19 January 2021

Violent crime in London: trends, trajectories and neighbourhoods

This report is a result of a project which aims to illustrate some of the ways in which police-recorded crime data could be combined with other sources to provide a deeper insight into the geographical and social distribution of violent crime. 

Generating evidence on the nature and distribution of violent crime, and what can be done about it, is a key problem currently facing police. While police-recorded data can offer important insights, it only captures part of the picture of violent crime. While we often reflect on what crimes are committed, we may think less about where or when they occur, or why. Combining police data with other sources offers us the potential to address some of these limitations, and maximise the utility of data collection and analysis strategies.

In Violent crime in London: trends, trajectories and neighbourhoods we set out to illustrate some of the ways combining different data sources might provide deeper insight into the geographical and social distribution of violent crime. In addition to merging datasets, we prioritised the use of robust analysis techniques. We used three different approaches – mapping, predictive models and trajectory models – to consider the patterning of police-recorded violent crime across London; and we used a novel machine learning technique to assess the potential impact of stop and search on crime.

Crime mapping provided the baseline analysis. Using crime data from London across five financial years (2013-2017), we aggregated annual counts of violent crime to the LSOA level (Lower-level Super Output Areas are a Census-based area level classification, with an average population size of around 1,500 people – there are 4,835 LSOAs in Greater London). These maps revealed the stark geographical clustering of violent crime, concentrated in a small number of LSOAs distributed across the capital.


What, though, predicts vulnerability to violent crime? To explore this question we combined the police recorded data with other sources: the 2011 Census, the 2015 Index of Multiple Deprivation, MPS stop and search data, and the MPS (now MOPAC) Public Attitudes Survey (PAS, 2007-2010). 

We defined high vulnerability to violent crime as being in the top quartile of LSOAs for violence in a given year. Using standard regression techniques, we found that greater deprivation in one year – across domains of income, health, housing, living environment and education – predicted a greater chance of being in the high vulnerability category the following year. These 'structural' neighbourhood characteristics seemed to be more predictive of violence than 'social' characteristics gleaned from the PAS (e.g. collective efficacy, the ability of people living in local areas to engage in informal social control).

Different areas are however likely to have different trajectories of crime. Group Based Trajectory Modelling (GBTM) is a statistical technique that identifies the number (and shape) of distinct crime trajectory patterns within a given dataset. We identified 17 groups of LSOAs, each with a distinct trajectory of violent crime between 2013 and 2017. Seven of these groups, representing just 17% of LSOAs, had a high and/or increasing level of violence. Most other areas, representing the bulk of the capital, were in trajectory groups where violence was low and saw little or no increase from 2013 to 2017. And when we looked to see what predicted membership of the seven high/increasing groups, we found that disadvantage was again, a key factor.

Finally, we looked at the potential effect of stop and search on violent crime. Merging in MPS stop and search data for 2016, we used a machine learning approach to create two groups of LSOAs. These two groups were notionally equivalent across a range of indicators, such as overall crime rates and deprivation, but in one, the use of stop and search was much higher than the other (equivalent to being in the top 10% of areas in terms of use of stop and search). We then looked to see whether LSOAs in the high stop and search areas had lower levels of violence in 2017.

Strikingly, however, violent crime in 2017 was higher in areas assigned to the high level stop and search category in 2016 than in otherwise equivalent areas assigned to the low level stop and search category. One explanation of this finding may be that stop and search acts as a 'leading indicator' for violence in areas on the cusp or in the early stages of an upswing. And even though we took overall crime into account, areas high in violence in 2017 are also likely to have been high in violence in 2016, attracting a higher level of stop and search that year.

Violent crime in London is therefore heavily clustered in a small number of areas that tend to be significantly more deprived than others. The increase in violent crime between 2013 to 2017 was similarly confined largely to a small number of areas, many of which were already relatively prone to violence. 

There are obvious similarities between our findings and work on crime hotspots. Like smaller hotspots, the distribution of violent crime over London, even at the larger LSOA level, seems to be relatively predictable. Violent crime concentrates in a relatively small number of areas, and this concentration is quite stable over time. 

The exploratory work we conducted may contain some important lessons for policing, in London and beyond. First, while deprivation does not determine levels of crime, there is a robust association. Resources should be targeted accordingly. As the economic fallout of Covid-19 bites, police and other services should keep a close eye on those areas most affected.

Second, the trajectory of crime in local areas seems to be an important consideration. Areas that have been low and stable in recent years seem likely to remain so, while identifying those on an upwards trajectory might, again, help target resource. To reiterate, we can be confident that deprivation will be a key factor distinguishing between areas on different trajectories.

Having the statistical tools and capabilities to do this type of analysis may therefore be an important capability for police. However, the GBTM was implemented in Stata, which is a proprietary statistical package not commonly available to police analysts; freeware packages such as R are also not often accessible. Leveraging the full potential of the data available may make obtaining the right tools an unavoidable necessity.

Third, our findings seem to underline that policing is unlikely to be the primary solution to violent crime, which is deeply embedded in poverty, deprivation and social exclusion. The analysis of stop and search described above is hardly conclusive. But it does chime with other research that has found limited evidence of stop and search having a deterrent effect on crime. Other policing tactics are of course available, but the link between poverty and violence is such that much wider, programmatic, interventions are needed.

The approaches we have outlined here should be amenable to police and other researchers operating anywhere in the UK. While relatively few places have a resource as robust as the PAS, the core datasets we used are duplicated in the data made available on police.uk and elsewhere. If the findings outlined above replicate to other contexts – and there is no reason to think they will not – then knowing more about the social and economic characteristics of areas is vital to understanding the distribution of violent crime and how it develops over time. Much further research is needed to fully unpick this relationship, of course, and perhaps most importantly turn understanding into operable policy interventions.

Police organisations will of course have access to much more detailed and fine-grained data, which could be used to increase the resolution of the analysis, evaluate specific policy interventions, and so on. Resources like the UCL Jill Dando Institute Research Laboratory (JDIRL), which we used when doing this research, may thus become increasingly important. The JDIRL is a secure computer facility that allows authorised access to sensitive datasets that would not otherwise be accessible. As analysis of complex datasets becomes increasingly important to policing, offering access to academics and other researchers should provide significant benefits in terms of the breadth, depth and utility of the outputs generated.


The full report was prepared for the College of Policing by Dr Alex Sutherland (Behavioural Insights Team); Professor Ian Brunton-Smith (University of Surrey); Oli Hutt and Professor Ben Bradford (University College London)

This report was prepared as part of the Vulnerability and Violent Crime Programme.  The Phase 2 intervention evaluations will be published in Spring 2021.   


If you would like to receive our regular College What Works email updates, contact us at whatworks@college.pnn.police.uk and ask to be added to our mailing list.