Introducing stats19

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  out.width = "100%"
)

Introduction

stats19 enables access to and processing of Great Britain's official road traffic casualty database, STATS19. A description of variables in the database can be found in a document provided by the UK's Department for Transport (DfT). The datasets are collectively called STATS19 after the form used to report them, which can be found here. This vignette focuses on how to use the stats19 package to work with STATS19 data.

Note: The Department for Transport refers to "accidents", but "crashes" is a more appropriate term, as emphasised in the "crash not accident" arguments of road safety advocacy groups such as RoadPeace. We use the term "accidents" only in reference to nomenclature within the data as provided.

The development version is hosted on GitHub and can be installed and loaded as follows:

# from CRAN
install.packages("stats19")
# you can install the latest development (discoraged) using:
remotes::install_github("ITSLeeds/stats19")
library(stats19)

Functions

The easiest way to get STATS19 data is with get_stats19(). This function takes 2 main arguments, year and type. The year can be any year between 1979 and 2017. The type can be one of accidents, casualties and vehicles, described below. get_stats19() performs 3 jobs, corresponding to three main types of functions:

Multiple functions (read_* and format_*) are needed for each step because of the structure of STATS19 data, which are divided into 3 tables, as described in an article that analyses 2005-2015 STATS19 data with reproducible R code by Christoph Freier (2018):

  1. "accident circumstances, with details about location, severity, weather, etc.;
  2. casualties, referencing knowledge about the victims; and
  3. vehicles, which contains more information about the vehicle type and manoeuvres, as well the some information about the driver."

The data can be downloaded for many years. Datasets since 1979 are broadly consistent, meaning that STATS19 data represents a very rich historic geographic, as stated in the DfT's road casualties report in 2017:

The current set of definitions and detail of information goes back to 1979, providing a long period for comparison.

Download STATS19 data

stats19 enables download of raw STATS19 data with dl_* functions. The following code chunk, for example, downloads and unzips a .zip file containing STATS19 data from 2017:

dl_stats19(year = 2017, type = "Accidents", ask = FALSE)

Note that in the previous command, ask = FALSE, meaning you will not be asked. By default you are asked to confirm, before downloading large files. Currently, these files are downloaded to a default location of tempdir which is a platform independent "safe" but temporary location to download the data in. Once downloaded, they are unzipped under original DfT file names. The dl_stats19() function prints out the location and final file name(s) of unzipped files(s) as shown above.

dl_stats19() takes three parameters. Supplying a file_name is interpreted to mean that the user is aware of what to download and the other two parameters will be ignored. You can also use years and type to "search" through the file names, which are stored in a lazy-loaded dataset called stats19::file_names.

You can find out the names of files that can be downloaded with names(stats19::file_names), an example of which is shown below:

stats19::file_names$dftRoadSafetyData_Vehicles_2017.zip

To see how file_names was created, see ?file_names. Data files from other years can be selected interactively. Just providing a year, for example, presents the user with multiple options (from file_names), illustrated below:

dl_stats19(year = 2017)
Multiple matches. Which do you want to download?

1: dftRoadSafetyData_Vehicles_2017.zip
2: dftRoadSafetyData_Casualties_2017.zip
3: dftRoadSafetyData_Accidents_2017.zip

Selection: 
Enter an item from the menu, or 0 to exit

When R is running interactively, you can select which of the 3 matching files to download: those relating to vehicles, casualties or accidents in 2017. You can also search for file names, as illustrated below:

d17 = "dftRoadSafetyData_Accidents_2017"
dl_stats19(file_name = paste0(d17, ".zip"))

Note that this is the same file (containing the 'accidents' table from 2017) downloaded previously with the command dl_stats19(year = 2017, type = "Accidents"). If the file is already present, dl_stats19() will skip the download process.

