Risk of bicycle collisions and ‘safety in numbers’: a natural experiment using the local introduction of e-scooters in England

Abstract

Background

Ecological studies hypothesise a ‘safety in numbers’ (SiN) effect whereby road safety for bicycles and other micromobility users improves as their numbers increase, due to behavioural changes of motorists. Causal interpretation of these studies is difficult due to confounding and reverse causation. The introduction of electric scooter (e-scooter) rental schemes in selected districts in England meant an increase in micromobility users in these areas, which presented an opportunity to test the SiN hypothesis using a natural experiment.

Methods

Time-series analysis of police data on road collisions in local authorities (LAs) in Great Britain, 2015–2023. Random-effects Poisson regression time-series models compared collision rates in LA districts with an e-scooter trial (n=41) versus matched control districts (n=41). Primary outcomes were all road collisions and bicycle collisions. Models adjusted for time; seasonality; baseline collision rate; COVID-19 period; and preintervention/postintervention period (proxied by intervention group/COVID-19 period interaction).

Results

The rate of bicycle collisions reduced following the introduction of the schemes, compared with control districts (incidence rate ratio (IRR) 0.78, 95% CI 0.68 to 0.89 during peak COVID-19; IRR 0.87, 95% CI 0.77 to 0.99 in the post-COVID-19 period). This effect was specific to bicycle collisions and strongest in the subgroup of serious/fatal collisions.

Conclusions

Findings suggest that the increase of a new and sustainable mode of transport, e-scooters, may have reduced bicycle collisions. This could have far-reaching benefits including reduced injuries, safer environments, and public health and environmental benefits if more people choose bicycles and micromobility over car transport. Findings should be verified in further work.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Systematic review and meta-analysis show consistent evidence for a ‘safety in numbers’ (SiN) effect, whereby road safety for bicycles and other active-mode and micromobility transports improves as their numbers increase, but these studies are mainly cross-sectional so causal interpretation is difficult.

WHAT THIS STUDY ADDS

• Our findings support and add to the existing evidence on SiN, using a stronger natural experiment study design, which overcomes some of the limitations of past evidence.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• This study supports the hypothesis that the introduction of electric scooters improves road safety for cyclists. This could reduce injuries and encourage more people to choose active-mode and micromobility transport options, bringing public health benefits from increased physical activity and environmental benefits from reducing car transport. These are important policy implications relevant to the Sustainable Development Goals; future research should confirm and extend these preliminary observations in other data sets and look at other indirect effects (both positive and negative) of such schemes.

Background

In road safety literature, the ‘safety in numbers’ (SiN) hypothesis is that the safety of a mode of transport improves as numbers increase. For example, a lower rate of collisions involving cyclists when there are more cyclists on the roads. This also applies to pedestrians and other active-mode or micromobility transport options. This has important policy implications for improving road safety and reducing injuries, but also has public health and environmental benefits if more people choose micromobility transport options over cars, all of which contribute to the Sustainable Development Goals.¹

A recent systematic review and meta-analysis found evidence consistent with the SiN hypothesis² from 45 international studies covering a variety of contexts; microlevel (specific junctions), mesolevel (parts of a city) and macrolevel (city-wide). Nearly all found evidence for SiN, although there was variation in the magnitude of effect estimates. The main proposed mechanism behind this association is that motorists are more careful and safer drivers in response to the higher volume of vulnerable road-users, but alternative explanations include changes to traffic patterns such as physical clustering at junctions; wider culture change; and demographic changes among cyclists (and users of other modes of transport).

The evidence for SiN relies mainly on cross-sectional studies (only one study was longitudinal). These studies cannot control for reverse causation—that is, that perhaps safer road environments lead to more cyclists. They also have limited ability to adjust for confounders, for example, road infrastructure like bicycle lanes which could encourage more cyclists and segregate cyclists from cars. Interestingly, a stronger SiN effect was observed in those studies that did control for more confounding factors.

