I have spatial point data (say addresses of all inhabitants of a metropolitan area or larger region) with context data for each point like age, gender, employment status, and so on.
For each spatial point, I would like to generate aggregate information about the "neighborhood" environment (with regard to age, gender, employment status), using kernel density estimation (KDE).
E.g., I would like to generate the share of unemployed persons in the vicinity of the spatial point.
So far I can do the kernel density estimation, generate a polygon map and attach the corresponding estimate to each coordinate pair of the point data. My research aim is use the characteristics of these small-scale neighborhood Environment as explanatory variables in regression models at the Level of individual persons.
Two issues are troubling me, though:
- I have point data from denser city centers but also more rural areas. How can I generate comparable KDE-estimates for the whole study area?
Using a fixed bandwidth for the whole study area will be suboptimal for all contexts.
Will using adaptive bandwidths generate comparable estimates, though?
- So far I only receive the estimated number of unemployed from the KDE.
Since this is a function of population density, I guess I have to divide the estimated number of unemployed by the estimated number of the total population, right?
If I do this, do both KDE have to use the exact same bandwidth to be comparable?
Some events like unemployment might be rare (= less dense point patterns) and thus would need to have a larger bandwidth, while the spatial patterns for the whole population should be denser and thus need a lower bandwidth. A common bandwidth could thus be suboptimal for one or both.
Should I compromise with a mean bandwidth or pick the largest common bandwidth or are KDE from varying bandwidth still comparable?
I hope I could convey what I want to do and why I want to do it. MDo you think my approach is feasible or do you have an alternative proposal?
I am new to KDE.