Read STATS19 data

In a similar approach to the download section before, we can read files downloaded using a data_dir location of the file and the filename to read. The code below will download the dftRoadSafetyData_Accidents_2017.zip file from the DfT servers and read its content. Files are saved by default in tempdir(), but this can be overridden to ensure permanent storage in a user-defined location.

d17 = "dftRoadSafetyData_Accidents_2017"
dl_stats19(file_name = paste0(d17, ".zip"))
crashes_2017_raw = read_accidents(data_dir = tempdir(),
                                  year = 2017,
                                  filename = "Acc.csv")

stats19 imports data with readr::read_csv() which results in a 'tibble' object: a data frame with more user-friendly printing and a few other features.

class(crashes_2017_raw)
dim(crashes_2017_raw)

There are three read_*() functions, corresponding to the three different classes of data provided by the DfT: 1. read_accidents() 2. read_casualties() 3. read_vehicles()

In all cases, a default parameter read_*(format = TRUE) returns the data in formatted form, as described above. Data can also be imported in the form directly provided by the DfT by passing format = FALSE, and then subsequently formatted with additional format_*() functions, as described in a final section of this vignette. Each of these read_*() functions is now described in more detail.

Crash data

After raw data files have been downloaded as described in the previous section, they can then be read-in as follows:

crashes_2017_raw = read_accidents(year = 2017, format = FALSE)
crashes_2017 = format_accidents(crashes_2017_raw)
nrow(crashes_2017_raw)
ncol(crashes_2017_raw)
nrow(crashes_2017)
ncol(crashes_2017)

What just happened? We read-in data on all road crashes recorded by the police in 2017 across Great Britain. The dataset contains r ncol(crashes_2017_raw) columns (variables) for r format(nrow(crashes_2017_raw), big.mark = ",") crashes.

This work was done by read_accidents(format = FALSE), which imported the "raw" STATS19 data without cleaning messy column names or re-categorising the outputs. format_accidents() function automates the process of matching column names with variable names and labels in a .xls file provided by the DfT. This means crashes_2017 is much more usable than crashes_2017_raw, as shown below, which shows some key variables in the messy and clean datasets:

crashes_2017_raw[c(7, 18, 23, 25)]
crashes_2017[c(7, 18, 23, 25)]

By default, format = TRUE, meaning that the two stages of read_accidents(format = FALSE) and format_accidents() yield the same result as read_accidents(format = TRUE). For the full list of columns, run names(crashes_2017).

# commented out as confusing...
key_patt = "severity|speed|light|human"
key_vars = grep(key_patt, x = names(crashes_2017_raw), ignore.case = TRUE)
random_n = sample(x = nrow(crashes_2017_raw), size = 3)
crashes_2017_raw[random_n, key_vars]
crashes_2017[random_n, key_vars]

Note: As indicated above, the term "accidents" is only used as directly provided by the DfT; "crashes" is a more appropriate term, hence we call our resultant datasets crashes_*.

Format STATS19 data

It is also possible to import the "raw" data as provided by the DfT. A .xls file provided by the DfT defines the column names for the datasets provided. The packaged datasets stats19_variables and stats19_schema provide summary information about the contents of this data guide. These contain the full variable names in the guide (stats19_variables) and a complete look up table relating integer values to the .csv files provided by the DfT and their labels (stats19_schema). The first rows of each dataset are shown below:

stats19_variables
stats19_schema

The code that generated these small datasets can be found in their help pages (accessed with ?stats19_variables and ?stats19_schema respectively). stats19_schema is used internally to automate the process of formatting the downloaded .csv files. Column names are formatted by the function format_column_names(), as illustrated below:

format_column_names(stats19_variables$variable[1:3])

Previous approaches to data formatting STATS19 data involved hard-coding results. This more automated approach to data cleaning is more consistent and fail-safe. The three functions: format_accidents(), format_vehicles() and format_casualties() do the data formatting on the respective data frames, as illustrated below:

crashes_2017 = format_accidents(crashes_2017_raw)

# vehicle data for 2017
dl_stats19(year = 2017, type = "vehicles", ask = FALSE)
vehicles_2017_raw = read_vehicles(year = 2017)
vehicles_2017 = format_vehicles(vehicles_2017_raw)

# casualties data for 2017
dl_stats19(year = 2017, type = "casualties", ask = FALSE)
casualties_2017 = read_casualties(year = 2017)

The package automates this two-step read_* and format_* process by defaulting in all cases to data_year = read_*(year, format = TRUE). read_* functions return, by default, formatted data. The two-step process may nevertheless be important for reference to the official nomenclature and values as provided by the DfT.

