As the name suggests, origin-destination (OD) data represents movement through geographic space, from an origin (O) to a destination (D). Sometimes also called ‘flow data,’ OD datasets contain details of trips between two geographic points or, more commonly, zones (which are often represented by a zone centroid). Most OD datasets refer to start and end locations with ‘ID’ columns containing character strings such as
zone1. These IDs refer to a geographic feature in a separate geographic dataset. Origin and destination locations are sometimes represented as geographic coordinates.
OD datasets typically contain multiple non geographic attributes. These usually include, at a minimum, the number of trips that take place from the origin to the destination over a given time period (e.g. a typical work day). Additional attributes can include breakdown by the mode(s) of transport used for the trips. Usually only a single mode is captured (trips made by a combination of cycle-train-walk modes are often counted only as ‘train’ trips). Additional disaggregations of overall counts may include trip counts at different time periods.
Many OD datasets omit information. If there is only one time period, then this resides in the metadata for the whole data set. There is rarely any information about the path taken between the start and end points. It is typically the job of the analyst to use a routing service (such as OSRM, Google Directions API, CycleStreets.net or OpenRouteService) or an assignment model (such as those contained in proprietary software such as SATURN and Visum) to identify likely routes with reference to shortest path algorithms or generalised cost minimisation algorithms (which account for monetary plus time and quality ‘costs’).
Despite the rather dull name, OD datasets are a vital part of the modern world: they underpin analysis and models that influence current and future transport systems. Historically, these models, and the OD datasets that drove them, were used to plan for car-dominated cities (Boyce and Williams 2015). Now that there is growing evidence of the negative impacts car domination, however, there is a strong argument for transport models being re-purposed. Origin-destination data can be part of the solution.
From a health perspective transport planning, supported by OD data and analysed primarily using proprietary software and opaque methods, has failed: roads are now the largest cause of death of young people worldwide, killing more than 1 million people each year (World Health Organization 2018). Even ignoring problems such as air pollution, obesity and climate change, it is clear that current transport systems are unsustainable. There are other reasons why transport data analysis and software are important (Lovelace and Ellison 2018).
The purpose of this vignette is to introduce OD data, an important component of many transport planning models, with examples based on data and functions from the stplanr package. The aim is to enable you to use OD data to inform more sustainable transport plans, for example by identifying ‘desire lines’ along which policies could cause a modal switch away from cars and towards lower energy modes such as walking, cycling, and public transport.
OD data can be accessed from a range of sources (we will see code that downloads many thousands of OD pairs later in this vignette). Some ‘data carpentry’ may be needed before the OD data is ready for analysis. This vignette does not cover cleaning OD data: we assume you know R and come with ‘tidy’ data (Wickham 2014), in which each row represents travel between an origin and a destination (typically zones represented by zone IDs), and each column represents an attribute such as number of trips or vehichle counts by mode or straight line distance.1
In simple terms OD data looks like this:
library(stplanr) library(dplyr) <- stplanr::od_data_sample %>% od select(-matches("rail|name|moto|car|tax|home|la_")) %>% top_n(n = 14, wt = all) class(od) #>  "tbl_df" "tbl" "data.frame" od#> # A tibble: 14 x 8 #> geo_code1 geo_code2 all train bus bicycle foot other #> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> #> 1 E02002361 E02002361 109 0 4 2 59 0 #> 2 E02002361 E02002393 94 0 17 0 10 1 #> 3 E02002363 E02002363 183 2 13 5 101 0 #> 4 E02002363 E02002384 92 1 13 2 21 0 #> 5 E02002363 E02002393 156 0 19 12 15 0 #> 6 E02002367 E02002393 88 0 17 4 16 1 #> 7 E02002371 E02002363 110 1 18 2 47 0 #> 8 E02002371 E02002371 220 1 28 1 116 2 #> 9 E02002371 E02002393 165 0 18 10 49 0 #> 10 E02002377 E02002377 129 0 11 1 79 0 #> 11 E02002377 E02002393 93 0 20 0 37 0 #> 12 E02002382 E02002393 94 0 9 1 44 3 #> 13 E02002384 E02002384 166 2 13 2 116 0 #> 14 E02002393 E02002393 265 4 15 0 185 6
Like all data, the object
od, created in the preceding code chunk, comes from a specific context: the 2011 UK Census questions:
od is a data frame containing aggregated answers to these questions (see
?pct::get_od() for details). It is implicitly geographic: the first two columns refer to geographic entities but do not contain coordinates themselves (OD coordinates are covered below). Other columns contain attributes associated with each OD pair, typically counting how many people travel by mode of transport. OD data can be represented in a number of ways, as outlined in the next sections.
