bca.simple option in parameterEstimates function

[Nika]

Hello,

I have a question on bca.simple option in parameterEstimates function. It says in the description that "The "bca.simple" option produces intervals using the adjusted bootstrap percentile (BCa) method, but with no correction for acceleration (only for bias)." I am a bit confused with the difference between that and the "regular" BC method, because I thought that the difference between BC and BCa is the correction for acceleration?

Another question is that I am using MLR estimator that uses robust standard errors (but doesn't influence parameter estimates, or so I understood) - is it ok to also use bca.simple method for estimating confidence intervals, or should I report the original ones?

Thank you in advance!

[Terrence Jorgensen]

I thought that the difference between BC and BCa is the correction for acceleration?

Yes.  The boot package offers BCa intervals (corrected for bias and acceleration).  I don't recall why, but lavaan only offers bias-corrected intervals

.

Another question is that I am using MLR estimator that uses robust standard errors (but doesn't influence parameter estimates, or so I understood) - is it ok to also use bca.simple method for estimating confidence intervals, or should I report the original ones?

I'm not sure why you need bootstrap intervals if you already requested robust SEs, but it is correct that MLR point estimates are the ML estimates.  The robust adjustment only changes SEs and chi-squared.

Terrence D. Jorgensen

Postdoctoral Researcher, Methods and Statistics

Research Institute for Child Development and Education, the University of Amsterdam

UvA web page: http://www.uva.nl/profile/t.d.jorgensen

[Nika]

Thank you for your quick reply!

Regarding the BC estimates, I thought as much, but the explanation provided (which I quoted earlier) is a bit confusing - thanks for the clarification.

As for the bootstrap intervals, I read somewhere that they are a better approximation of the confidence intervals than the ones calculated using robust SEs. My question was actually whether this was "cheating" in a way. I'm not sure how the math behind the program works - are the bootstrap intervals calculated using robust SEs and if so, would this mean that they were first "corrected" using robust correction and then again using bootstrapping. If that is the case, it sounds like "over-correction" to me?

[Terrence Jorgensen]

As for the bootstrap intervals, I read somewhere that they are a better approximation of the confidence intervals than the ones calculated using robust SEs.

Not necessarily.  Bootstrapping is a common recommendation for indirect effects only, since products of normally distributed statistics are not normally distributed.  If you don't want to test indirect effects, robust SEs would only be necessary if data are not normally distributed. robust SEs are more efficient than bootstrapping because the latter adds Monte Carlo error on top of sampling error.  If you do want to test indirect effects, I would still just use robust SEs for all other tests, but you can use the Monte Carlo CI method for testing the indirect effects.  That method at least allows you to minimize Monte Carlo error by using 10,000 or 100,000 samples, and it doesn't take hardly any time because it doesn't require the model to be fit each repeatedly like bootstrapping does.

http://quantpsy.org/medmc/medmc.htm

You can get the required information from your fitted lavaan model.  For example, if your paths are labeled "XM" and "MY", you can extract them:

coef(fit)[c("XM","MY")]

vcov(fit)[c("XM","MY"), c("XM","MY")]

are the bootstrap intervals calculated using robust SEs

Bootstrap CIs are based only on the bootstrap distribution of point estimates, not estimated SEs (robust or otherwise).

[Mikko Rönkkö]

Hi,

On 17 Jan 2017, at 8.01, Nika <nika....@gmail.com> wrote:

As for the bootstrap intervals, I read somewhere that they are a better approximation of the confidence intervals than the ones calculated using robust SEs. My question was actually whether this was "cheating" in a way. I'm not sure how the math behind the program works - are the bootstrap intervals calculated using robust SEs and if so, would this mean that they were first "corrected" using robust correction and then again using bootstrapping. If that is the case, it sounds like "over-correction" to me?

For a counter argument to that, you may want to also take a look at these two articles:

Koopman, J., Howe, M., & Hollenbeck, J. R. (2014). Pulling the Sobel test up by its bootstraps. In C. E. Lance & R. J. Vandenberg (Eds.), More statistical and methodological myths and urban legends (pp. 224–244). New York: Routledge.

Koopman, J., Howe, M., Hollenbeck, J. R., & Sin, H.-P. (2015). Small sample mediation testing: Misplaced confidence in bootstrapped confidence intervals. Journal of Applied Psychology, 100(1), 194–202.

Mikko

[Alex Schoemann]

If you're interested in Monte Carlo CIs with a lavaan model the function MonteCarloMed in the semTools package can compute Monte Carlo CIs and pull the required information from a lavaan object (assuming you used parameter labels in your model specification).