Methodological Issues in the Combined use of GPS, GIS and Accelerometry in Research on Greenspace and Physical Activity with AdolescentsJustine Geyer TSG/ESRC PhD Studentship

Child and Adolescent Health Research Unit (CAHRU)University of Edinburgh, Scotlandj.geyer@sms.ed.ac.uk 23rd February 2011

Green spaceNatural environments in urban settings

Quantity and proximity linked to higher levels of physical activity in adults and adolescents

Leisure time use of green space and its relationship to physical activity

Method35 adolescent girls and boys aged 13 and 15 years from 3 towns in Fife, ScotlandMonitoring of physical activity using Actigraph AM7164 uniaxial accelerometers 30s epoch, all waking hoursTracking of geospatial behaviour with Blackberry 8900 GPS enabled mobile phones using GPSlogger* - 30s epoch, kept with them at all times

* Sourced fromemacberry.com

Fife Council GIS green space layer used to assign green space land use code to GPS location data

AccelerometerGPS enabled Blackberry

Experiences - Fieldwork

Data LossGPSlogger and fiddling fingers!Applock

AcceptabilityIncentiveAcceptable lookPrimarily a phoneFamiliarity with rechargingLarge memory 1GB, >16,000 recordings

Dash to volunteerApproval of phoneGood variety of volunteers

LimitationsSeparate to accelerometer and not attached to bodyExpense limits sample size

Experiences - Data processing and analysisUsing GIS data to characterise GPS data from phones

Missing data and imputation

Joining GIS with GPSA GS code was assigned to all GPS data with a distance valueNon-GS (Home/street/shops) assigned a GS code usually a gardenRequirement to post process to reclassify and exclusion of garden exposure

Beach erroneously treated as non green space

Missing Data and ImputationAmount of missing data extremely variable between individuals ranging from 98.6% to 8.3% (includes low quality participants)

Exclusion of days where actigraph data >30mins unexplained gap or GPS >60 mins gap possible non-compliance

Total percentage missing data for aggregate dataset was 25% (before imputation)

Pattern of missing data frequency of durations, reasonable to impute?

Imputation if bounded by same code and within 35m of each other then impute based on last known land use coding

Is it reasonable to assume missing data is non-GS? proportion of boundary pairs that are in GS probability that longer missing sections are bounded by nonGS 15% of bounding GPS points were GS coded