The most useful way of representing OD data is the ‘long’ data frame format described above. This is increasingly the format used by official statistical agencies, including the UK’s Office for National Statistics (ONS), who provide origin destination data as a
.csv file. Typically, the first column is the zone code of origin and the second column is the zone code of the destination, as is the case with the object
od. Subsequent columns contain attributes such as
all, meaning trips by all modes, as illustrated below (we will see a matrix representation of this subset of the data in the next section):
1:3] od[#> # A tibble: 14 x 3 #> geo_code1 geo_code2 all #> <chr> <chr> <dbl> #> 1 E02002361 E02002361 109 #> 2 E02002361 E02002393 94 #> 3 E02002363 E02002363 183 #> 4 E02002363 E02002384 92 #> 5 E02002363 E02002393 156 #> 6 E02002367 E02002393 88 #> 7 E02002371 E02002363 110 #> 8 E02002371 E02002371 220 #> 9 E02002371 E02002393 165 #> 10 E02002377 E02002377 129 #> 11 E02002377 E02002393 93 #> 12 E02002382 E02002393 94 #> 13 E02002384 E02002384 166 #> 14 E02002393 E02002393 265
geo_code1 refers to the origin,
geo_code2 refers to the destination.
Additional columns can represent addition attributes, such as number of trips by time, mode of travel, type of person, or trip purpose. The
od dataset contains column names representing mode of travel (train, bus, bicycle etc), as can be seen with
names(od[-(1:2)]). These ‘mode’ columns contain integers in the example data, but contain characters, dates and other data types, taking advantage of the flexibility of data frames.
The ‘OD matrix’ representation of OD data represents each attribute column in the long form as a separate matrix. Instead of rows representing OD pairs, rows represent all travel from each origin to all destinations (represented as columns). The stplanr function
od_to_odmatrix() converts between the ‘long’ to the ‘matrix’ form on a per column basis, as illustrated below:
<- od_to_odmatrix(od[1:3]) od_matrix class(od_matrix) #>  "matrix" "array" od_matrix#> E02002361 E02002393 E02002363 E02002384 E02002371 E02002377 #> E02002361 109 94 NA NA NA NA #> E02002363 NA 156 183 92 NA NA #> E02002367 NA 88 NA NA NA NA #> E02002371 NA 165 110 NA 220 NA #> E02002377 NA 93 NA NA NA 129 #> E02002382 NA 94 NA NA NA NA #> E02002384 NA NA NA 166 NA NA #> E02002393 NA 265 NA NA NA NA
Note that row and column names are now zone codes. The cell in row 1 and column 2 (
od_matrix[1, 2]), for example, reports that there are 94 trips from zone
E02002361 to zone
E02002393. In the case above, no people travel between the majority of the OD pair combinations, as represented by the
NAs. OD matrices are a relatively rudimentary data structure that pre-date R’s
data.frame class. Typically, they only contained integer counts, providing small and simple datasets that could be used in 20th Century transport modelling software running on limited 20th Century hardware.
