Maps of inferred relative PD prevalence, England

Geographical variations in the prevalence of Parkinson's Disease are of interest for a number of reasons. They may give clues as to:
- the aetiology of the disease;
- the effectiveness of various therapies:
- what environmental toxins are involved in PD.
In short, if we understood the spatial differences in PD's prevalence, we would be much closer to finding a cure.

This post goes into some detail into how to use data sets containing spatial data for England that are open to the public to map PD.

As far as I'm aware, there are only two papers that give detailed maps of prevalence across a whole country. One by Willis et al. [1] for the US. And one by Pedro-Cuesta et al. [2] for Spain. I've found nothing similar for the UK.

I've used the data to produce "dot" maps of PD prevalence in England. See:
http://www.parkinsonsmeasurement.org...ceEngland1.htm

The maps are based on prescription data from June, 2012, from each of about 10,000 GP's surgeries in England. Some of the smallest surgeries will see very few PwP, so their results that may be misleading. To avoid this, only those practices writing at least 500 prescriptions in the month have been included in the analysis. This has left us with a final sample size of about 8400.

The maps show inferred relative PD prevalence (a term that is described below) which is a crude measure for true PD prevalence. It is the color of the dots which is interesting, not the number or dots (that reflects population density).

The UK government through its Open Data initiative [3] has now made it easy to, at least, make a start on getting spatial prevalence figures for England.

Amongst the many data sets that they have opened up to the public are ones relating to prescriptions.

The main condition for their use is that you attribute the source, which in this case is:
Contains Ordnance Survey data © Crown copyright and database right [2012]
Contains Royal Mail data © Royal Mail copyright and database right [2012]
Contains National Statistics data © Crown copyright and database right [2012]
Provided you do this the Open Government Licence gives you the right to use, distribute and adapt the data.

For this work I've used two data sets:
-Raw data for June 2012, from files T201206PDPIEXT.CSV and T201206ADDREXT.CSV obtained from [4].
"All prescribed and dispensed medicines (by chemical name), dressings and appliances (at section level) are listed for each GP practice. -For each GP practice, the total number of items that were prescribed and then dispensed is shown. -The total Net Ingredient Cost and the total Actual Cost of these items is shown." The second file gives addresses including, importantly, postcode.
-Positional data from the Ordnace Survey, Code-Point data set [5]. This gives Easting and Northing information for each postcode.

There are approximately 10,000 GP practices in England and each is in the dataset. From the point of view of mapping the density is excellent.

The crucial question is: How do we extract prevalence data?

Strictly speaking, we don't. Instead we look for a weaker measure: relative prevalence. This allows us, in this context, to answer questions like, Where in England is PD most common?, but not how many people in England have PD? Or how this compares with the US?

So, we're looking for a measure of relative PD prevalence. That's not in the database either. A proxy is used:

.................................................. NUMBER PARKINSON'S PRESCRIPTIONS
INFERRED RELATIVE PREVALENCE = ------------------------------------------
.................................................. ....TOTAL NUMBER PRESCRIPTIONS

I listed the "Parkinson's drugs" in a recent post on NeuroTalk [6].

Is this a perfect measure of the true rate of PD prevalence? Of course not, for instance:
- PwP not on prescription drugs are missed;
- doctors who prescribe two different drugs, rather than double the quantity of a single drug will get twice the inferred prevalence;
- doctors who prescribe small quantities, e.g. a week's supply rather than a month's, on each prescription will have higher rates;
- different local prescribing regimes will lead to differing inferred rates;
- demographics, like age profile, are missed;
- changes in the use of non-Parkinson drugs at place A will impact on the relative inferred rate at place B.
- some of the Parkinson drugs can be used for other conditions, etc..
But, although these problems will certainly add noise to the statistics, I don't think they will invalidate the whole approach.

And the results? It is early days yet, but some things can be seen:

Plotting the whole sample does not show large regional differences like those reported in the US. The most telling measure of this, albeit not a sufficient condition to disprove large variances in distribution, is that the "centres of gravity" of the source of all prescriptions compared with the PD prescriptions differ by only a few miles. See Map 1.

