I'm using CPS monthly individual data which comes with
wtfinl as a float weight.
I'm computing some annual statistic for unemployed. I do that by
- Sum up
wtfinlfor all the unemployed observations for a given year
- Divide all
wtfinlvalues by the annual sum, and then multiply the variable with this relative share
For example, if the variable is income, this gives me average income of unemployed.
Now, I would like to bootstrap the standard errors. I'm in Python. I
- draw samples from the full annual table, each row with equal probability
- Stop when number of rows is equal to the actual data
- I then apply the same estimator as before to the sample, each row weighted with
Now, when I compute the confidence bands, they're not centered around my actual estimator. Even worse, for one or two data points, the estimator is outside of the bands.
I suppose a alternative way to do the bootstrapping would be to
i. Draw from the annual table, each row with probability relative to
ii. Stop drawing once
wtfinl.sum() is equal to
wtfinl.sum() in original data?
iii. Then do all the computations without using weights.
I thought my way was better because (a) I use the exactly same estimator both times without changing weights, and (b) it is clearer: How would I do (ii) exactly? When I have room for
wtfinl=1 left, do I only draw from rows with
wtfinl <= 1?
However, given that my confidence bands dont line up with my original estimation, I'm befuddled. What should I do?