ConclusionsExciting method with great appeal to adolescents

Promising level of detail and accuracy for researchers

Fieldwork and data handling challenging

Caution required in reliance on results from this method

***

I have used a number of different methods in my research but today Im focusing on my experiences with the combined use of GPS, accelerometry and GIS, in particular how the fieldwork went and some of the issues encountered during data processing and analysis. Green space as Im sure many of you already know refers specifically to natural environments, such as parks, woodlands or beaches that occur within urban environments. Research has shown that the amount of green space in a neighbourhood, or living in proximity to it has been linked to higher levels of physical activity. My interest has specifically been in how green space is actually used by adolescents during their leisure or non-school time and how levels of use relate to physical activity.This research involved a group of 35 male and female adolescents aged 13 and 15 years, recruited from 3 towns in Fife, Scotland.They wore an accelerometer on their hip from waking till bedtime and only removed if it was likely to get wet or cause harm during an activity. The adolescents were asked to note down on/off times in a diary and activity was summarized every 30 seconds.Participants were also given on loan a Blackberry GPS enabled mobile phone and asked to keep it switched on and with them at all times. GPS data was set to be captured every 30 seconds by GPSlogger software to match to the 30 sec epoch used for the actigraph. A potential for 5 or 6 full days of monitor data was possible once the drop off and pick up days were excluded.Finally a GIS dataset from Fife Council which detailed all urban greenspace in Fife by type, location and size was used to assign a green space code to location data provided by the GPS phone. The different colours illustrated here represent different types of green space with the grey representing non green space such as civic space.In general the fieldwork went very well, however a number of issues are worth mentioning.Data lossI lost pretty much all the GPS data for the first 8 teenagers because they managed to accidentally switch off the GPSlogger application even though it was in a hidden folder and they were instructed not to access it. This was eventually resolved with an application that made theGPSlogger password accessible only called Applock.AcceptabilityIncentive - One of the primary reasons I used Blackberrys for the GPS unit was as an incentive, so they would appeal to teenagers in what could have been viewed as intrusive research.Cool Also I thought it would be more attractive than alternative GPS options. Some of the other choices would probably have been aesthetically unacceptable to a population group increasingly concerned with appearances.Not primarily a GPS Another advantage was that the primary function of the Blackberry was as a phone and not a GPS receiver so I envisaged that its purpose for the research would be quickly forgotten and so have little impact on normal behaviour.Finally the phone had good battery life and memory - A common problem with GPS units appears to be the battery life and to some extent the data storage capacity for several days worth of continuous free-living monitoring which can affect compliance and data quality. Mobile phones are almost ubiquitous within the teenage population, certainly in the UK and so it was thought that teenagers would be used to charging their phones regularly. The battery life with GPS on was somewhere between 18 to 24 hours. Also the phones had a 1GB memory card supporting the phone memory which proved more than adequate.The phones seemed to work well as an incentive to take part and not a deterrent. This is illustrated by the comment from one mother who said she had never seen her son volunteer for something so fast in her life. The teenagers were all really positive about the phones and appeared to attract a good spread of participants from different backgrounds both sexes and both age groups. On reflection I consider the Blackberrys worked well as the GPS unit but suffer a couple of significant limitations:The first is not being attached to the participant in the same way as an accelerometer, this introduces problems determining when a participant has been compliant and thus ultimately the quality of the data.The phones are expensive which impacts on sample size.As many of you will know, who have gone through the process of cleaning and merging accelerometer, GPS and GIS data this is an involved and very time consuming process. I dont have time today to go into the detail of all the processes I have gone through. Today Ill focus on some of the issues I encountered when assigning land use data to the GPS data and some of the issues with missing data.This slide shows an example of the merged GPS and accelerometer data. Highlighted in red are sections of missing GPS data and in this yellow column is the activity data highlighting where activity data is present but there is no matched GPS data.Here we have an example of one persons GPS data over several days overlaid onto the green space layer where everything in green has been coded as green space including private gardens and everything in white is nonGS.When the green space layer was joined to the GPS data a green space code was assigned to every GPS datapoint regardless of whether it was in greenspace or not. A distance was given for how near each point was to the assigned nearest greenspace parcel. So for example this point here (highlighted in blue) is part of a journey down a street but has been assigned a greenspace code and is just less than 3m from it. The result was that even when traveling down a street, sitting at home or shopping in a shopping centre, a green space code was still assigned usually a garden one. This meant that post processing of the greenspace codes was required to reclassify garden codes as non-GS and set a distance threshold that if exceeded also meant nonGS so that these points here would be correctly classified as not in GS.A consequence was that garden exposure could not be included as a form of green space exposure. Actually this highlights wider issues in how green space exposure is defined for the purposes of research but thats a whole other area.An additional problem only came to light whilst I was preparing for this presentation. This was that the green space layer had not coded the beach as green space which technically it should be and so these GPS recordings here (which are actually on a beach in one of the towns) were coded as nonGS. The general point here really is to illustrate how the nature of the GIS information has the potential to affect results and adds another layer of uncertainty. I chose to use this dataset because it provided such rich green space information yet it has come with a set of issues I didnt anticipate.The second main issue I wanted to briefly cover was in relation to missing data. Missing GPS data is widely recognized as an issue, especially when seeking to combine it with activity data from accelerometers. But how that is treated varies greatly in the literature and is dependant on what questions are asked of the data. A process that is still ongoing for me is to describe the nature of my missing data so I can evaluate the potential it has to influence my results. So far I have looked at the extent of missing data and am still going through the process of describing the pattern of it.The amount of missing data by individuals was extremely variable and ranged from between 98% at worst to 8% at best. This included participants where compliance was obviously an issue so an aggregate dataset was created excluding any day where GPS data had a gap > 60 minutes or unexplained gaps of >30mins actigraph data, as this was a possible indication of non-compliance.The total percentage of data missing from the aggregate set of best quality days was better at 25%.A preliminary check was then done to assess the frequency of missing data sections of different durations. A substantial proportion of missing data were under 10 mins duration and therefore it was deemed reasonable to impute these to reduce missing data based on two criteria:The data bounding either side of the missing section was the same GS coding and The two points were within 35m of each other suggesting the participant had stayed in same area based on walking speed of 4kmph and smallest time gap of 30 sec The next question to ask was whether it was reasonable or not to assume that the remaining missing data occurred purely because the participants were indoors or in a car or suchlike and therefore could safely be imputed as nonGS?The simple answer is no. This is because the Blackberrys recorded even when indoors and also nonGS included areas outdoors.

This is something I am still working on and the plan is to assess this by:1) Seeing what proportion of datapoints either side of a missing section are coded as GS or not2) Seeing if there is a pattern to the durations of missing data, in other words whether longer missing sections are more likely to be nonGS or not.The results of this will ultimately dictate how I proceed with analysis and whether such missing data can be ignored or needs to be taken into account in some wayAs a preliminary indication a simple analysis shows 15% of the boundary points are coded as in GS which suggests a need to further assess the impact on results of the missing data.In conclusion I think this is an exciting approach, appealing to adolescents in what could be considered intrusive research. I think it has the potential to offer some really useful insightful data for environment and health researchers. However fieldwork and data processing are challenging. Greater transparency is required in how data is proceesed and data quality remains to be fully evaluated before there can be confidence in the results produced from this style of investigation.