However, analyses based on on comparisons between the lowest and highest prevalence practices hint at spatial variations. For instance, Map 2 shows the bottom 5% and the top 15% of practices ordered by inferred rates, hint at variations, with the South West appearing to have higher prevalence rates. This may just reflect demographics.

Where to next? I'd like to normalize the data to account for age differences, look at more months of data and to do cluster analysis.

I'll be grateful for people's comments. If anyone needs help getting started data mining, I'll be happy to help.

References

[1] "Geographic and EthnicVariation in Parkinson Disease: A Population-Based Study of US Medicare Beneficiaries"
Allison Wright Willis, Bradley A. Evanoff, Min Lian, Susan R. and Brad A. Racette
Neuroepidemiology. 2010 April; 34(3): 143–151.
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2865395/

[2]"Spatial distribution of Parkinson's disease mortality in Spain, 1989-1998, as a guide for focused aetiological research or health-care intervention"
Jesús de Pedro-Cuesta, Eduard Rodríguez-Farré and Gonzalo Lopez-Abente
BMC Public Health, vol 9, Dec 2009.
http://www.biomedcentral.com/1471-2458/9/445

[3] http://data.gov.uk/

[4] http://data.gov.uk/dataset/gp-practice-prescribing-data

[5] http://www.ordnancesurvey.co.uk/oswe...int/index.html

[6] http://neurotalk.psychcentral.com/thread179755.html

John

Wow, John! That is impressive and rather, I daresay, pretty. You have created an aurora borealis of PD

It is very striking to see that the Wilis study challenges the long held and widely disseminated assertion that rural living or working in a rural environment increases the likelihood for PD. The map clearly shows the opposite. What I find really alarming is the Great Lakes "trail"- that is not just Midwestern US but more a huge sweeping arc along the states bordering Lake Erie extending into the Northeast; I wonder if this prevalence extends across the Canadian borders?

I was also surprised to see in vastly green areas pockets of PD or clusters show. IMHO, it is easy to surmise that in urban areas, air pollution is the likely culprit. That suffices if we want to say that many things cause PD and leave room for agricultural use of pesticides. My next question is do the clusters in the South and West US reflect areas of high pesticide use or runoff. Next question is pesticide sprayed in those areas? I think for all of us the likelihood that the toxins are airborne is a pretty safe bet.

My bet on yielding answers has always been on the small groups where the odds are really rare. To that end, I ran across a study in Canada of three clusters. Overwhelmingly they all support how much our immediate environments play a part, more so than air quality outdoors. Is it the air quality inside buildings that counts? These clusters support this theory. All people who developed PD in these cases shared common work space; a poorly ventilated tv studio, a garment manufacturing facility, and a portable classroom over a filled in waste dump. BTW, I think that first group included MJF.

There is an even bigger takeaway here. From the Canadian researchers:

Furthermore, familial occurrence does not necessarily mean genetic causation, because family members share their environment and their genes. One study demonstrated that the risk for development of PD in a child from a parent-child cluster depended on the child's age when the parent started to show symptoms rather than the parent's age; younger children had greater risks. This finding suggests environmental risk factors.

John, these researchers toy with answering your question from awhile ago. If we remain in the environment that triggered PD are we doing further harm? Their conclusion surprised me:

[B]This pattern suggests an environmental cause that exerts its influence during a short time during the period of the shared environment and results in a cascade of events culminating in the clinical manifestations of PD after a long delay.[/B] One study15 used a mathematical model based on observations relating to clinical deficits to indicate the most likely pathogenesis is an event that kills some neurons and damages others in such a way that their life expectation is reduced, or an event that triggers a mechanism that kills healthy neurons at a constant rate. These models best explain the occurrence of PD clusters such as ours.

Clustering of Parkinson Disease: Shared Cause or Coincidence?
JAMA Network. 2004.

It also looks like latency of exposure period might yield some valuable clues. The TV crew was the youngest onset and had the shortest latency to PD of 9 years. High levels of Carbon Dioxide were found in air samples.

