Identifying the appropriate spatial resolution for the analysis of crime pattern...

2.1 Crime at place, and the concentration of crime events

A substantial body of recent research has found that crime concentrates at ‘micro places’ and, in general, a small unit of analysis is the most appropriate for many quantitative environmental criminology studies [5, 8, 10–13]. These findings, that have emerged through a number of disparate research efforts, were synthesised in a recent systematic review [13] that found consistent support for this “Law of Crime Concentration at Place”, defined by [11] as follows:

“for a defined measure of crime at a specific microgeographic unit, the concentration of crime will fall within a narrow bandwidth of percentages for a defined cumulative proportion of crime” [11]

These findings align both with established environmental theories—that highlight the local nature of criminal events [14–16]—and with the observations of other spatial phenomena where aggregate analyses have been found to hide important lower-level patterns [1, 17]. Although care must be taken when analysing crime at the smallest spatial units (e.g. addresses) because the sparsity of some crime types relative to the number of observation points can artificially inflate the observed concentration [18], broadly most studies report support for the finding of crime concentrates at micro places, and at higher levels than at larger spatial scales [2, 4].

However, determining the appropriate spatial scale is non-trivial for at least three reasons. Firstly, spatial units could be addresses, street segments, or small grid cells, etc. [13], and although smaller units of analysis are typically better able to identify crime concentration [10], does this hold for all crime types (or the number of events in an area to be more general) and all areas? Secondly, high-resolution data that are required for studies at the level of micro-places, such as individual addresses, are harder to obtain than more aggregate data. Thirdly, the difficulty in acquiring fine scale data is further compounded by possible small number problems that occur when rare crime events are analysed at small spatial scales [5]. One could, therefore, ask the question: at what point does it become unnecessary to obtain finer scale data? In other words, are address-level data required for studies of all crime types/counts in all regions, or might more aggregate data be sufficient? What matters is which spatial resolution is appropriate for the analysis of a particular phenomena given the nature of the spatial pattern. Or, as [5] put it: “The question then is how small can you go before [the potential problems outweigh] the benefits of more homogeneous areas?” For some phenomena, large neighbourhoods might be an appropriate unit of analysis, whereas for others it might be necessary to use a much higher resolution geography.

To the best of our knowledge, only one paper has investigated a similar research question. [19] attempt to find the spatial unit at which open source geomasked crime data are sufficiently close to the true underlying point pattern (actual police data) to enable meaningful analysis. They use a similar process to that outlined here, albeit only applied to administrative boundaries. Importantly, they found that smaller units of geography (e.g., postal codes) exhibited significantly different spatial patterns when comparing geomasked to non-geomasked crime data. As such, when using geomasked crime data there is a lower bound to which a researcher can trust that the spatial patterns represent reality. Here we investigate the threshold at which finer scales of spatial resolution do not meaningfully add to the analysis when working with various spatial data aggregations.

2.2 Goodness-of-fit for spatial point patterns

The method proposed here requires a metric to quantify the difference between two point patterns using an area-based test. Therefore this section reviews some of the most commonly used approaches and outlines the gap that the new method proposed here has been designed to fill.

One way of comparing two point patterns directly is to calculate statistics that describe their first- or second-order properties and then compare the values of these statistics. A common example are nearest-neighbour statistics that examine the average distances between points. For example, the Clark and Evans R statistic [20] is the ratio mean minimum distance in a point pattern to that which would be expected under complete spatial randomness. A significance test has also been developed [21]. Therefore by calculating R for two point patterns, the degree of clustering can be compared. An improvement over the R statistic is Ripley’s K function [22] which, rather than only considering a single minimum distance for each point in a pattern, takes all of the neighbours that are within a given distance, d, into account. The value of K at distance d is then calculated as the mean of all counts divided by the overall point density. This provides more information about the point pattern as it estimates the amount of clustering present at different distances. Again, graphs of Kd for two point patterns can be compared to estimate similarity between disparate data sets. In addition, the K statistic can be adapted such that rather than examining distances between points of the same type, it can compare distances between points of different types (i.e. comparing two different crime types simultaneously); termed ‘cross-K’ [23]. There are a number of related statistics that measure slightly different properties of point patterns; see [24] for detail.

