Extracting trends

One particularly major modern bias that exists in the datasets being studied by EngLaId is that of where archaeological activity tends to take place in England.  In particular, commercial archaeology tends to take place more commonly in areas where more modern development takes place (especially relevant to our AIP and HER datasets) and also there are obvious biases in the PAS towards areas which are more popular with metal detectorists /  have better conditions for metal detecting / where Finds Liaison Officers are more well-established.

It is necessary to somehow quantify these biases, so that it can be discerned whether patterns discovered are more likely to be a true reflection of past activity or more likely to be an artifact of modern archaeological activity.  Amongst others, Andrew Bevan has done some very interesting work in this area using kernel density estimates (Bevan 2012), but I have my own bias towards a slightly different approach: trend surface modelling.

The trend surface is, essentially, an attempt to model underlying trends in a point based dataset based upon numerical values attached to those points.  The algorithm creates a polynomial surface which tries to reflect that trend.  It is then possible to use that surface to test whether individual points within the original data either fall above or below the expected value (bucking the trend), or fall more close to the trend itself.  This would then require explanation.  The first step, however, is creating the trend surface.

As a measure of where commercial archaeological interventions have taken place since the onset of PPG16 in 1990, I first extracted AIP data for each county of England via their website, restricting my query to the years 1990 to 2010 and to field evaluations and post-determination / non-planning events.  These spreadsheets were converted, combined, and imported into ArcGIS using my script.  The resulting points were then counted per 1 x 1 km grid cell and a 12th order (the most complex surface available in ArcGIS) trend surface created:

trend_AIP
12th order polynomial linear trend surface for field evaluations and post-determination / non-planning events for 1990 to 2010 recorded by the AIP.

Essentially, this map then shows the expected amount of interventions per square kilometre across England.  Although the values are small (0-2+), the pattern looks convincing to me, with obvious peaks in London, Kent, the eastern south coast, the Bristol region, South / East Yorkshire and Tyneside.  There are also clear troughs across most upland regions of England.  It is particularly obvious how much commercial archaeological work has taken place in London over the past twenty years.

I then repeated this task for the PAS data for all time periods (as of August 2012):

trend_PAS
12th order polynomial linear trend surface for all PAS finds up to August 2012.

Again, this map shows the expected number of finds per square kilometre across England.  Here we see obvious peaks in East Anglia, Kent, the eastern South coast and the Isle of Wight, western Cornwall, Northamptonshire, and Humberside / Lincolnshire.  Again, there are obvious troughs across the uplands of England and also in some areas of dense settlement (Tyneside, Essex, Medway).  As stated above, there are several factors at play here, but the pattern seems a convincing and useful one to my eyes.

The next stage for the PAS data would be to construct further trend surfaces for each of our time periods and see how these compare against the overall trend across all periods.  In this way, it ought to be possible to pick out areas for further study which show particular peaks within a single time period that are not present in the overall trend.  Whether this works, only time will tell!

Chris Green

References:

Bevan, A. 2012. “Spatial methods for analysing large-scale artefact inventories.”  Antiquity 86, pp. 492-506.

Author: Chris Green

Postdoctoral Researcher (GIS)

8 thoughts on “Extracting trends”

  1. Chris this looks amazing! What does a combined map of both datasets look like? I think it would be interesting for a better understanding of where there is a greater paucity of data.

Leave a Response

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s