Dear all,
I want to relate sales growth in firm*city*year (FCY) cells to the firm's prior market share in that city, while controlling for firm*year fixed effects to capture amongst others changes in the firm's production costs.
I've started with Sergio Correia's -reghdfe- as follows:
But I'm still facing 2 challenges:
1. Sales growth in some FCY cells, in which the firm sold only say 1 unit initially, is super noisy such that getting reasonable summary statistics would require winsorizing or trimming more heavily than I feel I can easily justify. To start with I have persuaded myself that trimming (dropping the outliers entirely rather than replacing them all with say P90 or P95) is sensible. But instead of trimming even more, I chose to weight each cell observation in the regression by the number of sales observations I have inb that cell: :
As far as you can tell from the admittedly limited context I've given does that sound sensible or is there a better way to deal with outliers than trimming or weighted regressions?
2. From my understanding, the default way to compute standard errors given the absorption of firm*year fixed effects should be
. Unfortunately that kills the statistical significance at conventional levels for some of the coefficients. Does that mean the effect simply is not statistically significant, or would you see reasons for clustering differently, or bootstrap standard errors? In the latter case, how can I get around the constraint that "bootstrap: reghdfe..." does not seem compatible with the weighting?
Thank you so much,
PM
I want to relate sales growth in firm*city*year (FCY) cells to the firm's prior market share in that city, while controlling for firm*year fixed effects to capture amongst others changes in the firm's production costs.
I've started with Sergio Correia's -reghdfe- as follows:
Code:
reghdfe salesgrowth mshare, absorb(i.firmid##i.year) vce(cluster i.firmid##i.year)
1. Sales growth in some FCY cells, in which the firm sold only say 1 unit initially, is super noisy such that getting reasonable summary statistics would require winsorizing or trimming more heavily than I feel I can easily justify. To start with I have persuaded myself that trimming (dropping the outliers entirely rather than replacing them all with say P90 or P95) is sensible. But instead of trimming even more, I chose to weight each cell observation in the regression by the number of sales observations I have inb that cell: :
Code:
reghdfe salesgrowth mshare [aweight obs], absorb(i.firmid##i.year) vce(cluster i.firmid##i.year)
2. From my understanding, the default way to compute standard errors given the absorption of firm*year fixed effects should be
Code:
vce(cluster i.firmid##i.year)
Thank you so much,
PM
