I provide statistical support for a public health department. As you might imagine, we put together a lot of maps on a regular basis. For me, maps are just another kind of data visualization - useful for getting a feel for the data, for generating and checking hypotheses, etc. But we don't often follow through on actual modeling and hypothesis testing.

How do you/your organization go about this? What does a workflow that includes inference look like? Who's involved? What tools do you use? What would it ideally look like, if you had your way?



To be clear, I'm curious about different strategies for going from spatial data to formal, statistical tests of hypotheses about what's going on in the world. For example, let's say I'm trying to target an educational campaign to increase tuberculosis testing. I (personally) would map out the cases of TB against covariates of interest (say, median income or percent foreign-born residents) and try to see if there were any patterns.

I might or might not find any; but I would ultimately build a model to estimate the association between those covariates and the number of demographics. This is a critical step because of how good humans are at finding patterns where none exist, or finding uninteresting ones. I know how to do this on my own, but I'm curious about how different organizations institutionalize it (if at all).

  • Great question! – whuber Dec 21 '10 at 17:19
  • Are you saying that you need to have a workflow so that if there's an outbreak of some disease for which a limited supply of vaccine is available, you need to be able to show that you are optimally distributing the vaccine? – Kirk Kuykendall Dec 30 '10 at 16:34
  • Broadly, I'm just interested in how people incorporate statistical inference into their mapping processes. What you describe is certainly one possible scenario, but there are lots of others and I'm not even especially interested in responses from epidemiology. – Matt Parker Jan 3 '11 at 22:16

Very interesting question!

Firstly, your question alludes to what I term 'data mining' and I think its worth restating the problem explicitly as some people here may not have got it: with any data set (doesn't have to be spatial) to achieve a statistically valid relationship the convention is that it must be at or above 95% probability. However, if you do 20 tests then chance are high that at least one of the 'statistically valid' results you obtain is due to pure chance. So its bad practice to play around with a data set (in GIS it would be mapping it out) to visualise many possible relationships between variables, find an interesting one and plug in the stats and quote the result as if this was the only test you had done. You can still use the result but you have to take account of the number of tests you've done.

Is that what you were driving at?

Your question appears to ask how people formalise avoiding this problem. My answer is that the 'not at all' option you mention is common. Medical statisticians (e.g. my girlfriend) in my experience apply a much higher standard of rigour to this sort of process than is found in other areas, I suspect all sorts of data mapping outside of public health is done without any sort of formal consideration of the problem with stats formulas being blindly applied without understanding the process properly. A geological example comes to mind:

I read a peer reviewed paper where authors looked at how borehole yield (amount of water that could be pumped) related to geological and spatial influences in Africa e.g. thickness of gravel layer that was dug through before bed rock was hit. The idea was to help borehole drillers so that they could target the best locations for boreholes. The authors blatantly mined the data combining all sorts of variables to see which ones came up with a 95% confidence level and (I assume) none of the reviewers had questioned the validity of the results. Their conclusions were therefore completely untrustworthy.

Hope that's of interest

  • Can you explain a little more why the paper you describe is untrustworthy? It's not obvious to me why this is the case. If the relationship exists statistically, does it matter what 'mental model' you used to achieve identifying it? I understand that it does not explain the mechanism, but that is a separate issue. – djq Jan 30 '11 at 12:27

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.