It varies with Reynolds number (rho .TAS. L / mu), which is a function of density. But, the dependency is small.
My suggestion would be that you look at the variation in "a" with Reynolds number, over the range of tested conditions, and include that (along with IAS accuracy, altitude accuracy, Nz accuracy...) in a reasonably robust error analysis of the eventual derived AoA. Using multiple test points, you can then plot through well argued error bars.
Look at the paper
here for one example of how to do the error analysis in a derived quantity. It's a bit labour intensive to derive, but then easily automated on a spreadsheet.
G