Although ‘OD matrix’ is still sometimes used informally to refer to any OD datadset, the long OD pair representation is recommended: OD matrices become unwieldy for large OD datasets, which are likely to be sparse, with many empty cells represented by NAs. Furthermore, to represent many attributes in matix format, multiple lists of OD matrices or ‘OD arrays’ must be created. This is demonstrated in the code chunk below, which represents travel between OD pairs by all modes and by bike:
lapply(c("all", "bicycle"), function(x) od_to_odmatrix(od[c("geo_code1", "geo_code2", x)])) #> [] #> E02002361 E02002393 E02002363 E02002384 E02002371 E02002377 #> E02002361 109 94 NA NA NA NA #> E02002363 NA 156 183 92 NA NA #> E02002367 NA 88 NA NA NA NA #> E02002371 NA 165 110 NA 220 NA #> E02002377 NA 93 NA NA NA 129 #> E02002382 NA 94 NA NA NA NA #> E02002384 NA NA NA 166 NA NA #> E02002393 NA 265 NA NA NA NA #> #> [] #> E02002361 E02002393 E02002363 E02002384 E02002371 E02002377 #> E02002361 2 0 NA NA NA NA #> E02002363 NA 12 5 2 NA NA #> E02002367 NA 4 NA NA NA NA #> E02002371 NA 10 2 NA 1 NA #> E02002377 NA 0 NA NA NA 1 #> E02002382 NA 1 NA NA NA NA #> E02002384 NA NA NA 2 NA NA #> E02002393 NA 0 NA NA NA NA
odmatrix_to_od() can converts OD matrices back into the more convenient long form:
odmatrix_to_od(od_matrix) #> orig dest flow #> 1 E02002361 E02002361 109 #> 9 E02002361 E02002393 94 #> 18 E02002363 E02002363 183 #> 26 E02002363 E02002384 92 #> 10 E02002363 E02002393 156 #> 11 E02002367 E02002393 88 #> 20 E02002371 E02002363 110 #> 36 E02002371 E02002371 220 #> 12 E02002371 E02002393 165 #> 45 E02002377 E02002377 129 #> 13 E02002377 E02002393 93 #> 14 E02002382 E02002393 94 #> 31 E02002384 E02002384 166 #> 16 E02002393 E02002393 265
A common, and sometimes problematic, feature of OD data is ‘intra-zonal flows.’ These are trips that start and end in the same zone. The proportion of travel that is intra-zonal depends largely on the size of the zones used. It is often useful to separate intra-zonal and inter-zonal flows at the outset, as demonstrated below:
<- od %>% filter(geo_code1 != geo_code2)) (od_inter #> # A tibble: 8 x 8 #> geo_code1 geo_code2 all train bus bicycle foot other #> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> #> 1 E02002361 E02002393 94 0 17 0 10 1 #> 2 E02002363 E02002384 92 1 13 2 21 0 #> 3 E02002363 E02002393 156 0 19 12 15 0 #> 4 E02002367 E02002393 88 0 17 4 16 1 #> 5 E02002371 E02002363 110 1 18 2 47 0 #> 6 E02002371 E02002393 165 0 18 10 49 0 #> 7 E02002377 E02002393 93 0 20 0 37 0 #> 8 E02002382 E02002393 94 0 9 1 44 3 <- od %>% filter(geo_code1 == geo_code2)) (od_intra #> # A tibble: 6 x 8 #> geo_code1 geo_code2 all train bus bicycle foot other #> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> #> 1 E02002361 E02002361 109 0 4 2 59 0 #> 2 E02002363 E02002363 183 2 13 5 101 0 #> 3 E02002371 E02002371 220 1 28 1 116 2 #> 4 E02002377 E02002377 129 0 11 1 79 0 #> 5 E02002384 E02002384 166 2 13 2 116 0 #> 6 E02002393 E02002393 265 4 15 0 185 6
Intra-zonal OD pairs represent short trips (up to the size of the zone within which the trips take place) so are sometimes ignored in OD data analyis. However, intra-zonal flows can be valuable, for example in measuring the amount of localised transport activity and as a sign of local economies.
Another subtly with some (symetric, where origins and destinations can be the same points) OD data is that oneway flows can hide the extent of bidirectional flows in plots and other types of analysis. This is illustrated below for a sample of the
<- od_data_sample[c(1, 2, 9), 1:6]) (od_min #> # A tibble: 3 x 6 #> geo_code1 geo_code2 all from_home light_rail train #> <chr> <chr> <dbl> <dbl> <dbl> <dbl> #> 1 E02002361 E02002361 109 0 0 0 #> 2 E02002361 E02002363 38 0 0 1 #> 3 E02002363 E02002361 30 0 0 0 <- od_oneway(od_min)) (od_oneway #> # A tibble: 2 x 6 #> geo_code1 geo_code2 all from_home light_rail train #> <chr> <chr> <dbl> <dbl> <dbl> <dbl> #> 1 E02002361 E02002361 109 0 0 0 #> 2 E02002361 E02002363 68 0 0 1
Note that in the second dataset there are only 2 rows instead of 3. The function
od_oneway() aggregates oneway lines to produce bidirectional flows. By default, it returns the sum of each numeric column for each bidirectional origin-destination pair.
The previous representations of OD data are all implicitly geographic: their coordinates are not contained in the data, but associated with another object that is geographic, typically a zone or a zone centroid. This is problematic, meaning that multiple objects or files are required to fully represent the same data. Desire line representations overcome this issue. They are geographic lines between origin and destination, with the same attributes as in the ‘long’ representation.
od2line() can convert long form OD data to desire lines. The second argument is a zone or a centroid dataset that contains ‘zone IDs’ that match the IDs in the first and second columns of the OD data, as illustrated below:
<- zones_sf z class(z) #>  "sf" "data.frame" <- od2line(flow = od_inter, zones = z) l #> Creating centroids representing desire line start and end points.
