022 and income data released incrementally from 2023. Consequently, social assistance planning in São Paulo (population 11.45 million) relies on income maps that are 12–15 years out of date. While administrative registries like Cadastro Único are continuously updated, they only reflect families reached by active outreach, potentially missing the most vulnerable populations. There is a need for a low-cost, high-frequency proxy to bridge the intercensal gap and identify registration gaps.

Methodology
The study proposes using crowd-sourced Points of Interest (POIs) from Google Maps as a proxy for household income across 26,625 census sectors in São Paulo.

Data Collection:
- Target Variable: Average gross nominal monthly income of the household reference person from the 2022 IBGE Census. The raw income distribution is highly right-skewed; therefore, the natural log of income is used for modeling to normalize residuals, with raw Reais (R$) metrics reported for interpretation.
- Predictor Data: POI data was retrieved via the Google Places API in April–May 2026. The authors selected 47 "anchor" primary types based on hypothesized relevance to income (e.g., banks, retail, healthcare, education).
- Query Design: For each census sector, up to 20 POIs were queried from each of three groups within a 2 km radius of the sector centroid, capped at 60 POIs per sector. This resulted in a sparse $26,625 \times 386$ count matrix (386 distinct subtypes derived from the 47 anchors).
Feature Engineering:
- To address high dimensionality and sparsity, the authors applied two decomposition methods: Principal Component Analysis (PCA) and Non-Negative Matrix Factorization (NMF).
- PCA: Applied to log-transformed, mean-centered data to maximize variance.
- NMF: Applied to square-root transformed data to stabilize variance, enforcing non-negativity for parts-based decomposition.
- Embeddings were tested at ranks $K = 5, 10, 15, 20, 50$ , alongside models using raw 386-dimensional counts.
Modeling and Validation:
- Spatial Validation: To prevent data leakage caused by the overlapping 2 km query windows of neighboring sectors, a rigorous spatial validation design was employed. Two parallel stripes (approx. 15% of sectors) were withheld as a test set. A 3 km buffer around these stripes excluded sectors from both training and testing. The remaining pool was split into a training set and a validation set, where the validation set contained 20% of the training sectors, ensuring income distribution representativeness.
- Algorithms: A sweep of seven regression models (Ridge, Bayesian Ridge, Random Forest, Extra Trees, Gradient Boosting, LightGBM, XGBoost) was trained on the decomposed features and raw counts.
- Evaluation: Performance was measured using $R^2$ , Mean Absolute Error (MAE), and Absolute Relative Error (ARER$) on the held-out test set. Uncertainty was quantified via bootstrapping (2,000 draws), and spatial autocorrelation of residuals was assessed using Moran's I.

Key Results

Predictive Performance: The best-performing model (NMF with $K=50$ features and Gradient Boosting Regressor) achieved a held-out $R^2$ of 0.646 (95% CI [0.62, 0.67]). Performance was stable across decomposition methods and feature extraction techniques, with the top five finalists showing overlapping confidence intervals.
Dimensionality: Low-dimensional embeddings (even at $K=5$ ) retained most of the predictive signal, performing comparably to the full raw feature set.
Interpretability: Decomposition analysis revealed interpretable latent factors:
- Negative Income Signal: High loadings for "Church" and "Bar" POIs, which correlate with lower-income sectors. This aligns with literature on the territorial expansion of evangelical organizations in peripheral, lower-income areas.
- Positive Income Signal: High loadings for "Doctor," "Dentist," and "Parking" POIs, which concentrate in affluent areas. This reflects the spatial accessibility of high-complexity healthcare and commercial destinations.
Limitations in Prediction: The model exhibits systematic errors at the tails of the income distribution, under-predicting the highest income sectors and over-predicting the lowest. Residuals show significant positive spatial autocorrelation (Moran's I = 0.334), indicating that POI composition alone misses spatially structured information, particularly in very high-density affluent areas (where the 60-POI cap may be insufficient) and sparse low-income areas.

Significance and Claims
The paper claims that commercial, crowd-sourced geospatial data can serve as a practical, low-cost complement to conventional income statistics during intercensal periods.

Policy Utility: Unlike census maps which are static snapshots, POI-based proxies are "living signals" that can be re-queried monthly or quarterly to track neighborhood changes, flag shifting conditions, and cross-check administrative registries.
Accessibility: The approach relies on interpretable models (tree-based regressors on decomposed features) and low computational costs, making it readily adoptable by under-resourced policy teams compared to "black box" deep learning or satellite imagery approaches.
Theoretical Contribution: The study extends the logic of Economic Complexity—inferring development from the structure of what is present in a territory—to the sub-municipal scale. It demonstrates that the mix of institutions and services (POIs) physically present in a neighborhood carries a strong signal of household income.
Future Directions: The authors suggest extending this framework to multidimensional poverty metrics (education, health, capabilities) and combining POI data with satellite imagery or transportation data to address tail-level prediction errors.

The study does not claim to replace census data or administrative registries but rather to provide a timely, independent signal to inform social policy and resource allocation in the absence of frequent official statistics.

Technical Summary: Crowd-Sourced Geographies of Income in São Paulo

More like this