I work on my research in finance concerning pricing of green bonds and I am running a two stage model.
Stage 1 regression is an unbalanced panel fixed effects estimation. For each of my 100 green bonds in the sample, I compare its daily ask yield to an ask yield of a similar conventional bond and estimate the difference while controlling for the difference in their liquidity:
(yield.green - yield.conventional) ~ (liquidity.green - liquidity.conventional)
I am particularly interested in the individual effects captured by the intercepts - for each of my 100 green bonds, I save the estimated individual fixed effect which tells me if that bond sells for a higher price than its conventional counterpart. I call this fixed effect a "green premium".
Stage 2 model is on OLS regression of "green premium" on bond's characteristics, e.g.:
green.premium ~ maturity + issue.amount + currency + sector
Now the problematic part comes: Since my dependent variable "green premium" is an estimate coming from the Stage 1 regression, I need to deal with heteroskedasticity, although Breusch Pagan test does not reveal any signs of heteroskedasticity.
Is any of these two remedies a good solution for my case:
Use Weighted Least Squares - for this case, shall I use the inverse of the squared standard error of the indivudual fixed effects from Stage 1 - 1/(SE^2)- as the weightes?
Use White's SE?