Public policies, including those to encourage active modes of transport, are rarely tested using conventional trial methodology, often because this is not feasible. This makes causal interpretation difficult. However, it is now well established that natural experiment evaluations can provide strong evidence for making causal claims to guide policymakers.^3–5

This study tests the SiN hypothesis using the natural experiment of the introduction of electric scooter (e-scooter) rental schemes in selected urban areas in England in 2020, following the start of the COVID-19 pandemic in the UK. These schemes represented an increase in micromobility transport in intervention areas that did not occur in other control areas. (It is currently illegal in the UK to ride a private e-scooter on the road or in public spaces, so numbers outside of formal trial areas are expected to be much lower.) On the road, e-scooters are comparable to bicycles in terms of their size, speed and road position; we postulated that the increase in e-scooter numbers would lead to a reduced rate of bicycle collisions in trial areas as predicted by the SiN hypothesis.

Methods

Aims

The primary aim was to identify whether local authority (LA) districts that introduced e-scooter schemes showed a relative reduction in the risk of bicycle collisions, compared with similar districts without e-scooter schemes.

Study design

This was a time-series analysis with a natural experiment design, set in Great Britain from 2015 to 2023.

Data

All data sources for this study are publicly available. Data on official e-scooter trial areas came from the Department for Transport, Arup and NatCen report on a national evaluation of e-scooter trials.⁶ Road safety data on collisions reported by police came from the UK government Department for Transport STATS-19 statistics.⁷ Our focus was recorded collisions, and the term ‘bicycle collision’ meant any collision involving a bicycle. Details of associated casualties/injuries were beyond the scope of this project. Data covered the period January 2015 to June 2023. The final 6 months were provisional data; all other data were final release. Population denominator data came from the Office for National Statistics (ONS)⁸ or other publicly available sources.⁹

Outcomes and exposures

The main outcomes were the rate of all reported collisions and the rate of reported bicycle collisions. Secondary outcomes were rates by other vehicle types (e-scooters, cars, motorcycles). The exposure of interest was e-scooter rental schemes. LA ONS districts (administrative, geographical areas of local government¹⁰) were classified as intervention areas (those with an official e-scooter rental scheme) or control areas (those without).

Statistical methods and models

Analysis was performed at the LA ONS district level. Matched control districts were selected 1:1 using propensity score matching (PSM)¹¹ on district population size and baseline rate of collisions in January 2019. One intervention district (Birmingham) was excluded due to failure to match with an appropriate control. City of London (intervention district) was excluded due to its relatively low population denominator, but high number of commuters and thereby collisions, resulting in an artificially high collision rate. Eight other intervention districts were excluded due to district boundary changes, which invalidate the timeseries because of changing baseline populations (online supplemental appendix 1).

Random-effects Poisson regression time-series models were used to compare the rate of collisions in intervention versus matched control districts. These models account for observations within areas being more similar than observations between areas, and observations closer in time being more similar than observations further apart in time (autocorrelation). Separate models were developed for each outcome: all collisions; bicycle; e-scooter; bicycle and e-scooter combined; car; and motorcycle (subdivided by power: 50 cc; 125 cc to>500 cc).

Models adjusted for season; secular time trend (year); and baseline rate of collisions prior to the introduction of the schemes. Other potential confounders, including the number of cars/vans registered per district, total vehicle miles and length of road network were considered, but were excluded due to collinearity with other variables and high levels of missing data. The COVID-19 period was included as an independent variable and as an interaction term with intervention group. This interaction term represents a proxy for preintroduction/postintroduction of the schemes, as all e-scooter schemes were introduced during the peak period of lockdowns and restrictions (online supplemental figure 1). The COVID-19 variable was coded as pre (pre-March 2020); during (March 2020 to June 2021, peak period of restrictions, and the period in which all e-scooter schemes were introduced); and post (July 2021 to June 2023). Subgroup analyses were performed by severity of collision (slight vs serious/fatal). Cluster robust SEs were used in all models. Stata V.18 (StataCorp, College Station, Texas, USA) was used for all data analysis.