A summary of the outputs for each of the three tables is shown below.

summarise_stats19 = function(x) {
  data.frame(row.names = 1:length(x),
    name = substr(names(x), 1, 19),
    class = sapply(x, class),
    n_unique = sapply(x, function(v) length(unique(v))),
    first_label = sapply(x, function(v) substr(unique(v)[1], 1, 9)),
    second_label = sapply(x, function(v) substr(unique(v)[2], 1, 9))
  )
}
knitr::kable(summarise_stats19(crashes_2017), 
             caption = "Summary of formatted crash data.")
knitr::kable(summarise_stats19(vehicles_2017), 
             caption = "Summary of formatted vehicles data.")
knitr::kable(summarise_stats19(casualties_2017), 
             caption = "Summary of formatted casualty data.")

For testing and other purposes, a sample from the accidents table is provided in the package. A few columns from the two-row sample is shown below:

key_patt = "severity|speed|light|human"
key_vars = grep(key_patt, x = names(stats19::accidents_sample_raw), ignore.case = TRUE)
knitr::kable(stats19::accidents_sample_raw[, key_vars])

Casualties data

As with crashes_2017, casualty data for 2017 can be downloaded, read-in and formatted as follows:

dl_stats19(year = 2017, type = "casualties", ask = FALSE)
casualties_2017 = read_casualties(year = 2017)
nrow(casualties_2017)
ncol(casualties_2017)

The results show that there were r format(nrow(casualties_2017), big.mark=",") casualties reported by the police in the STATS19 dataset in 2017, and r ncol(casualties_2017) columns (variables). Values for a sample of these columns are shown below:

casualties_2017[c(4, 5, 6, 14)]

The full list of column names in the casualties dataset is:

names(casualties_2017)

Vehicles data

Data for vehicles involved in crashes in 2017 can be downloaded, read-in and formatted as follows:

dl_stats19(year = 2017, type = "vehicles", ask = FALSE)
vehicles_2017 = read_vehicles(year = 2017)
nrow(vehicles_2017)
ncol(vehicles_2017)

The results show that there were r format(nrow(vehicles_2017), big.mark=",") vehicles involved in crashes reported by the police in the STATS19 dataset in 2017, with r ncol(vehicles_2017) columns (variables). Values for a sample of these columns are shown below:

vehicles_2017[c(3, 14:16)]

The full list of column names in the vehicles dataset is:

names(vehicles_2017)
# old code to be up-dated
d14 = "Stats19_Data_2005-2014"
crashes_2005_2014 = read_accidents(data_dir = d14)
crashes_2005_2014_f = format_stats19_2005_2014_ac(crashes_2005_2014)
d15 = "RoadSafetyData_2015"
crashes_2015 = read_accidents(data_dir = d15, filename = "Accidents_2015.csv")
crashes_2015_f = format_stats19_2015_ac(crashes_2015)
d16 = "dftRoadSafety_Accidents_2016"
crashes_2016 = read_accidents(data_dir = d16, filename = "dftRoadSafety_Accidents_2016.csv")
crashes_2016_f = format_stats19_2016_ac(crashes_2016)
all_crashes = rbind(crashes_2015_f, crashes_2016_f, crashes_2017_f)
table(ac$Accident_Severity)

Creating geographic crash data

An important feature of STATS19 data is that the "accidents" table contains geographic coordinates. These are provided at ~10m resolution in the UK's official coordinate reference system (the Ordnance Survey National Grid, EPSG code 27700). stats19 converts the non-geographic tables created by format_accidents() into the geographic data form of the sf package with the function format_sf() as follows:

crashes_sf = format_sf(crashes_2017)

The note arises because NA values are not permitted in sf coordinates, and so rows containing no coordinates are automatically removed. Having the data in a standard geographic form allows various geographic operations to be performed on it. Spatial operations, such as spatial subsetting and spatial aggregation, can be performed, to show the relationship between STATS19 data and other geographic objects, such as roads, schools and administrative zones.

