How to compute and winsorize/trim growth rates

[Peter Meier]

Dear listers,

for an empirical analysis I need to compute year-on-year growth rates.

I have tried
CODE]gen growth_percent = 100*D.vol/L.vol [/CODE], g
gen

growth_sym = 100*D.vol/(0.5*L.vol+0.5*vol)

, and

gen growth_ld = 100*D.ln_vol

As some of these measures, in particular the percentage-based one, retain very big outliers, I have also tried to winsorize or trim at different percentiles.

While luckily direction and significance of the resulting regressions are very robust, point estimate sizes are not and I am unsure which formula and winsorization are the most appropriate ones a priori or which criteria I should use to choose.

Any guidance on this would be much appreciated!

Thank you, PM

[Nick Cox]

If growth rate has a very long-tailed distribution, which I do believe, the answer lies in a more suitable metric -- not in using growth rate and arbitrarily discarding the highest or lowest values because they are awkward for analysis.

For example, if the outcome is always positive, consider working on a log scale. But you seem to have done that already!