Sensitivity analyses

Sensitivity analyses included: (1) Falsification tests replacing intervention districts with randomly-selected control districts, and by substituting a fake intervention time period (January to December 2018) for the real intervention time period. (2) Control-group sensitivity analyses were performed on the bicycle collisions model, with PSM controls selected based on baseline rate of bicycle collisions rather than baseline rate of all collisions. (3) COVID-19-period sensitivity was performed using a four-level variable, splitting the original post-COVID-19 period into two (post year 1: July 2021 to June 2022, post year 2: July 2022 to June 2023) to identify more recent changes in trends.

Role of the funder

The funder was not involved in the study design; the collection, analysis or interpretation of data; in the writing of the report; or in the decision to submit the paper for publication.

Patient and public involvement

Cyclists were involved in the initial discussion of the research proposal. As a secondary data-analysis study, patients and the public were not involved in the study conduct. Two cyclists read and gave feedback on the manuscript, and larger patient and public involvement groups will be involved in dissemination including via general public focus groups, and specifically targeting cycling clubs and networks.

Ethics

Ethical approval was not required as the study involved secondary analyses of publicly available data.

Results

The data set included 372 LA ONS districts, of which 41 were included as intervention districts and 41 matched control districts. The mean population was slightly higher in intervention and matched control districts compared with Great Britain overall. Between January 2015 and June 2023, the mean monthly rate of all police-recorded collisions per 100 000 population for Great Britain overall, intervention and matched control districts was 14.7 (SD 6.8), 14.0 (SD 5.4) and 13.5 (SD 6.1). For bicycle collisions, the respective figures were 2.0 (SD 2.1), 2.5 (SD 2.5) and 1.7 (SD 1.6). For e-scooter collisions, the respective figures were 0.04 (SD 0.21), 0.10 (SD 0.35) and 0.03 (SD 0.14). Collision severity patterns were equivalent between the groups and representative of Great Britain overall (table 1). Within intervention districts, the e-scooter trials started between July 2020 and June 2021 (median start month October 2020, IQR September 2020 to January 2021 (online supplemental table 1 and figure 1)).

Table 1. Police-recorded road collisions in Great Britain 2015–2023^*.

	Great Britain overall†	E-scooter intervention districts	Matched control districts‡
Number of local authority ONS districts	372	41	41
Mean population per district (SD)	177 923.4 (122 113.5)	194 329.3 (102 550.2)	189 775.2 (100 298.9)
Mean number of cars/vans available (SD)	89 570.6 (54 913.4) 32/372 missing data	92 315.5 (42 432.2) 0/41 missing data	95 165.22 (52 972.3) 4/41 missing data
Total collisions	917 264	110 532	102 313
Fatal	12 161 (1.3%)	1249 (1.1%)	1533 (1.5%)
Serious	166 847 (18.2%)	19 315 (17.5%)	18 515 (18.1%)
Slight	738 299 (80.5%)	89 968 (81.4%)	82 245 (80.4%)
Mean number of collisions per district per year (SD)	318.6 (272.8)	336.9 (215.2)	311.9 (214.2)
By vehicle type§: car	271.2 (226.4)	290.6 (181.6)	272.7 (187.7)
Bicycle	47.8 (59.1)	62.6 (59.0)	38.8 (28.4)
Motorcycle	48.2 (57.9)	42.2 (28.8)	44.1 (40.2)
Van	31.5 (26.5)	29.3 (19.1)	31.2 (25.1)
Taxi	11.4 (25.3)	13.2 (19.4)	7.9 (9.8)
HGV	14.1 (12.3)	13.2 (10.1)	15.3 (13.2)
E-scooter	1.1 (4.7)	2.6 (9.7)	0.7 (2.2)
Mean rate of all collisions per 100 000 population per month	14.7 (6.8)	14.0 (5.4)	13.5 (6.1)
Mean rate of bicycle collisions per 100 000 population per month	2.0 (2.1)	2.5 (2.5)	1.7 (1.6)
Mean rate of e-scooter collisions per 100 000 population per month	0.04 (0.21)	0.10 (0.35)	0.03 (0.14)

^*January 2015 to June 2023. 2023 data are provisional, all other data are final release.

^†LA ONS districts recorded in police data with population denominator data available.

^‡Propensity score matched on population size and baseline rate of collisions in January 2019.

