I have been provided a dataset of drink driving incidents aggregated into postcodes. This is where the incident happened, not where the driver lives.

Postcode 1 - 12 incidents
Postcode 2 - 163 incidents

This is for an entire state in Australia, so just showing total counts will pretty much be a population map.

Therefore I will need to do a bit of normalisation.

With postcodes not being equal-area, my obvious choices are either area or population (which I have the information for.

Population might not be accurate, as there might be factors like particular rural areas being notorious for drink driving due to no public transport, or other factors.

Area seems incorrect, as lets say we have two equal area postcodes, one rural and one urban, is it fair to normalise by area alone?

I then considered other options such as calculating the total lengths of roads within each suburb and using that.

I don't have traffic data, but in an ideal world, would using some sort of traffic count data be also an option?

The essential question I want the map to answer is, "Where are the hotspot areas for drink driving within this state of Australia?"

How would you go about this?

Final result http://au.news.yahoo.com/qld/video/watch/21428474/drug-and-drink-driving-hotspots-revealed/

The news guy did not quite follow the script/advice I provided, and I really wanted to avoid 3D, but you know media.

  • 3
    Whenever I see postcodes being proposed for use for anything other than improving the efficiency of postal delivery, my heart sinks because, while convenient, it is like using a hammer to drive home a screw - sure it will work after a fashion but the result will be nasty because that's not what a hammer was designed for. I don't think postcode districts are really the best way to do this. You may be forced to use postcode because lazy bureaucrats only collected the data in that way but if you can get access to the original point data, you can start doing proper hot spot analysis. – MappaGnosis Jan 30 '14 at 8:10
  • 1
    If available one possibility is to normalize by total number of accidents: 'number of driving incidents / total number of accidents' – Jens Jan 30 '14 at 8:13
  • @MappaGnosis - 100% agree, but yes. No point data collected. – Simon Jan 30 '14 at 9:20
  • @Jens - Dont have this data, especially not specific accidents caused by drunk drivers. Dont want to debate, but I dont think this would be the best to normalise by. Could be that loads of people drink drive somewhere, but never crash... – Simon Jan 30 '14 at 9:22
  • 1
    There is a probably interesting, related question on Stackexchange Open Data: How to normalize the data when mapping crime reports? – lavarider Apr 25 '14 at 19:36

The data, drink driving incidents per postcode, comprises information of very different types. Some of which correlate, and some considerations arise:

  • You are dealing with absolute values (number of incidents). You may prefer relative ones: Like a percentage of cars with drunken drivers with respect to the total number of cars on the road. I think the goal is to approximate a value like this as best as we can, since it would show the desired hot-spot areas pretty well.

  • Even if more detailed data is not available, it would be helpful to know: How were these incidents sampled? The more precisely we know this, the more precisely we can normalize the dataset. Ideally, we would know the total number of drivers that have been controlled by traffic checks. Obviously, not every single car was checked for drunken drivers, so your data depends on which cars have been picked for checking: For example, if police picks cars randomly at a certain road, you would use the traffic count of this particular road to calculate a percentage. (Depending on sampling style, your data may be under-represent certain roads, time of day etc. We probably do not know these details, but we should not increase the bias of the data by incautious normalization. Let's assume that your data is unbiased.)

  • Most likely you do not know sampling details mentioned above. - But which data resembles the conditions mentioned above best? I think an (unbiased) total number of cars on the roads, i.e. absolute traffic count data would be perfect.

  • If absolute traffic count data is not available, you could use proxy data instead. However, try to find a test area where some kind of traffic data is available, so you can evaluate your approximation. It does not have to be your sample area, just a place somewhere with great data availability to cross-check your method.

  • To look for appropriate proxy data, think of what correlates with absolute traffic. You came up with the total length of roads, I think maybe a road density is also interesting (i.e. road length divided by area), because it does not introduce area correlation. Then combine this with population to get a traffic proxy (normalize by road density * population). Just an idea, though.

  • You should not normalize by area alone, since traffic intensely correlates to population: you stated that showing total counts resembles a population map. On the other hand, you fear that normalization by population is not accurate. Probably, normalizing by population alone is not a bad start, and a better traffic proxy than area alone . However:

  • Discover correlation in a more impartial way: Find out which data correlates with other data and calculate correlation coefficients. (There are lots of methods, Pearson's r is a start which can easily be calculated with Excel's built-in functions for example.) Make a table with all available types of data (incidents, population, area, road length, road density, etc - and, for reference, traffic, if available). Calculate correlation for each pair of data types.

  • Do not normalize against weird things. You have to normalize against traffic (or a traffic proxy) first! Also, be careful not to normalize against data, where you want to "proof" correlation later on (e.g. by presenting some type correlation coefficient): For example, do not normalize against number of pubs if you want to map if many people drive drunk near pubs.

For visualization of your map, another GIS.SE thread come into my mind: What Makes a Map Beautiful? - A value by alpha map could be very interesting for this kind of data.

| improve this answer | |

Well, I think part of that may depend on exactly what you're looking to get out of the results, but if you're not careful in normalizing, you could also skew the results toward what you want it to show instead of what the facts show.

For example, if you want to know if more people drive drunk in one area than another to help police know where to spend their time patrolling and enforcing drunk driving laws, then you might normalize by road length because police are probably going to enforce such laws by sitting along the road way and monitoring a situation. This way they would know per mile of road, which area is going to have the most offenders and they can spend limited police resources patrolling the highest risk areas.

However, if you are trying to understand the data to help a public education campaign know where to spend limited money on education efforts to curb drunk driving, you may want to look at normalizing by population so you target areas where a greater percentage of the people targeted are likely to be drunk drivers.

Or, if for example you are trying to work with the Bars, clubs, restaurants, etc... that serve alcohol to institute programs that reduce drunk driving, then you might want to see if you can find any data (maybe from some sort of alcoholic beverage regulatory agency if you have one there) that would help you normalize by places that sell alcoholic beverages. That way you would know which area's businesses are the likely the greatest cause of drunk driving.

These are just some thoughts that came to me off the top of my head, but I hope they get across the idea.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.