The main drawback with these approaches is that they do not provide any information about the differences in the locations of the points. For example, two very different datasets could have similar R or K values if the degree of clustering was the same but the clusters were located in different places. The use of cross-K would mitigate this problem as it compares two datasets directly, but as it is a global statistic it does not provide information about where the differences between the two datasets occur in space. As an alternative, rather than analysing the point patterns directly it is possible to first aggregate the data to some neighbourhood to create a table or matrix. Administrative boundaries are typically used as the unit of aggregation, but this is not essential. Aggregation has the advantage that, once the data are in tabular form, a range of well-developed goodness-of-fit statistics can be used and, in addition, the cell differences can be mapped to show how the differences vary across space. Relevant statistics include include:

• residual sum of squares—calculated by simply summing the square of the difference between the number of events in each areal unit;
• R2—which represents the percentage agreement between a model and the observed data;
• (Standardised) Root Mean Square Error—which has been found to be comparable across spatial systems and not sensitive to large deviations from the mean [25];
• Andresen’s S [7–9]—which is adapted for use here.

The main drawback with the typical application of these metrics is that they are usually applied at a single resolution and provide no information about whether the scale of spatial aggregation is appropriate. To overcome this drawback, [6] outline a multiple-resolution method that was developed to measure the similarity between a model of a spatial process and real data. The method was developed for the field of ecology, which is a discipline that has a history of work into appropriate cell (‘quadrat’) sizes. For example, in the 1950s, [26] experiments with the use of different cell sizes with the aim of identifying an optimal size for use in sampling populations of plant species. The authors find that species density is extremely important in the determination of an appropriate sampling cell size, which echoes some of the results discussed here in Section 5.1. More recently, [27] and [21] also discuss the implications for changing cell sizes during sampling. The method that this work builds on most directly, as developed by [6], has been used to assess the reliability of land use change models (e.g. see [28–31]). It has also been used in the field of environmental criminology to assess the reliability of agent-based models [32, 33] but is otherwise not widely used as few studies report to undertake multi-scale validation [34]. By testing the similarity of real v.s. simulated data at multiple resolutions, it is possible to show how well different models perform across different scales. This paper will build on the method proposed by [6] to develop a technique that can be used to assess how similar two data sets are at a given resolution, and estimate the resolution that is the most appropriate for studies of the phenomena.

3.1 Multi-Scale Error Analysis (MSEA) method

To answer the question of the appropriate spatial scale, as defined above, this paper presents a new method that builds on a multiple resolution goodness-of-fit procedure [6] and combines it with Andresen’s S Index [7–9, 35] for measuring similarity. The aim is to identify the spatial resolution at which no further disaggregation of areal units leads to new information regarding the spatial patterns within the data: the appropriate scale of analysis. In other words, at what point is there little or no value added in using smaller spatial areal units of analysis? The analysis source code (in the R language) is available as supporting information (S1 File) and online at https://github.com/nickmalleson/Multi-Scale-Error-Analysis/. The source code and required data are both open source, so the analysis presented here can be repeated in its entirety using the source file.

3.2 Statistical significance and quantifying local ‘similarity’

As spatial units decrease in size, there will be fewer points per cell than compared to larger units. It is not possible to determine whether the number of events that occur in a cell from two point patterns are significantly different if the numbers of events are too small. What should be done with grid cells with such low counts such that they cannot be distinguished with any statistical significance? If these grid cells are simply marked ‘similar’ they will inflate the global similarity artificially, resulting in perfect similarity when using very small grid cells. For this reason these grid cells are removed from the analysis completely, thus focussing on the cells for which meaningful differences can be found. To make the decision on which cells to remove, the standard recommendation for the Chi-squared test is used as follows.

Consider the case where there are n spatial units, U1, U2,…, Un and two point patterns: X1 and X2. It is then possible to estimate the expected number of points from X1 or X2 in any given unit, Ui, by multiplying the row and column marginals and dividing by the grand total. For example, to calculate the expected number of events from X1 in unit Ui: (2) where X(i) denotes the number of points from process X in unit i. When the number of units, n, is large, it will be the case that for many units U1, U2,…, Un the number of points per unit will be very low, Ui ≪ n. Following standard recommendations the threshold used here is set a 5, such that any cell with less than an expected frequency of 5 (either or ) is excluded. This is a conservative test [38], so the cut-off of 5 could be relaxed in future work.

