search

jepusto / lmeInfo

Information matrices for fitted nlme::lme and nlme::gls models
https://jepusto.github.io/lmeInfo/
4 stars 2 forks source link

Incomplete Fisher_info(..., type = "observed") #53

Open DrJerryTAO opened 1 year ago

DrJerryTAO commented 1 year ago

Hi @jepusto,

I tried your Fisher_info() in {lmeInfo} and found that the function is not defined when type = "observed" although the help file specifies that "observed" is an option. Currently, it produces an error "object 'info' not found" because a script section for if(type == "observed"){} is missing.

jepusto commented 1 year ago

Thanks for pointing this out. As you noted, the package does not currently support observed information matrices. We will update the documentation to correctly represent this.

DrJerryTAO commented 1 year ago

I used varcomp_vcov(..., type = 'expected') to get standard errors of error correlation coefficient, variance structure parameters, and residual standard deviation (i.e. sigma) in generalized least squares and linear mixed-effects models. The default type is expected, and I noticed that the "expected" standard error is smaller than the "average" one for error correlation coefficient but larger for variance parameter and sigma. If available, which of these three types is best? Are they on the "natural" scale or some unrestricted estimands?

I understand that parameters like sigma have skewed sample distribution, so my plan is to generate a t ratio like sigma/se_sigma and use student bootstrapping to build confidence intervals. Do you know if this is a good use of the vcov of the variance component?