^§Not mutually exclusive categories.

e-scooter, electric scooter; HGV, Heavy Goods Vehicle; LA, local authority; ONS, Office for National Statistics.

Total collisions

Prior to the introduction of the e-scooter schemes, intervention and control districts had equivalent rates of all police-recorded collisions. In control districts, collisions reduced during the peak COVID-19 period compared with prepandemic (incidence rate ratio (IRR) 0.80, 95% CI 0.77 to 0.83, p<0.001), but following the end of restrictions, returned to prepandemic rates (IRR 1.04, 95% CI 0.98 to 1.11, p=0.20). During the intervention period, there was no evidence of a difference in the rate of total road collisions in intervention versus control areas (IRR 0.98, 95% CI 0.92 to 1.04, p=0.45 during peak COVID-19, IRR 0.96, 95% CI 0.89 to 1.04, p=0.30 post-COVID-19). There was evidence of seasonal fluctuations (higher collision rates in summer and autumn), and a slight overall trend of decreasing rates of collisions per year (IRR 0.96, 95% CI 0.95 to 0.96, p<0.001) (table 2, figure 1).

Table 2. All police-recorded road collisions in Great Britain 2015–2023^*.

	IRR	95% CI	P value
Comparability of groups preintervention
Matched control district	1.00	–	–
Intervention district	1.03	0.92 to 1.15	0.58
COVID-19 period (in controls)
Pre (January 2015 to February 2020)	1.00	–	–
During (March 2020 to June 2021)	0.80	0.77 to 0.83	<0.001
Post (July 2021 to June 2023)	1.04	0.98 to 1.11	0.20
Main intervention effect†
Intervention areas x March 2020 to June 2021	0.98	0.92 to 1.04	0.45
Intervention areas x July 2021 to June 2023	0.96	0.89 to 1.04	0.30
Time (calendar year)	0.96	0.95 to 0.96	<0.001
Season
Winter	1.00	–	–
Spring	0.98	0.97 to 1.00	0.03
Summer	1.06	1.04 to 1.08	<0.001
Autumn	1.12	1.11 to 1.14	<0.001

^*January 2015 to June 2023. 2023 data are provisional, all other data are final release.

^†Measured by interaction between intervention group and COVID-19 period (proxy for active intervention period).

Figure 1. All police-recorded road collisions 2015–2023. e-scooter, electric scooter.

Bicycle collisions

Prior to the introduction of the e-scooter schemes, intervention districts had a higher rate of bicycle collisions compared with control districts (IRR 1.48, 95% CI 1.10 to 1.99, p=0.01). In control districts, there was no evidence that the rate of bicycle collisions changed during the pandemic. However, in intervention districts, there was evidence for a relative reduction in bicycle collisions following the introduction of the e-scooter schemes (IRR 0.78, 95% CI 0.68 to 0.89, p<0.001 during peak COVID-19). This remained reduced over time (IRR 0.87, 95% CI 0.77 to 0.99, p=0.04 in the post-COVID-19 period). Again, there was evidence of seasonal fluctuations and a slight declining trend in the number of collisions per year (IRR 0.97, 95% CI 0.96 to 0.98, p<0·001) (table 3, figure 2).

Table 3. Police-recorded bicycle collisions in Great Britain 2015–2023^*.

	IRR	95% CI	P value
Comparability of groups preintervention
Matched control district	1.00	–	–
Intervention district	1.48	1.10 to 1.99	0.01
COVID-19 period (in controls)
Pre (January 2015 to February 2020)	1.00	–	–
During (March 2020 to June 2021)	1.07	0.97 to 1.17	0.17
Post (July 2021 to June 2023)	1.04	0.95 to 1.14	0.39
Main intervention effect†
Intervention areas x March 2020 to June 2021	0.78	0.68 to 0.89	<0.001
Intervention areas x July 2021 to June 2023	0.87	0.77 to 0.99	0.04
Time (calendar year)	0.97	0.96 to 0.98	<0.001
Season
Winter	1.00	–	–
Spring	1.20	1.16 to 1.25	<0.001
Summer	1.40	1.33 to 1.48	<0.001
Autumn	1.38	1.33 to 1.43	<0.001

^*January 2015 to June 2023. 2023 data are provisional, all other data are final release.