3.3 Data

The data used here represent calls or complaints made to the Vancouver Police Department. They are publicly available through the City of Vancouver Open Data Catalogue (https://data.vancouver.ca/datacatalogue/crime-data.htm) and are available from 2003 to the present under various crime categories. The data are anonymised prior to being made availability on the portal. Here, the following four crime categories have been chosen:

1. Breaking and entering—residential (BNER);
2. Breaking and entering—commercial (BNEC);
3. Theft from vehicle (TFV);
4. Theft of bike (TOB).

These four types have been chosen because: they are voluminous (each has at least 2000 events per year) so there is value in applying quantitative methods; they represent crimes that are likely to be reported due to the need to make insurance claims so should be largely representative of the true crime distribution; they have very different spatial patterns (some are much more heavily clustered than others) so represent a good test of the method.

For each crime type, two data sets are required for a comparison. Here, data from 2015 will be compared to those for 2016. It was decided to use 12 months worth of data to ensure that there are a sufficiently large number of points for meaningful quantitative analysis and because 12 months is a time period that is commonly chosen in the field to mediate the impacts of seasonality on crime patterns. It is important that during the comparison period (Jan 2015–Dec 2016) the crime patterns do not change radically, otherwise the method will identify these changes rather than the inherent randomness that will always be present in the study of social phenomena. During the time period chosen there were no major policy or policing interventions that would suggest the patterns should be radically different. Although it is beyond the scope of this paper to attempt to separate the influence of the random fluctuations in the pattern from any longer-term systemic changes, future work will attempt to derive a more precise definition of the point at which wider systemic changes obscure the natural fluctuations (i.e. more precisely defining the point that they become ‘radically different’).

Fig 2 illustrates the density of each of the crime types in 2015 and 2016. Although most are similar (as required), residential burglary (BNER) does appear to have undergone some changes in the two years. It is also much more widely distributed throughout the study area than the other crime types. Both of these factors have interesting implications for the results, as Section 5 will discuss.

To assist with interpreting the size of the cells in the different grids, six different grid resolutions are overlaid on to some of the results graphs in the later sections. The resolutions of these grids range from approximately 6Ha (slightly smaller than a dissemination area; the smallest Canadian census unit) to 100Ha (roughly equivalent to a census tract). These resolutions do not influence the analysis itself, but just illustrate how the size of the cells in different grids relate to the size of city blocks.

4.1 A simulation of point similarity

In order to better understand how the method will behave when applied to real spatial data, it is first applied to some simulated point patterns. Five different point patterns are created using an inhomogeneous Poisson process as implemented using the spatstat::rpoispp R function. The formula used to determine the density of the points per unit area is: (3) where a is a balancing factor used to determine the approximate number of points created (larger values for a result in more points), and b and c determine the distribution of the points in the x and y planes. Larger b and c values mean that the points are more clustered to the lower end of the scale. For example, the first point pattern created (P1) has b = c = 1 so points are weakly clustered in the lower left corner. As b and c increase the points become more clustered. The clustering itself is not important, what matters is the difference between the patterns; this can be controlled by varying b and c. a is chosen so that each pattern has approximately 3,000 points, which is a similar number of events to the crime data. Table 1 describes the point patterns that were created to run the simulations. P1 and P2 are the most similar as they are created using identical parameters. P3, P4, and P5 become increasingly dissimilar.

Four experiments are conducted in total. In each experiment, P1 is compared to one of the other simulated patterns. Fig 3 illustrates the density of each of the point patterns and Fig 4 illustrates the variation in similarity.

