Dutch road safety policy is founded on the Safe Systems vision called Sustainable Safety which was introduced at the beginning of the nineties (Koornstra et al., 1992; Wegman et al., 2008). Agreement on implementation of the vision was reached in 1998, resulting in the construction of large 30 km/h zones. The so-called access roads in these areas have a residential function where slow traffic and motorized traffic mix, for which 30 km/h is considered a safe speed. To clearly signal the 30 km/h speed limit, no dedicated bicycle infrastructure is built, and no road markings are applied, following the principle of ‘self-explaining roads’ (Theeuwes & Godthelp, 1995). Intersections are preferably yield-to-the-right (CROW, 2021), requiring drivers to slow down at most junctions and check for road users approaching from the right. It also occurs that the right of way is arranged with exit structures: speed hump-like elevated section at the end of a street. Traffic passing the exit construction from the side street has to yield. By 2008, the speed limit on 85% of the roads in built-up areas classified as access roads had been reduced to 30 km/h (Weijermars & Wegman, 2011).

On access roads the volume of motor vehicles should be capped at approximately 3,000 per day, while town centres and shopping areas may accommodate a maximum of 5,000 vehicles daily (Van Minnen, 1999). Road authorities started to reduce speed limits from 50 km/h to 30 km/h on busier roads with a through function in recent years. CROW (2023) has published preliminary design guidelines for this so-called ‘distributor road 30’. We do not focus specifically on these roads that are proportionally still rare (see Andriesse, 2021).

Figure 1 illustrates examples of built environment characteristics observed on 30 km/h roads in the Netherlands that are discussed in this section and included in the study.

We conducted our study on a street level, a collection of road sections with the same street name. Firstly, streets are often built at once and subject to similar maintenance, so the street characteristics and surroundings likely match. Secondly, motorists maintain speed when driving straight from one road section to the next.

We utilized the 85^th percentile driving speed, as determined by the NRTP using Floating Car Data (FCD) purchased from Be-Mobile for the year 2023. Kijk in de Vegte (2022) compared the FCD estimates to driving speeds measured at 143 locations using loop detectors, road tubes, and radar. On average, the FCD estimate of the 85^th percentile speed for FCD road segments of 50 m (or shorter to fit between intersections) only differed by 3 km/h from the loop detector measurements. In this study, we used the dataset provided by NRTP in which the speed data is aggregated to NWB road sections yielding a more accurate speed measure.

For our analysis, the data were further aggregated to the street level by calculating the weighted average of road section speeds, using their respective lengths as weights. An alternative method involves dividing the total street length by the travel time, calculated based on the 85th percentile speed for each road section. These two aggregation methods yield nearly identical results (correlation = 0.997). To ensure consistency with the aggregation approach used for other variables, we adopted the weighted average method. Streets shorter than 50 meters were excluded to ensure that the 85^th percentile driving speed estimate was based on data from at least two FCD road segments.

The NRTP provides road feature datasets linked to the NWB via linear referencing (see NDW, 2023). For example, a speed hump is recorded as extending from 34 m to 40 m along a road section with ID 345678. This allowed us to count features such as speed humps and calculate their density per 100 m (representing a fairly high density for these measures) or the proportion of road section length covered by features, such as the percentage of a road section with parking spaces. For parking spaces, the proportion of length refers to the average of the left and right sides of the road meaning that the maximum proportion is 1. A summary of the characteristics and how NRTP collected the data is provided in Table 1.

Building data, mandatory to be collected by each municipality, were extracted from the Basic Register of Addresses and Buildings of September 2024 (version september 2024; Kadaster, 2024a). The area along the carriageway was identified through the NWB line geometry projected at the centre of road areas as registered in the large-Scale Topography Registry, and by using the NRTP pavement width dataset. Flat-end buffers were used along the NWB line geometry, extending half the road width plus 4 m. Before buffering, 10 m at the beginning and end of each section were excluded because of the imprecise location of junctions where road sections start and end. Junctions in NWB still need to be projected in the middle intersection areas in the large-Scale Topography Registry. Since only a narrow parking lane and pavement would fit between the road and any buildings, buildings within these buffers are considered close to the carriageway. The number of buildings within the buffers was counted to determine the density of premises near both sides of the carriageway per 100 m.

