The increasing availability of Volunteered Geographic Information (VGI) has enabled new approaches for modeling urban dynamics, particularly in road safety studies on speed management[1]. Despite the growing use of OpenStreetMap (OSM), there is still limited evidence regarding its adequacy to support spatial models of operating speed in urban environments. Across multiple studies, road width, number of lanes, segment length, and distance to intersections consistently predict higher operating speeds. Longer continuous segments and multi‑lane roads increase the 85th-percentile speed [2]. Narrower streets and visually constrained environments reduce speeds [3]. Curves and gradients lower speeds, particularly on residential streets [4,5]. At the same time, naturalistic driving datasets, such as the Brazilian Naturalistic Driving Study (NDS-BR), provide detailed observations of driver behavior but often lack structured spatial variables describing the built environment [6]. This gap raises the following research question: to what extent can OSM-derived variables explain spatial variations in operating speed, and how suitable are they for spatial regression modeling in speed management analysis?
The main objective of this study is to evaluate the explanatory capacity of OSM-derived variables in modeling the 85th-percentile operating speed (V85) of private vehicles and to compare the performance of global spatial regression models. Specifically, the study assesses whether OSM data can provide statistically and spatially meaningful predictors and examines the implications of their limitations for urban road safety studies.
The analysis was conducted in Curitiba, Brazil. NDS-BR collected data from 52 drivers’ routine activities to study road safety for approximately two weeks for each driver. The study utilized GPS devices to gather data without any interference or interaction from researchers regarding routes or destinations between the years of 2019 and 2023. A regular grid of 200 × 200 meters was used as the unit of analysis, integrating operating speed data with socioeconomic variables from the Brazilian Institute of Geography and Statistics (IBGE) and built-environment variables extracted from OSM. This is the standard grid made available from IBGE for statistical analysis and allows similar approaches to be used in other areas of the country. The dependent variable (V85) was calculated from observed vehicle speeds within each grid cell. Due to the nature of the data acquisition process, some cells within the grid were not fully populated with information. The total number of grid cells containing information on inhabitants and households within the entire boundary of Curitiba was 8,956. For analysis and interpretation, grid cells lacking recorded speed data were excluded, resulting in a total of 5,456 grid cells. Independent variables included IBGE database such as population and number of households and OSM database that includes road network length (interpreted as street density as it refers to the sum of all road segments within each grid square), bus stops, number of intersections, traffic signals and speed cameras.
The software utilized for this research were all free and open-source such as R (for statistical analysis and data manipulation) QGIS (for data visualization) and GeoDa (for spatial statistics). Three regression approaches were implemented: Ordinary Least Squares (OLS), Spatial Lag Model (SLM), and Spatial Error Model (SEM). Spatial dependence was modeled using a contiguity-based spatial weights matrix of first order. The dataset was split into training (75%) and testing (25%) subsets for validation. Multicollinearity diagnostics led to the exclusion of highly correlated variables. High values were identified for households and inhabitants (0.97) and for intersections and extension (0.55). So, alternative model specifications were tested.
The OLS results indicated a statistically significant model (p-value < 0.001), but with low explanatory power (R² approx. 0.10), suggesting that linear relationships alone are insufficient to explain spatial variations in operating speed. This is common in complex behavioral phenomena, or in the present case where there is a spatial component that is not adequately accounted for by the model. Among the predictors, road network density and traffic signals showed consistent negative associations with V85, indicating that denser and more regulated urban environments tend to reduce operating speeds. However, variables such as bus stops and speed cameras were not statistically significant (p-value > 0.05) and were excluded from subsequent analyses.
Spatial models improved model fit substantially. The Spatial Lag Model presented the best performance, with lower Akaike Information Criterion (AIC) and higher log-likelihood values compared to OLS and SEM, indicating that spatial interaction effects play a significant role in explaining operating speed. Nevertheless, even with spatial dependence accounted for, the overall explanatory power remained limited, with prediction errors (RMSE approx. 14.8) indicating considerable unexplained variability.
The values obtained (R², RMSE) are likely influenced by the grid configuration used, reflecting the Modifiable Areal Unit Problem. Since each grid cell may aggregate road segments with distinct geometric and functional characteristics, this can reduce the model’s explanatory power by masking local variability. So far, the results highlight both the potential and the limitations of OSM data in road safety modeling. On one hand, OSM provides relevant spatial variables such as network density and traffic control elements that are statistically associated with operating speed and can support spatial regression analyses. On the other hand, the absence of critical attributes identified in the literature (e.g., lane width, road geometry, traffic volume, and detailed functional classification) restricts the explanatory capacity of the models. This suggests that OSM, in its current state, is better suited for exploratory and large-scale analyses rather than high-precision predictive modeling.
From a scientific perspective, this study contributes by providing an empirical assessment of the suitability of OSM data for spatial modeling of operating speed, explicitly quantifying its limitations within a regression framework. From a practical standpoint, the findings inform researchers and practitioners about the conditions under which OSM data can be reliably used in urban safety studies and highlight the need for data enrichment strategies, whether through improved mapping practices or integration with complementary datasets.
Future work will focus on incorporating local spatial models, such as Geographically Weighted Regression, and machine learning approaches, as well as evaluating the impact of OSM data completeness and quality on model performance.