This makes me even more curious about those pockets of red on the US map and I am sure you have a correlate on yours.

Laura

for the u.s., you have to keep in mind that the factors that caused pd might have occurred years ago and have no relationship to where the medicare recipient currently lives. you tend to move to the city to be near healthcare/support services when you retire. just food for thought.

I had thought of that before too, Soccertese, and I just assumed that this had been accounted for in the epidemiology studies, but I don't think it has?!?? Basing it prescriptions and medicine means nothing than here are people now showing prevalence. This sort of information as presented is misleading. Without knowing where all these people lived decades before diagnosis, we cannot conclude much of anything on etiology. This is yet another vote for more cluster studies.

i look at myself as an example, there are numerous things that might have set the stage for my pd, playing football, working on a fishing boat and breathing diesel fumes and gorging everyday on salmon , taking a systemic fungicide for athelete's foot, working in a lab, chickenpox in my 30's. seems like with the lag time you have to code for all sorts of factors to find a correlation. the fact that it rarely runs in families or communities implies to me it has to be more a combination of causative factors and their magnitude.
you have to overwhelm your repair mechanisms.

i think resources better spent on early detection so you're closer to the "start" and can test neuroprotective treatments when they can help the most.

Don't throw the epidemiological baby out with the bath water of complexity.

The buses in London used to be mainly double deckers (that is they had two floors with stairs in beween). They had both a driver, who sat in a cab, and a conductor, who went around collecting the fares. It was noticed in the 1940s that drivers had more heart disease than conductors. We are all familiar now with the importance of exercise.

Is the situation with Parkinson's more complcated? Yes, of course it is. (If it were easy, we'd have the answers by now.) But, to our advantage we have computers that can process data billions of times faster than what people could do in the 40s. And, I think, more importantly we have vastly improved access to data via the internet and we have a huge number of people with the equipment, a laptop or PC, to do the research. So, it's now possible for someone in Stafford, or Seattle or Singapore, to access billions of pieces of Parkinson's data and data mine it. The probability that any one one piece of work finds something interesting is, I think, low. But if hundreds of people do it, I'm optimistic that something significant can be found.

Let's look at two of the issues raised in the replies.

Firstly, there's the issue of where a possible cause of the disease isn't included in the present study and goes undetected. An example of salmon eating was given. I have no idea of whether or not salmon eating has a role in the aetiology of PD. But, for the sake of argument, let's suppose it does. Would I detect that using the English data set? No. It would just be noise. But, that's a problem with the data set, not the methodology. In such circumstances, I suspect that an epidemiological study using data from the fishing villages along the coast of Washington, BC and Alaska would give some very interesting results.

Secondly, there's the issue of people moving house after contracting the disease. This tends to reduce the magnitude of the variations from place to place. But it is a matter of noise. Suppose we had two towns A and B, where A was 10:90 (PD, not PD) and B was 20:80 and suppose 20% of the people moved to the other place that would leave us with 12:88 at A and 18:82 at B. Interestingly, we would treat the relative prevalence of 1.61 as the real value when it's really 2. Effects such as this make it more important that we look for spatial variations.

John

john,
i was referring only to the medicare statistics, didn't read it thoroughly enough to see if any conclusions were drawn, not throwing the baby out with the bathwater. i fully appreciate the use of statistics but i also realize that you have to know the limitations of a study before you draw conclusions, especially when presenting them to laypersons.
and yes, i've heard of computers.

obvious clusters don't need major statistical studies , doctors would notice them, families would notice them, communities would notice them, professions would notice them. you see that with areas impacted by environmental pollution. and i imagine the cause and affect with those bus drivers didn't require a major statistical study.

i live in washington state which has the highest incidence of MS in the country, of course i want to know what's causing it as a person and as a taxpayer. and only statistics will find a disease influenced by multiple genes. just yesterday a correlation between autism and auto exhaust was reported, the closer you lived to a highway, the greater your chance of gettng autism. only a major study would have determined that.
http://www.globalpost.com/dispatch/n...aust-new-study