I tried to look at whether there is cluster of psychiatric patients in a county. I have census tract data with a column showing how many psychiatric patients are there in a census tract. There are many (more than a half) census tracts without any psychiatric patients.

I used both fixed distance band and spatial weights matrix to do the hot spot analysis. When used fixed distance band, I used spatial auto correlation to generate the distance, and I chose the max peak, (because the when first peak occured, some features had no neighbor). Both fixed distance band and spatial weights matrix methods yielded the similar result (tracts with 0 value were marked as hot spots).

When examine the result, I found that many polygons marked as hot spot were actually with 0 value, which should have been cold spots. I understand that hotpost analysis calculates the values within a certain defined areas instead of the original polygons, but still don't totally understand the result and how to explain it.


1 Answer 1


Remember that with Hot pot Analysis it not the individual unit (census tract in your case) that is the hot spot, it is the tract and all of its neighbors (as defined by the conceptualization you choose) that constitute the hot or cold spot - as a group their "sum" of psychiatric patients is significantly higher than the number of patients expected in a group of that many tracts by chance. Therefore, the center value can indeed be a zero value if it is surrounded by tracts with high or low enough values to create a significant sum.

  • Thank you very much! But does it me I didn't use the appropriate distance threshold when doing the analysis?
    – Elaine
    Sep 17, 2014 at 20:39
  • The distance threshold you use depends on the nature of your data. If you expect to see clustering tracts with high or low numbers of psychiatric patients within a certain distance - for instance, if you expect that clustering won't occur past a radius of 50 Kilometers around any one center point, than a fixed distance of 50 Kilometers or less might work, but really you should test this a range of different fixed distances, because your data may show different trends depending on how you define the neighbors for the test.
    – GISdan
    Sep 18, 2014 at 4:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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