With all point patterns, the mean similarity initially increases. This is an unintuitive result as similarity could be expected to decrease with cell size because it becomes less likely to find points from both datasets in the same cells. The increase occurs because with very large cells, some (i.e. those near the origin) will have a very large number of points in them when compared to cells that are located nearer the outskirts of the simulation boundary. We observe an increase in the proportional difference with increasing number of points, such that if there are cells with very large numbers of points, they will have a large statistically significant proportional difference. This reduces the similarity for large grids. After reaching a peak, the mean similarity then remains relatively consistent or begins to decrease. Importantly, it is immediately obvious which point patterns are the most similar. In simulation 5 (which compares the least similar point patterns—P1 and P5) the similarity takes considerably longer to reach its peak, does not reach such a high level, and declines afterwards.

4.2 Expected crime frequency

Before reporting the results of the method applied to the real crime data, it is useful to better understand how the spatial scale (i.e. grid resolution) affects explanatory power. Fig 5 illustrates the mean expected value at each resolution. This is constructed by first calculating the expected value (discussed in Section 3.2) for each cell, and then calculating the mean expected value across all cells at that resolution. Assuming that crimes are reasonably uniformly distributed across the study area, then grids with a higher mean value have a greater number of cells with a sufficient number of points to be able to calculate similarity. Those with lower means will have fewer cells with sufficient explanatory power. Whilst not directly relevant to the main results of the paper, the distribution of expected values could, potentially, reveal interesting information about the underlying spatial distribution of the crime patterns. For example, where a crime pattern exhibits a larger mean than median, this implies that the points are clustered; the mean is skewed by a few cells with very large numbers of points in them.

Fig 6 provides more information about the expected frequency by illustrating the proportion of cells at each resolution that have a sufficiently large expected value in order to be able to estimate the similarity between the two point patterns. This is useful because it allows for statements such as “at resolution a, b percentage of all cells are used to estimate the overall similarity”. Where b starts to become very low then it becomes difficult to make statements with confidence about overall similarity because there will be few cells actually used in the calculation. Again, this is not directly relevant to the results here, but could be useful as a means of estimating the spatial unit of analysis at which there will be, on average, an insufficient number of events to test properly for local spatial differences a priori.

4.3 Overall similarity changes

Having outlined the proportion of cells that, at each resolution, have a sufficiently high number of points to make a comparison between the two crime patterns, Fig 7 illustrates how the mean similarity changes with resolution. With all crime types, the mean similarity initially increases in a similar manner to that of the simulations. After reaching a peak, the mean similarity then remains relatively consistent or (as in the case of residential burglary) decreases. These results are discussed in detail in Section 5.1.

5.1 The most appropriate spatial scale

Identifying the most appropriate scale is non trivial. As the size of the spatial units changes, there are two distinct statistical processes that take place. These, as illustrated in Fig 8, are:

1. Initially, the point pattern similarity is low, and it increases as the number of cells increases. This is because when there are very large spatial units, some cells have very large numbers of points in them. We observe an increase in the proportional difference with increasing number of points, such that these cells with large numbers of points (compared to others on the outskirts of the study area that have fewer) have a statistically significant proportional difference that is not necessarily large in magnitude. This reduces the similarity for large grids.
2. At the other end of the scale, where there are many small cells, there will be few points in each cell. This has the effect of widening the confidence intervals and subsequently making it less likely that there is a significant difference. This artificially amplifies the similarity. To remedy this we remove individual cells that have insufficient explanatory power (as per Section 3.2), otherwise a point pattern would converge on total similarity (Sg → 1). The removal of these cells means that although the similarity does not tend to 1, at the highest resolution grids only a small proportion of cells contribute information to the global similarity measure.

The end result of these two processes are the graphs that were illustrated in Fig 7. The most appropriate spatial scale can be found somewhere between the two extremes of the graph. The pertinent question therefore becomes: where, in the region between the two extremes, does the most appropriate spatial scale lie? Interestingly, the behaviour of the global similarity varies by the crime type in question, and therefore the explanation needs to be tailored for each one.