In 2024, NRTP partnered with the Dutch Cyclist’s Union to incorporate bicycle infrastructure presence and road access for cyclists from the Cyclist’s Union route database (Fietsersbond, 2024). NRTP linked the Cyclist’s Union data to all road sections with a 30 km/h speed limit based on distance and direction of road sections in both datasets. Since cycling is typically allowed on all 30 km/h roads in the Netherlands, we assumed a road section had separate bicycle tracks if cycling on the carriageway was forbidden, according to the Cyclist’s Union route database. Additionally, visually marked bicycle lanes as recorded by the Cyclist’s Union, were adopted for this study. The proportion of street length with bicycle infrastructure was determined using this information.

We intersected 30 km/h streets with neighbourhood data obtained from Statistics Netherlands (CBS, 2024b) to determine the address density per square km in the area surrounding the street. Further, we calculated the proportion of street length falling within areas designated as shopping zones or industrial areas through land use data from Statistics Netherlands (CBS, 2024a).

NRTP estimated traffic volumes for all NWB road sections in 2023 using FCD purchased from TomTom. While Be-Mobile data was used for speeds, it did not include probe numbers. The estimates are based on the number of vehicles in the FCD sample (known as probes) and account for variations in coverage across the road network, such as lower coverage on lower-ranked roads (see Hastig, 2024). A comparison with loop detector measurements showed the estimated traffic volume differed less than 10% at 44% of loop detectors and less than 50% at over 90% (NDW, 2024). In this study, we aggregate the average daily traffic volumes of road sections to streets, with the values weighted by the length of each road section.

According to flow theory, driving speed typically decreases as traffic volume increases. However, Andriesse (2021) found the contrary for even the most busy Dutch 30 km/h streets. This can be explained by the fact that these streets operate well below their capacity. Nevertheless, it is unlikely that speed continues to increase proportionally as traffic volume grows. To account for this tapering effect, we use the natural logarithm of traffic volume as an independent variable. Similarly, while street length is associated with increased driving speed (Rahim & Daniel, 2023) this increase is likely to flatten of as street length grows further. Therefore, both traffic volume and street length are log-transformed. The coefficient in the regression model should be interpreted as the change in the 85^th percentile driving speed when these log-transformed variables increase by one unit. This corresponds to the original variables increasing by a factor of e, the base of the natural logarithm ( e≈ 2.72). For example, when a variable increases from 1 to e, its natural logarithm increases from 0 to 1; from e to e², it increases from 1 to 2, etc.

Out of the 175,000 30 km/h streets longer than 50 m, we included 159,000 in our study, comprising a total street length of 47,000 km. Streets were excluded due to missing speed data or the absence of neighbouring streets, as the lack of neighbours prevents the incorporation of spatial influences in spatial regression modelling.

Figure 2 shows the distribution of 85th percentile driving speeds by length and frequency. The distribution of 85th percentile driving speeds based on road length is somewhat skewed towards higher values, with fewer streets experiencing low speeds. The distribution based on the number of streets is less skewed, as speeding is less prevalent on shorter streets, which occur more frequently. The average 85th percentile driving speed is 32.2 km/h, with a standard deviation of 4.6 km/h. Moran’s I for the 85th percentile driving speed is 0.32 (p = 0.001), indicating positive spatial autocorrelation, likely due to drivers maintaining speeds as they move from one street to another.

Table 2 presents descriptive statistics, including the mean, standard deviation, and correlations of all variables with the 85^th percentile driving speed. Variables are listed in descending order of the absolute value of their correlations. Notably, the proportion of street length with closed pavement, the logarithm of motor vehicles per day, and the logarithm of street length exhibit positive correlations exceeding 0.3. Conversely, the proportion of road section length within the total street length with a school address is the only variable with a correlation that is not statistically significant. The averages reveal substantial variation in the frequency of characteristics. For instance, raised intersections, with an average density of 0.28 per 100 m, are relatively common. In contrast, features such as cycle tracks and school zones, with an average proportion of street length of just 1%, can be considered rare.