The preceding code chunk created a zones object called
z, the coordinates of which were used to convert the object
l, which are geographic desire lines. The desire line object is stored in as a geographic simple features object, which has the same number of rows as does the object
od and one more column:
class(l) #>  "sf" "data.frame" nrow(od) - nrow(l) #>  6 ncol(l) - ncol(od) #>  1
The new column is the geometry column, which can be plotted as follows:
By default, plotting
l shows the attributes for each line:
Because these lines have a coordinate reference system (CRS) inherited from the zones data, they can also be plotted on an interactive map, as follows (result only shown if webshot is installed):
library(leaflet) leaflet() %>% addTiles() %>% addPolygons(data = l)
Note that in some OD datasets there may be IDs that match no zone. We can simulate this situation by setting the third origin ID of
nomatch, a string that is not in the zones ID:
$geo_code2 <- "nomatch" odod2line(od, z) #> Creating centroids representing desire line start and end points. #> Error: 1 non matching IDs in the destination. ID on row 3 does not match any zone. #> The first offending id was nomatch
You should clean your OD data and ensure all ids in the first two columns match the ids in the first column of the zone data before running
The minimal example dataset we’ve been using so far is fine for demonstrating the key concepts of OD data. But for more advanced topic, and to get an idea of what is possible with OD data at a city level, it helps to have a larger dataset.
We will use an example dataset representing commuting in London, accessed as follows (note: these code chunks are not evaluated in the vignette because it starts by downloading 2.4 million rows and could take a few minutes to run). First, we can use the
pct package to download official data from the UK (note the addition of the % active column):
library(dplyr) # get nationwide OD data <- pct::get_od() od_all nrow(od_all) # > 2402201 $Active <- (od_all$bicycle + od_all$foot) / od_all$all * 100 od_all<- pct::get_centroids_ew() %>% sf::st_transform(4326) centroids_all nrow(centroids_all) # > 7201 <- pct::pct_regions %>% filter(region_name == "london") london <- centroids_all[london, ] centroids_london <- od_all %>% od_london filter(geo_code1 %in% centroids_london$msoa11cd) %>% filter(geo_code2 %in% centroids_london$msoa11cd) <- od_all[ od_london $geo_code1 %in% centroids_london$msoa11cd & od_all$geo_code2 %in% centroids_london$msoa11cd, od_all]
Now that we have the input OD data (in
od_london) and zones (population-weighted centroids in
cents_london in this case), can can convert them to desire lines:
<- od2line(od_london, centroids_london) desire_lines_london nrow(desire_lines_london) # > 352654
Even after filering flows to keep only those with origins and destinations in London, there are still more than 300k flows. That is a lot to plot. So we’ll further subset them, first so they only contain inter-zonal flows (which are actually lines, intra-zonal flows are lines with length 0, which are essentially points) and second to contain only flows containing above a threshold level of flows:
<- 20 min_trips_threshold <- desire_lines_london %>% filter(geo_code1 != geo_code2) desire_lines_inter <- desire_lines_london %>% filter(geo_code1 == geo_code2) desire_lines_intra <- desire_lines_inter %>% filter(all >= min_trips_threshold) desire_lines_top nrow(desire_lines_top) # > 28879
If we do any analysis on this dataset, it’s important to know how representative it is of all flows. A crude way to do this is to calculate the proportion of lines and trips that are covered in the dataset:
nrow(desire_lines_top) / nrow(desire_lines_london) # > 0.08189046 sum(desire_lines_top$all) / sum(desire_lines_london$all) # > 0.557343
This shows that only 8% of the lines contain more than half (55%) of the total number of trips.
Once you have an OD dataset of a size that can be plotted (20,000 desire lines is quick to plot on most computers) a logical next stage is to plot it, e.g. with
You may be disapointed by the result, which is more of a ‘hay stack’ plot than an intuitive illustration of flows across the city. To overcome this issue, you can set the aesthetics to emphasize with important flows, e.g. by line width in
sf’s plotting system:
<- desire_lines_top$all / mean(desire_lines_top$all) / 10 lwd $percent_dont_drive <- 100 - desire_lines_top$car_driver / desire_lines_top$all * 100 desire_lines_topplot(desire_lines_top["percent_dont_drive"], lwd = lwd, breaks = c(0, 50, 70, 80, 90, 95, 100))