^†Measured by interaction between intervention group and COVID-19 period (proxy for active intervention period).

Figure 2. All police-recorded bicycle collisions 2015–2023. e-scooter, electric scooter.

E-scooter collisions

There was a significant increase in the rate of e-scooter collisions following the onset of the e-scooter schemes (IRR 195.0, 95% CI 39.3 to 965.6, p<0.001 during peak COVID, IRR 212.0, 95% CI 39.6 to 1134.5, p<0.001). As expected, intervention districts had far higher collision rates compared with control districts, and this differential increased over time. However, there was no statistical evidence for this difference, likely due to the very small total numbers of police-recorded e-scooter collisions. In contrast to other vehicle types, the overall time trend was for increasing rates of e-scooter collisions per year (IRR 1.18, 95% CI 1.06 to 1.32, p=0.002) (online supplemental table 2).

Other vehicles

As cars accounted for the majority of collisions, findings for cars alone were equivalent to the total collisions model. Motorcycles were similar, except that for larger motorcycles, collision rates increased postpandemic (IRR 1.32, 95% CI 1.17 to 1.49, p<0.001). The model using a combined outcome of bicycle plus e-scooter collisions was comparable to the bicycles-only model (with effects slightly attenuating over time), suggesting that the relative reduction of cycle collisions did not come at the cost of increased rate of e-scooter collisions (online supplemental table 2) as it was possible that an observed reduction in cycle collisions could be due to users switching from bicycles to e-scooters.

Subgroup analysis by severity

For all serious/fatal collisions, there was around a 10% sustained reduction in rates in intervention areas following the onset of the e-scooter schemes. For serious/fatal bicycle collisions, there was a 20% sustained reduction in rates following the onset of the e-scooter schemes in both the COVID-19 and post-COVID-19 periods (IRRs 0.78 and 0.80, respectively). In control areas, there was evidence of an increase in serious/fatal bicycle collisions during the peak pandemic period that appeared to return to pre-COVID-19 levels after the end of restrictions (online supplemental table 3, figures 2 and 3).

Sensitivity analysis

Falsification tests with (1) randomly-selected control districts replacing intervention districts and (2) a fake intervention time period reassuringly showed no evidence of any ‘intervention’ effect (online supplemental table 4a,b). Other sensitivity analyses were consistent with main findings (online supplemental table 4c–f).

Discussion

The introduction of e-scooter schemes was associated with, on average, around 20% relative reduction in bicycle collisions in intervention compared with matched control districts. This reduction attenuated slightly over time, but was sustained beyond any temporary effects of the COVID-19 pandemic and was also seen by severity of collisions. These findings were robust to falsification and sensitivity analyses and are consistent with the SiN hypothesis, that roads become safer for bicycles when numbers of non-car road users increase. The use of a natural experiment design provides stronger evidence than prior research.

As e-scooter schemes were introduced shortly following the start of the COVID-19 pandemic, the timeline of the latter represented a proxy for the former. It is possible that the observed reduction in bicycle collision rates in intervention districts was a differential effect of the pandemic in intervention and control districts, rather than an effect of the e-scooter schemes themselves. However, this is the less likely explanation given that (1) there was variation in start dates of the schemes, not all were introduced at the outset of the pandemic, and (2) the apparent protective effect was sustained beyond the pandemic, when traffic patterns otherwise returned to normal. (See online supplemental figure 4 for hypothetical visualisation of these differences.)

The decline in collisions during the peak pandemic period was no doubt due to lockdowns and travel restrictions. The increase in motorcycle collisions postpandemic might reflect an increase in riders/new riders on the road, as some people switched from public transport and cars to motorcycles following the pandemic.^{12 13} The increase in serious/fatal bicycle collisions during peak COVID-19 in control areas could reflect the increase in cycling, especially among more inexperienced cyclists, that occurred during the pandemic.^14–16

The evidence for an increase in e-scooter collisions over time in intervention districts is also as expected, given the policies to promote their use would have increased numbers of users. This finding is consistent with other evidence on emergency department presentations with e-scooter injuries.¹⁷ Department for Transport analysis indicates a pattern in total e-scooter casualties significantly increasing from 2020, followed by more stable levels from 2021 to 2023.¹⁸ Use of private e-scooters on public roads is currently illegal in the UK, hence e-scooter collisions in non-trial areas are likely to be under-reported in police records. This ‘contamination’ of exposure in the control areas would, if anything, lead to an underestimation of the difference in risk between trial and control areas.

