Is it possible to construct a difference-in-difference regression with only 4 data points, representing
For example, Wooldridge describes a setup where there is a random sample of units (e.g. firms) in two periods of time under two different treatments. Say, firms in California in 2012 and 2013 and firms in Wisconsin in the same years. Further, there is a change whereby, say California in 2013 changed their tax policy. So as he describes, it sounds like we are not tracking the same units per state in the two years but a different sample per state per year (i.e. 4 distinct samples).
This leads to my question: using the example above, can this same setup be used where you have a summary of the outcome variable for each state and year (not samples from each)? So, there are only 4 data points, say, total revenue of fast food restaurants in California in 2012 and 2013 (and the same for Wisconsin).
Thus the regression uses 4 data points, but clearly a standard error on the interaction term cannot be estimated.