Lets say I am running a regression on annual data at the county level and I would like to include some controls based on census data.
Some of my counties of interest are below 65k population, so ACS data is only available in the form of 5-year estimates.
How should I go about integrating the 5-year estimates with the rest of my data? I see several possibilities:
First is to use 1-year estimates where available and 5-year estimates for small areas. This seems wrong to me as numerous sources caution against comparing the two, so I will take it off the table.
That leaves me with using the 5-year estimates everywhere, and gives me several options for how to merge the data.
- Take the last year of each 5-year period, and then carry the earliest value backwards for the first 4 years. (or the reverse with the first value)
- Take the midpoint where available and then carry the end value forwards for the last 2 years and the beginning value backwards for the first 2 years.
- Take non-overlapping periods. E.g. use 2006-2010 and 2011-2016 and throw out the overlapping results.
- Fit a simple model to the 5-year estimates and use the model to predict annual values.
I am leaning towards #1 since it mostly avoids introducing future information into an annual observation, but I have seen multiple official publications that use the 2008-2012 5-year estimate as a stand in for 2010 when comparing to prior decenial censuses (which leans me towards #2). I don't like how much trend information #3 throws away and #4 seems overly complicated just to introduce some control variables to a model.
Are there any papers or textbooks that have already answered this question? My own searching has been unable to turn up anything on how to best use ACS data in an annual time-series context. Instead I have just found things focusing on choosing an estimate for a single area or making comparisons (but not when combined with existing annual data)