We began the regression analysis with an OLS model on the 85^th percentile driving speed, including all variables listed in Table 2 but without interaction terms. All GVIF values remained below 1.5, indicating that multicollinearity is unlikely. Because the model showed residual spatial autocorrelation (Moran’s I =0.24, p<0.001) and was supported through the LM test (p<0.001), we fitted a spatial error model. The overall model fit was with a pseudo-R² of 46% better than the OLS model with an R² of 39%; further supported through the AIC scores (AIC_SER= 804106, AIC_OLS= 816480). The spatial error coefficient λ was 0.32 suggesting moderate spatial dependence in the error terms. Table 3 presents the main results of the spatial error model for 85^th percentile driving speed. The coefficients and confidence intervals provide insight into how each variable is related to driving speed.

The variables are presented in Table 3 in the same order as in Table 2. The direction of all regression coefficients aligns with the direct correlations reported in Table 2. Once again, the proportion of street length with a school address is the only variable not statistically significant. The following variables are associated with higher 85^th percentile driving speeds: closed pavement, higher traffic volumes, longer street lengths, bicycle lanes and tracks, longer average road section lengths, wider carriageway widths, a greater proportion of street length through industrial zones, separated carriageways, exit constructions on side roads, and a greater proportion of street length through school zones with signs and/or markings. The association with school zones is weak, showing only a 0.4 km/h increase in speed for streets that are entirely within such zones. Nonetheless, this positive association is contrary to our expectations, as school zones are typically designed to reduce speed.

In contrast, the following variables are associated with lower 85^th percentile driving speeds: premises close to the carriageway, curves (especially tight curves), high address density, parking spaces along the street, exit constructions on the street, speed humps, raised intersections, and a greater proportion of street length through shopping areas. Although road narrowing density also shows a statistically significant negative association, the effect is negligible with only 0.2 km/h lower speed per road narrowing per 100 m.

The association between the 85^th percentile driving speed within a 30 km/h zone and the two intersection measures—exit constructions and raised intersections—seems complex. Our results indicate that a density of one speed hump per 100 m is associated with a greater speed reduction than one raised intersection on the same street. However, a raised intersection not only benefits the street where it is located but also has an impact on the intersecting streets, broadening its effect. Within a 30 km/h zone, raised intersections can achieve a comparable reduction in the 85^th percentile speed as speed humps. A raised intersection is also a valuable speed-reducing measure as it lowers speed at a location where road users interact.

Exit structures are also applied at intersections. On the street with the exit structure, speeds typically decrease. However, traffic on adjacent streets, where priority is granted, often experiences an increase in speed. When both the street with the exit structure and the adjacent priority street are quiet and short, these effects tend to offset each other. The overall effect is generally a speed reduction when both are longer or busier streets with closed pavement. However, longer streets are more frequently prioritised due to exit structures on quieter side streets, which can lead to an increase in average speeds within the 30 km/h zone: according to our results the speed increase on long, asphalt priority roads outweighs the speed reduction on short, open-pavement streets with exit structures. In the Dutch context, many exit structures connect 30 km/h streets to 50 km/h priority roads. In these cases, the exit structure slows traffic on the 30 km/h street, while drivers on the 50 km/h road maintain their speed due to priority signage, also without an exit structure. This configuration usually reduces speed within the 30 km/h zone. In summary, exit structures are particularly effective for lowering speed when used as entrances to 30 km/h zones.

Many factors on 30 km/h streets make a difference particularly when a street entices high driving speeds due to its greater length, high traffic volume and closed pavement. For measures designed to cause vertical acceleration like speed humps, raised intersections, and exit constructions, this can be explained by the associated discomfort among drivers remaining under a threshold acceptable for drivers while driving no faster than 30 km/h. We also observe an interaction with the presence of buildings close to the road and parking bays, which often coexist with parked vehicles. Through optic flow, these elements contribute to a perception of higher speed. However, we lack a specific explanation for why this perception would be stronger in certain environments, as suggested by the interaction effect.

