Closed ajdamico closed 1 year ago
SSW (1992, Sec. 5.2) has a table that shows the impact of the Bias Ratio ($\text{BR} [\widehat{\theta} ] = \frac{\text{B}[ \widehat{\theta} ]}{\sqrt{\text{Var}[ \widehat{\theta} ]}}$) on confidence statements. As long as BR < 20% and the normal approximation is valid, impact is minimal.
However we work with something else. PRB is not ideal for that either, but it is a good measure. The one we use is what I call Squared Bias Component, the fraction of the MSE attributed to Squared Bias. If that is below 5% (or 10%), we should be good.
This can recreate the table in SSW:
# bias ratio, bias component and coverage probability functions
br.fun = function(sbc) sqrt( sbc / ( 1 - sbc ) )
sbc.fun = function(br) br^2 / ( 1 + br^2 )
p0.fun <- function( bratio , alpha = .05 ) {
z.calc <- qnorm( 1 - alpha/2 )
lt <- pnorm( z.calc - bratio , 0 , 1 , lower.tail = FALSE ) # lower tail
ut <- pnorm( - z.calc - bratio , 0 , 1 , lower.tail = TRUE ) # upper tail
1 - lt - ut # complement
}
# recreate table
br.vec <- c( 0 , .05 , .10 , .20 , .30 , .50 , 1 )
round( cbind( "BR" = br.vec , "SBC" = sbc.fun( br.vec ) , "P0" = p0.fun( br.vec , .05 ) ) , 4 )
# BR SBC P0
# [1,] 0.00 0.0000 0.9500
# [2,] 0.05 0.0025 0.9497
# [3,] 0.10 0.0099 0.9489
# [4,] 0.20 0.0385 0.9454
# [5,] 0.30 0.0826 0.9396
# [6,] 0.50 0.2000 0.9209
# [7,] 1.00 0.5000 0.8299
This gives an idea about the effect of the bias on confidence statements. However, this is true when the normal approximation is valid -- i.e., large enough sample size for a CLT to hold. If the normal approximation is not good, these coverage probabilities will be misleading.
One solution is to drop PRB (it is used in the zenga paper, I believe) and just go with SBC. Requires some rewriting, but not too problematic.
so just going with this change? https://github.com/guilhermejacob/context/pull/24/commits/6193f27ad673bac80fbe906e0c3c84e18832b5e5
you say <1% in this part of the book.. but
watts pov. gap ratio
andtheil(poor)
are at 4% and 3% ...is that expected? does it make sense to clarify how/why the PRB below 1% only applies towatts
andfgt0
?For the variance estimators, we estimate the PRB using the code below. Note that the bias is still relatively small, with absolute values of the PRB below 1%.