Since you have 100,000 points there is no real meaning to your claim that there is insufficient data. Your aggregation of the data has produced a data set that is no longer useful. But dont throw away all the data!

I think you need to have a better idea of what you are trying to do and to seek the best way to do it. When you said you had gridded the data into cells, I was expecting perhaps quadrat analysis (how are the points in the cells). Say all the diseased trees are in one cell -- very non random, very indicative of negative auto-correlation ... and not too difficult to discern. The issue would be if there are a spatial distribution of diseased trees among the cells. Is your goal to find a spatial process (contagion?) for the disease process? How do these results change as you modify the spatial grids?

What about k-functions -- measuring the distance to nearest neighbors, and using what is called a marked process (mark = HEALTHY, or DISEASED).

Right now you are holding onto some ideal data analysis instead of planning what you may be able to do with the data.

 Hi!

I agree with Mr O'Kelly. By gridding the data you are in danger of compromising your results due to the MAUP (Modifiable Areal Unit Problem). It is important to know what you are aiming for.

A good first step would be to try performing a point pattern analysis, to check whether your data follow a clustered, random or dispersed distribution. This can be done by testing for CSR (Complete Spatial Randomness) by performing tests such as the Quadrat Method, Nearest Neighbor Index or G-function. Remember always to do a significance test for the produced result of every method. 

Since the Quadrat Method as well as the Nearest Neighbor index have issues regarding MAUP and, my suggestion would be to use the G-function. But, that also depends on what you are aiming for. If you choose the Quadrat Method, you will have to test the effect the size of the quadrat has on the result. Does the type of distribution change, if you change the quadrat size? What is a representative quadrat size for your study? etc. Moreover variations of the positioning of points within the quadrat or between neighboring quadrats will not be accounted for if you choose this method. The Nearest Neighbor Index will only detect clustering on a fine spatial scale, since only the nearest neighbor of every point is considered.

You can find more info about the G-function here: http://www.people.fas.harvard.edu/~zhukov/Spatial4.pdf

As a second step, you could compute a correlogram or variogram of your data, to check for spatial autocorrelation. By examining either you could see, if you indeed have spatial autocorrelated data, and, if this is the case, at what distance does the spatial dependence cease to exist. This can be done by checking the Sill and the Range (the distance at which the curve reaches the sill) of your function curve in the variogram. This defines the maximum distance at which there is any spatial dependence. The distance between your points in every grid should not exceed this distance. 

Another way to determine the size of your grids, after you have checked the above, would be to see if your data is clustered by implementing a non-hierarchical clustering method such as K-means or K-medioids (since you have a big dataset). If you find any clusters, you could use the within cluster distance between your points to help you determine a suitable grid size.

Hope this was helpful.

Good Luck!

Best regards,

Karolina