• Residential Burglary (BNER). With BNER, there is a peak in similarity when cells are approximately 20 hectares in size. This size, which is close to twice the square area of a dissemination area (the smallest Canadian census neighbourhood) appears to be the most appropriate spatial scale at which to analyse the phenomenon. Of course, using larger units will hide lower-level patterns so generally smaller is better (the contemporary ‘crime at place’ literature is clear on this). But, importantly, these results suggest that when using units that are much smaller than 20 hectares the noise, that will always be present with such spatial phenomena, is captured instead of the signal. In other words, there are so few points in cells below this size that we can expect significant volatility and, subsequently, unreliability in analyses.
  It is, however, important to recall that the 2015 and 2016 BNER point patterns are not as similar as those of the other crime types. Therefore some of the dissimilarity is caused by wider spatial variations across the two years (i.e. crime hotspots moving) rather than purely due to the underlying randomness in residential burglary. Were the patterns more similar then the most appropriate spatial scale would be smaller. Future work can investigate this particular case in more detail, particularly with respect to better understanding why the spatial distribution across the two years is not more consistent.
• Theft from Vehicle (TFV). In this case, the similarity increases until the spatial units are approximately 20ha and then plateaus. With TFV, therefore, there appears to be no disadvantage, in terms of signal v.s. noise, to using spatial units that are smaller than 20ha. However, it is worth noting that as the cell size decreases there will be fewer cells with sufficient explanatory power to contribute to the global similarity. For example, from Fig 6, approximately 65% of all cells have sufficient power when the size of the spatial units is 20ha, but this decreases to around 40% with cells of 10ha (roughly equivalent to a dissemination area).
• Commercial Burglary (BNEC) and Theft of Bike (TOB). As with BNER and TFV, the similarity increases rapidly initially but then, rather than plateauing, it continues to increase gradually. This suggests that, even with cells of only 6ha, there is still a signal that can be identified above the background noise. However, as with the other crime types, the number of cells with a sufficiently large number of crimes will be very low at these fine resolutions. This result is most likely to be a consequence of the degree of clustering that these two crime types exhibit; a point that will be elaborated on further shortly.

A further important consideration is the number of points in a dataset and, similarly, the point density. As the number of points increases then, if everything else remains equal, the appropriate spatial scale becomes finer. This is because with two similar patterns, more points means that the cell size will be able to become quite small before noise begins to hide the signal. This is important because phenomena that occur frequently, or at least frequently in a specific area (e.g. a highly clustered phenomena) can be reliably disaggregated to quite small cells, at least in the areas where the phenomena occurs densely. The ‘appropriate spatial scale’ not only depends on the abundance of the phenomena itself, but also the degree to which it is clustered.

5.2 Spatial distribution of similarity

Section 5.1 discussed the potentially most appropriate spatial scale for each crime type. But what does this look like in practice? To provide some context, this section explores the spatial distribution of the local similarity measures at a few different scales. The focus is on residential burglary (BNER) and theft of bike (TOB) exclusively, as these two crime patterns have very different spatial structures and, hence, demonstrated very different similarity profiles (as per Fig 7).

To begin with, it is important to consider that that BNER is much more widely distributed throughout the study area than TOB (as was evident in Fig 2). This is entirely expected given the contemporary environmental criminology explanations for observed crime patterns and their relationship with the distribution of potential targets. In short, residential properties are widely distributed, but bicycles are much more likely to be available (and accessible) outside transport hubs or places of employment. The interesting thing to note about these different degrees of clustering is the impact that they have on the most appropriate spatial resolution.

Fig 9 illustrates six grids at the illustrative spatial scales that were chosen in Section 3.3. It shows the local similarity values for BNER and TOB at the different resolutions, where 1 implies similarity in the cell and 0 implies no discernible difference. Cells with insufficient explanatory power are also highlighted.

It is immediately apparent that, as outlined earlier, the number of cells with sufficient explanatory power for burglary drops rapidly once cells become smaller than 20ha. With cells of 10ha, there are very few cells that have a sufficiently large number of points to determine, with confidence, whether the two point patterns are similar or dissimilar.

Regarding TOB, although the majority of the study area does not exhibit sufficiently large expected values, there is a consistent spatial pattern in (dis)similarity in the central business district and surrounding transport hubs, even down to cells that are under 10ha (less than a single dissemination area). In these areas, the appropriate spatial scale is very small indeed. These results therefore suggest that there is value in obtaining extremely high resolution data because even with very small spatial units there is still sufficient signal above the background noise to warrant empirical analysis.

