Bootstrap Confidence Intervals BCa

[JB]

Dear all,
I need bootstrap confidence intervals for a model. I do not know a lot about it but from what I have read, among the best options are the bias-corrected percentile method and the bias-corrected accelerated percentile method.
If I am correct, the bias-corrected corresponds to the boot.ci.type="bca.simple" in lavaan? I'm a little confused by the meaning of the a, standing for adjusted or accelerated. The boot.ci function in package boots says the BCa is adjusted bootstrap percentile and in lavaan: "bca.simple" option produces intervals using the adjusted bootstrap percentile (BCa) method, but with no correction for acceleration (only for bias)." Are the bca.simple in lavaan and BCa in boot.ci the same and correspond to the bias-corrected percentile method without acceleration, or is the BCa in boot.ci also accelerated?
The bias-corrected accelerated method seems better when boostrapped parameters are skewed(?), which is my case. However, it is not implemented yet in lavaan and seems very complicated . I found a reference that seems to provide a simpler way to compute bias-corrected accelerated but it is way above my math and coding abilities. But if anyone is interested, here is the reference:
Diciccio and Romano, 1988, A review of bootstrap confidence intervals. Journal of the royal statistical society. 50(3), 338-354.
Many Thanks!!!
JB


 

[yrosseel]

On 02/18/2014 01:38 PM, JB wrote:
> If I am correct, the bias-corrected corresponds to the
> boot.ci.type="bca.simple" in lavaan?

Not really.

> function in package boots says the BCa is adjusted bootstrap percentile
> and in lavaan: "|bca.simple"|option produces intervals using the
> adjusted bootstrap percentile (BCa) method, but with no correction for
> acceleration (only for bias)."

Exactly. So they are not the same. The "bca.simple" option was done to
mimic the behaviour of Mplus. Full BCa is currently not implemented in
lavaan.

Yves.