Shang Ha writes,

I am helping one of my colleagues draft a policy memo. He uses a cross-sectional time-series dataset, composed of about 30 monthly/bimonthly surveys (n is around 1,300 per each) in California.

His dependent^H^H^H^H^H^H^H^H^H outcome variable is individuals’ attitudes toward state government policies. Individuals are nested in counties and county-level variables (e.g., racial composition) are also included in the models. Furthermore, he gathered a variable on media coverage (on policy issues) per month, matched with each survey.

It seems like this data have three-level (level-1: individuals, level-2: counties, and level-3: monthly surveys). Can we run a simple three-level regression with this dataset? I am particularly uncertain how we can deal with autocorrelations at level-3. If you know any examplary papers using similarly structured data, please let me know.

My reply: Yeah, this is a big problem. I think you can run a 3-level (non-nested) model, for example using lmer() or Stata or even Bugs. I’d start by including varying intercepts (and maybe slopes) by county and by time, and then also including a linear time trend. You can then look at residuals and see if more is needed. In Bugs you can fit an autoregressive model without too much trouble, but I’d start by seeing if you can get away with a trend. No exemplary papers yet; we’re trying to write one!