Is there a rule of thumb for the selection of the radius (cut-off) regarding spatial correlation? I use Conley standard errors in my models.

I could not find explanations on the internet and my economics professor told me that is quite 'arbitrary'.

CodePudding user response：

In my experience your professor is accurate that there is not specific guideline.

It makes sense when you consider that all data are not equal, and this is especially true with spatial data. Even within economics your data points can vary wildly within the same variable across different geographic spaces.

Think of a 100km cut-off in the heartland of the US Agricultural belt, maybe meaningful. Then think or a 100km cut-off at the convergence of the corners of Portugal, Spain and Monaco, where you have different cultural, industrial, political, religious and social factors forming their economies. It stands to reason that there might still be correlation across this area, but it might fall off more quickly.

Assumably you are using this because you know that there is is spatially influenced correlation, you might consider testing the strength of correlation at a variety of distances to come up with a starting point for this particular dataset.

There is no leakage in compensating for correlation, as their might be in building a predictive model, so it is meaningful to consider an iterative approach to solving for it if no good starting point is known.