The various similar interaction effects can be summarised as creating a greater difference on long, busy streets with closed pavement. Conversely, they translate to less difference on short, quiet streets with open pavement. This may be explained by the theory of self-explaining roads, along with 30 km/h being perceived as relatively slow and as a lower threshold for the fastest 15% of motorists. Considering the length of road sections, we also observe this pattern in the distribution of the 85^th percentile driving speed, where the range of speeds above 30 km/h is greater than that below 30 km/h. If a street conveys characteristics of a 30 km/h zone—open pavement, quiet and short—this 30 km/h speed might represent what these motorists perceive as an appropriate target, thereby diminishing the effects for those in our model in such a low-speed environment.

A key strength of this study is its large sample size, which provides substantial statistical power. However, its cross-sectional design limits our ability to establish causal relationships. For example, speed-reduction measures might be implemented in response to high driving speeds, making these measures more frequent in areas associated with higher speeds. Unless all relevant conditions are statistically controlled for, this may lead to an underestimation of the measures’ true effects (Elvik, 2011). In contrast, a before-after study would avoid this limitation by comparing speed in the same locations before and after an intervention.

Another limitation is the inconsistent data quality, which may have affected the effect estimates. For instance, we know that the database of speed humps, exit constructions, and raised intersections is 82% complete (Mieras & Drolenga, 2023). Such incomplete data can blur the estimated effects, potentially underestimating the true effect. We recommend further investigation into data quality and repeating a similar study in the future with improved data, both in the Netherlands and internationally, to enable comparisons. For example, in the Netherlands, NRTP is working to enhance the quality of the aforementioned database, which will allow for a more accurate assessment of the effects of speed humps, exit constructions, and raised intersections.

Finally, we assume that the measures are implemented uniformly. However, in practice, there is variation. For example, speed humps differ in height, and we would ideally account for such variations in our model as taller humps are likely to be more effective at reducing speed. This will become feasible with the collection of more detailed data, which should therefore be considered.

New research questions emerge based on the needs of road authorities. Road authorities are increasingly lowering speed limits to 30 km/h on busier through roads. In response, CROW (2023) has published preliminary design guidelines for these so-called ‘distributor roads 30.’ Research is needed to assess how effectively these guidelines encourage drivers to adhere to the 30 km/h speed limit. As more such roads are implemented, opportunities for research will also increase.

This study has shown the value of examining speed behaviour for roads with a specific speed limit. Similar large-scale speed research is recommended for roads with other speed limits. In the Netherlands, while 30 km/h roads serve as access roads in urban areas, 60 km/h roads function as access roads in rural areas, where cyclists are typically mixed with motorized traffic. From a road safety perspective, investigating speed behaviour on these roads in relation to their characteristics is equally important as on 30 km/h roads.

In addition to understanding the relationship between measures and driving speed, our study suggests several options for practice. Despite the good coverage of floating car data for 30 km/h roads, there are roads where insufficient data is available or cannot be appropriately linked to the national road database. Driving speed on these streets can be estimated using a model like the one developed in this study. The model could also be applied to evaluate the design of expanding residential areas, where opportunities to adapt the built environment are most abundant. Furthermore, a modelling approach for 85^th percentile driving speed, such as the one used in the current study, could be employed to visualize and quantify potential speed reductions from various measures for existing streets, providing valuable insights to inform road safety policy.

Paul Schepers: Conceptualization, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review & editing. Werner van Loo: Conceptualization, Data curation, Validation. Wouter Mieras: Conceptualization, Methodology. Hans Drolenga: Conceptualization. Dick de Waard: Conceptualization. Marco Helbich: Methodology, Writing—review & editing.

Data used for this study can be downloaded or requested from:

The authors have no conflict of interests to declare.

This research paper has benefited from the use of ChatGPT, a generative AI tool, to enhance the readability and language of the text. The AI was utilized exclusively to improve clarity, coherence, and grammar, with all substantive content, analysis, and interpretations remaining the sole responsibility of the authors.

This study did not require formal ethical approval, as it exclusively utilized existing open-source data. The data analysed were publicly available and did not include any personally identifiable information or sensitive content. All research activities adhered to ethical standards for the use of secondary data.

Handling editor: Haneen Farah, Delft University of Technology, the Netherlands

Reviewers: Rune Elvik, Institute of Transport Economics, Norway; George Yannis, National Technical University of Athens, Greece