# Regression with autocorrelation given by a graph

Supppose I would like to fit a standard regression model of the form $$y = \alpha + X \beta + \epsilon$$, but my observations come from a networked population (eg. social network). I have the graph $$G$$ representing connections between observations and it's reasonable to assume that errors will be autocorrelated between connected observations.

How can I get valid standard errors of the $$\beta$$'s ?

• post your data ........ – IrishStat Jul 31 '19 at 20:46
• I don't have any data which I could share. I don't really see what would it bring. I think the setup is quite easy to imagine: just an ordinary table with accompanying graph representing links betwen observations from the table. Alternatively, you can think of it as an attributed graph with each vertex $v_i$ of the graph having a vector $x_i$ and a scalar $y_i$ as attibutes. It's a question which came to my mind while I was working on one of the projects at my work. – Wassermann Aug 1 '19 at 17:19