4

I have to test which factors influence game damage in fields. I mapped areas with damage and those without. It was not always possible to map 100% of a field, so there are also areas where it is unsure. Since the "unit" damage is not objective because it is not possible to determine where one ends and the next starts, I put a grid over the area and calculated for every cell independently the distance to different structures (forest, roads etc.). The resulting data look like this:

| damage |  id | dist_forest | dist_maiz | dist_roads |...
|0       |   51|           30|         20|          70|...   
|0       |   51|           20|         10|          60|...   
|0       |   52|           60|         10|          80|...   
|0       |   52|           40|         70|          10|...   
|0       |   52|           20|         60|          50|...   
|1       |   53|           10|         10|          50|...   
|1       |   53|           05|         20|          30|...   
|1       |   54|           20|         30|          20|...   
|1       |   54|           30|         20|          90|...   
|1       |   54|           40|         10|          10|...

(I have about 100 individual polygons which lead to 100000 lines when resolved in square meters)

I wanted to use a binary logistic regression with random effects. To solve the problem with the non-independent data I was adding the id of the damage-polygons as a random factor. The resulting model was like this:

glm <- glm(damage ~ dist_forest + dist_maiz + dist_roads + (1|cat), family=binomial(logit),data=data)

The problem ist now, that all my parameters are all highly significant. I already asked that here and it was suggested to use a special model for spatial data.

if data are (substantially) spatially dependent then standard significance tests don't apply any way!

Does anybody have any suggestions how to proceed further?

Community
  • 1
meles
  • 151
  • 1
  • Just because your data are spatial it doesn't necessarily follow that they are spatially dependent. For an explanation of spatial dependence see http://stats.stackexchange.com/questions/18406/what-is-the-difference-between-spatial-dependence-and-spatial-heterogeneity. – Adam Bailey Jan 16 '14 at 07:51

2 Answers2

2

You may want to try something like Indicator Kriging. Where you plot the locations where "damage" occurred and then run a kriging model. Then, you can extract the calculated probability of "damage" at the location of each feature and fit a logistic model based on distances to the remaining features. This way, you know that the distance to one of your features is 0 so you are effectively controlling for a feature in each instance.

Anyway, the above is just a sketch of some spatial modelling you could do. I'd look into Kriging/Geostatistical or spatial stastical models if you are modeling geometric relationships.

0

I solved the Problem now with a Generalize Additive Modell (GAM). I was including the term s(lat,long) to consider the spatial dependency of the points.

meles
  • 151
  • 1