The relative reduction in all severe/fatal road collisions in intervention districts following the start of the trials suggests protective benefits not just to cyclists, but also to other road users. This should be verified in other data sets. As some police forces have changed their method of recording collision severity over this period, ideally future work will include adjustment for this.¹⁹

This study used a natural experiment design, which minimises some of the problems with reverse causation and confounding faced by cross-sectional studies, which make up the majority of work on this topic.² Sensitivity analyses and falsification tests indicated the main findings were robust. A strength of the routinely collected police data is that it is generalisable across the UK, and can minimise measurement bias. A limitation is that it only covers a small proportion of total collisions, and there will be vastly higher numbers of collisions (particularly less severe and single-vehicle collisions) that are not reported to police and so not captured in this data set. A recent report comparing hospital versus police data on road collisions found that 26% of hospital records on serious bicycle injuries and 33% on serious e-scooter injuries were recorded in police data.²⁰ Another report found that<10% of casualties from e-scooter collisions were recorded in police data.²¹ Future work should triangulate different data sources to improve capture of collisions (and include analysis of injuries resulting from collisions). However, our hypothesis is specifically targeted on collisions as it argues that the behaviour of motor vehicle drivers may be altered; other causes of collisions (such as potholes) should not be altered by this intervention. Differences in police recording practices (and data accuracy/completeness) between districts mean that at an individual-district level, data are not perfectly comparable—however, this is unlikely to affect overall comparisons.

Exposure measurement had some misclassification, as e-scooter trials tended to take place in certain areas within a district, rather than across the whole district. It was not possible in the scope of this project to achieve more finely-grained exposure areas. In theory, this misclassification might weaken any real effects, so the true effects may be stronger. However, drivers exposed to e-scooter areas will likely also drive within the rest of their LA, meaning that any benefit of an e-scooter scheme could apply to a wider geographical area if it changes drivers’ behaviour. A benefit of conducting the analyses at district level is that there is less chance of spill-over (‘contamination’) between intervention and control areas, which would have been likely at smaller aggregation.

Intervention districts were different from control districts before the intervention and were likely targeted by e-scooter companies specifically for having features expected to be associated with their popularity (eg, size, population age and density, traffic infrastructure). Although we have tried to minimise confounding, and used sensitivity analyses to confirm that results were the same with different control groups, there will be residual confounding and future work should adjust for other factors where possible. In particular, we did not have data on traffic volume and the denominator of journeys (eg, number of trips/miles travelled). However, our findings are consistent with those from other studies that did include estimates of numbers of cycle journeys.^{2 22–27} The number of privately owned e-scooters in use is also not known, and as there are likely to be many in both intervention and control districts, this contamination may dilute effect estimates. Further research should include evaluation of other impacts of e-scooter schemes, including casualties and injury severity, collisions with pedestrians, and qualitative work to understand perceptions of e-scooters by other road users. Over the time period of the study, it was assumed that other modes of transport have not systematically changed between intervention and control areas; as it is possible they could have changed over time, this too should be verified in future work.

These results, if verified, could have important implications for policy and practice. Policies to encourage active-mode and micromobility transport options should be supported if they can be demonstrated to improve safety as well as reducing injuries. This could lead to public health and environmental benefits from increased physical activity (improving cardiorespiratory health, obesity, life expectancy) and reducing carbon emissions.

Conclusions

The introduction of e-scooters appears to have resulted in reduced bicycle collisions, reducing injuries and promoting safer environments. If verified in further work, this could give public health and environmental benefits if more people choose active-mode over car transport.

Data availability statement

Data are available upon reasonable request.