# At what level should I cluster my standard errors and why?

I have a yearly panel data in which each observation is a pair of monitoring stations (stations measuring water quality in rivers) one located upstream and the other downstream, each station in the pair is in the same river and in a different but neighboring county. I am running the following regression:

pollution^{downstream}_{i,t} = \beta_1 pollution^{upstream}_{i,t} + \beta_2 \text{county alignment}_{i,t} + \gamma_i + \lambda_t + u_{i,t}

Where $$pollution^{downstream}_{i,t}$$ is the pollution measured by the downstream monitoring station in station pair $$i$$ at time $$t$$, $$pollution^{upstream}_{i,t}$$ is the pollution measured by the upstream monitoring station in station pair $$i$$ at time $$t$$, \text{county alignment}_{i,t} is my "treatment" and is a dummy variable which equals 1 when the mayor in the neighboring municipalities where the pair of stations is located belongs to the same political party. Lastly, $$\gamma_i$$ and $$\lambda_t$$ are station pair fixed effects and time fixed effects.

I am trying to figure out at what level should I cluster my standard errors and the intuition behind it in order to get the right conclusions about my treatment coefficient ($$\beta_2$$). Right now I am clustering at the station pair level, but I am not sure if I should cluster at the downstream station level, county pair level, or downstream county level? Which level do you think I should cluster on and why?

New contributor
Quinoba is a new contributor to this site. Take care in asking for clarification, commenting, and answering. Check out our Code of Conduct.