An example of an administrative zone dataset of relevance to STATS19 data is the boundaries of police forces in England, which is provided in the packaged dataset police_boundaries. The following code chunk demonstrates the kind of spatial operations that can be performed on geographic STATS19 data, by counting and plotting the number of fatalities per police force:

library(sf)
library(dplyr)
crashes_sf %>% 
  filter(accident_severity == "Fatal") %>% 
  select(n_fatalities = accident_index) %>% 
  aggregate(by = police_boundaries, FUN = length) %>% 
  plot()

Of course, one should not draw conclusions from such analyses without care. In this case, denominators are needed to infer anything about road safety in any of the police regions. After suitable denominators have been included, performance metrics such as 'health risk' (fatalities per 100,000 people), 'traffic risk' (fatalities per billion km, f/bkm) and 'exposure risk' (fatalities per million hours, f/mh) can be calculated [@feleke_comparative_2018; @elvik_handbook_2009].

The following code chunk, for example, returns all crashes within the jurisdiction of West Yorkshire Police:

west_yorkshire =
  police_boundaries[police_boundaries$pfa16nm == "West Yorkshire", ]
crashes_wy = crashes_sf[west_yorkshire, ]
nrow(crashes_sf)
nrow(crashes_wy)

This subsetting has selected the r format(nrow(crashes_wy), big.mark = ",") crashes which occurred in West Yorkshire.

Joining tables

The three main tables we have just read-in can be joined by shared key variables. This is demonstrated in the code chunk below, which subsets all casualties that took place in West Yorkshire, and counts the number of casualties by severity for each crash:

library(tidyr)
library(dplyr)
sel = casualties_2017$accident_index %in% crashes_wy$accident_index
casualties_wy = casualties_2017[sel, ]
cas_types = casualties_wy %>% 
  select(accident_index, casualty_type) %>% 
  group_by(accident_index) %>% 
  summarise(
    Total = n(),
    walking = sum(casualty_type == "Pedestrian"),
    cycling = sum(casualty_type == "Cyclist"),
    passenger = sum(casualty_type == "Car occupant")
    ) 
cj = left_join(crashes_wy, cas_types)

What just happened? We found the subset of casualties that took place in West Yorkshire with reference to the accident_index variable. Then we used the dplyr function summarise(), to find the number of people who were in a car, cycling, and walking when they were injured. This new casualty dataset is joined onto the crashes_wy dataset. The result is a spatial (sf) data frame of crashes in West Yorkshire, with columns counting how many road users of different types were hurt. The joined data has additional variables:

base::setdiff(names(cj), names(crashes_wy))

As a simple spatial plot, we can map all the crashes that have happened in West Yorkshire in 2017, with the colour related to the total number of people hurt in each crash. Placing this plot next to a map of West Yorkshire provides context:

plot(
  cj[cj$cycling > 0, "speed_limit", ],
  cex = cj$Total[cj$cycling > 0] / 3,
  main = "Speed limit (cycling)"
  )
plot(
  cj[cj$passenger > 0, "speed_limit", ],
  cex = cj$Total[cj$passenger > 0] / 3,
  main = "Speed limit (passenger)"
  )

The spatial distribution of crashes in West Yorkshire clearly relates to the region's geography. Car crashes tend to happen on fast roads, including busy Motorway roads, displayed in yellow above. Cycling is as an urban activity, and the most bike crashes can be found in near Leeds city centre, which has a comparatively high level of cycling (compared with the low baseline of 3%). This can be seen by comparing the previous map with an overview of the area, from an academic paper on the social, spatial and temporal distribution of bike crashes [@lovelace_who_2016]:

knitr::include_graphics("wy-overview.jpg")