5.3 Is ‘smaller’ really ‘better’?

On the whole, recent research in environmental criminology has pointed towards the importance of ‘micro places’ in crime analysis [5, 10–13]. It is important to explain here that these findings do not reject that assertion. High quality, disaggregate data will always be preferable to more aggregate data. At the very least, disaggregate data can always be aggregated. For example, current work that examines the impacts of the road network [40, 41] on crime occurrence would most likely obfuscate important micro-level features were the data aggregated to neighbourhoods. This situation holds in other fields.

This work does not recommend aggregating point data to the ‘most appropriate’ spatial unit. Rather, the purpose of this work is to provide a framework that can be used to, among other things, explore the scale at which the broader spatial patterns start to become overshadowed by random noise. The is only relevant for research that is specifically interested in spatial patterns, not for similar research that, although spatial in nature, does not explicitly consider space. In the circumstances where the broader spatial patterns are of interest, the method presented here can be useful in two respects. Firstly, in the cases where data that have been aggregated to a spatial neighbourhood are available, this method can be used to estimate whether there would be value in attempting to obtain finer-scale data. The method could be performed by the data controllers to look for evidence for the need to release higher resolution data to the analyst. Secondly, for studies that do require aggregate data (such as those that explore neighbourhood-level phenomena) it can provide evidence for the most appropriate spatial unit to use in the aggregation—i.e. not so large that the scale causes lower-level patterns to be smoothed out, but not so small that there are insufficient events per unit to make any statistically significant assertions.

5.4 Meaningful spatial units

Before concluding the paper, it is worth briefly discussing the implications of using rectangular grid cells as the underlying spatial unit. A body of work in environmental criminology, as well as in other geographical fields, has attempted to identify the most meaningful spatial unit for the analysis of spatial phenomena. For work on social phenomena, these units are typically defined in a way that associates them explicitly with the underlying space, i.e. one that connects with how the infrastructure shapes people’s travel patterns, communication, sense of community, etc. For example, recent work on street segments has been particularly illuminating [40–44]. A regular grid of cells, such as that used here, does not meaningfully link to the underlying phenomena. This is a drawback because it is well known that the real underling geography will have significant effects on the level of crime. Therefore future work could design an algorithm to create more meaningful units at various different resolutions—using (for example) physical boundaries, social data, and land-use information—in a similar manner to those that are used to create census and other administrative boundaries currently. Such an effort is well beyond the scope of this work though, and it is worth pointing out that similarly meaningless spatial units, such as postcodes, are also used regularly and do not reflect real ‘neighbourhoods’. Finally it is also worth mentioning that the use of grids will also probably lead to ‘zoning effects’ [5] because the zones have not been drawn with the deliberate intention of maintaining homogeneity as is typical with the creation of administrative boundaries. Again, future work that explored more nuanced mechanisms for creating appropriate / ‘meaningful’ spatial units should account for zoning effects as well.

S1 File. Source code (R markdown).

The empirical work outlined here can be conducted, in full, using the R source code included in the publication and also available at: https://github.com/nickmalleson/Multi-Scale-Error-Analysis. It is in ‘R Markdown’ format so can either be used to generate an html file with text and graphics included, or the R commands can be run independently. The script depends heavily on the sppt R package [35]. The script attempts to download the required crime data (that are open source), so all experiments outlined here can be run in their entirety. The only data that are required are Vancouver administrative boundaries that are used for drawing maps. They are available here: https://github.com/nickmalleson/Multi-Scale-Error-Analysis/tree/master/vancouver_boundaries. The files must be downloaded and placed in a directory called vancouver_boundaries in the same location as the script. The easiest way to run the script is to load it into RStudio and run it from there. Alternatively, it can be run from a bash command line to generate an html file using this command (all on a single line):

Rscript -e “require(knitr); require(markdown);

knit(‘multi_scale_error_analysis.Rmd’,

‘multi_scale_error_analysis.md’);

markdownToHTML(‘multi_scale_error_analysis.md’,

‘multi_scale_error_analysis.html’);”

https://doi.org/10.1371/journal.pone.0218324.s001

(ZIP)