In addition to the Total number of people hurt/killed, cj contains a column for each type of casualty (cyclist, car occupant, etc.), and a number corresponding to the number of each type hurt in each crash. It also contains the geometry column from crashes_sf. In other words, joins allow the casualties and vehicles tables to be geo-referenced. We can then explore the spatial distribution of different casualty types. The following figure, for example, shows the spatial distribution of pedestrians and car passengers hurt in car crashes across West Yorkshire in 2017:

library(ggplot2)
crashes_types = cj %>% 
  filter(accident_severity != "Slight") %>% 
  mutate(type = case_when(
    walking > 0 ~ "Walking",
    cycling > 0 ~ "Cycling",
    passenger > 0 ~ "Passenger",
    TRUE ~ "Other"
  ))
ggplot(crashes_types, aes(size = Total, colour = speed_limit)) +
  geom_sf(show.legend = "point", alpha = 0.3) +
  facet_grid(vars(type), vars(accident_severity)) +
  scale_size(
    breaks = c(1:3, 12),
    labels = c(1:2, "3+", 12)
    ) +
  scale_color_gradientn(colours = c("blue", "yellow", "red")) +
  theme(axis.text = element_blank(), axis.ticks = element_blank())

It is clear that different types of road users tend to get hurt in different places. Car occupant casualties (labelled 'passengers' in the map above), for example, are comparatively common on the outskirts of cities such as Leeds, where speed limits tend to be higher and where there are comparatively higher volumes of motor traffic. Casualties to people on foot tend to happen in the city centres. That is not to say that cities centres are more dangerous per unit distance (typically casualties per billion kilometres, bkm, is the unit used) walked: there is more walking in city centres (you need a denominator to estimate risk).

To drill down further, we can find the spatial distribution of all pedestrian casualties, broken-down by seriousness of casualty, and light conditions. This can be done with tidyvers functions follows:

table(cj$light_conditions)
cj %>% 
  filter(walking > 0) %>% 
  mutate(light = case_when(
    light_conditions == "Daylight" ~ "Daylight",
    light_conditions == "Darkness - lights lit" ~ "Lit",
    TRUE ~ "Other/Unlit"
  )) %>% 
  ggplot(aes(colour = speed_limit)) +
  geom_sf() +
  facet_grid(vars(light), vars(accident_severity)) +
  scale_color_continuous(low = "blue", high = "red") +
  theme(axis.text = element_blank(), axis.ticks = element_blank())

Time series analysis

We can also explore seasonal and daily trends in crashes by aggregating crashes by day of the year:

crashes_dates = cj %>% 
  st_set_geometry(NULL) %>% 
  group_by(date) %>% 
  summarise(
    walking = sum(walking),
    cycling = sum(cycling),
    passenger = sum(passenger)
    ) %>% 
  gather(mode, casualties, -date)
ggplot(crashes_dates, aes(date, casualties)) +
  geom_smooth(aes(colour = mode), method = "loess") +
  ylab("Casualties per day")

Different types of crashes also tend to happen at different times of day. This is illustrated in the plot below, which shows the times of day when people who were travelling by different modes were most commonly injured.

library(stringr)

crash_times = cj %>% 
  st_set_geometry(NULL) %>% 
  group_by(hour = as.numeric(str_sub(time, 1, 2))) %>% 
  summarise(
    walking = sum(walking),
    cycling = sum(cycling),
    passenger = sum(passenger)
    ) %>% 
  gather(mode, casualties, -hour)

ggplot(crash_times, aes(hour, casualties)) +
  geom_line(aes(colour = mode))

Note that bike crashes tend to have distinct morning and afternoon peaks, in-line with previous research [@lovelace_who_2016]. A disproportionate number of car crashes appear to happen in the afternoon.

Further work

There is much potential to extend the package beyond downloading, reading and formatting STATS19 data. The greatest potential is to provide functions that will help with analysis of STATS19 data, to help with road safety research. Much academic research has been done using the data, a few examples of which are highlighted below to demonstrate the wide potential for further work.

The broader point is that the stats19 package could help road safety research, by making open access data on road crashes more accessible to researchers worldwide. By easing the data download and cleaning stages of research, it could also encourage reproducible analysis in the field.

There is great potential to add value to and gain insight from the data by joining the datasets with open data, for example from the Consumer Data Research Centre (CDRC, which funded this research), OpenStreetMap and the UK's Ordnance Survey. If you have any suggestions on priorities for these future directions of (hopefully safe) travel, please get in touch on at github.com/ITSLeeds/stats19/issues.

References



Try the stats19 package in your browser

Any scripts or data that you put into this service are public.

stats19 documentation built on Feb. 15, 2019, 5:05 p.m.