Parameter Estimates:
Standard errors Standard
Information Expected
Information saturated (h1) model Structured
Standard errors can be only either standard or Robust, correct?Information can be only either "Expected" and "Observed". Given that we can manipulate that, under what circumstances would I choose one or the other?What is this "Saturated (h1) model"?? and what does it mean "structured"?
Thank you for your valuable help.Kind regardsMichael.The baseline model is the model that only includes intercepts and variances.
Example:
library(lavaan)
data("PoliticalDemocracy")
d <- PoliticalDemocracy
model <- '
# measurement model
f1 =~ x1 + x2 + x3
f2 =~ y1 + y2 + y3 + y4
# regressions
f1 ~ f2
'
fitted_ml <- lavaan::sem(model, d)
summary(fitted_ml, fit.measures = TRUE)
Part of the output:
Model Test Baseline Model:
Test statistic 406.880
Degrees of freedom 21
P-value 0.000
We can fit the baseline model ourselves by specifying intercepts only (watch out for the defaults - I think that you'd need to specify more if you used cfa()):
baseline_model <- '
x1 ~ 1
x2 ~ 1
x3 ~ 1
y1 ~ 1
y2 ~ 1
y3 ~ 1
y4 ~ 1
'
fitted_baseline <- lavaan::sem(baseline_model, d)
summary(fitted_baseline)
This model just estimates intercepts and variances:
Model Test User Model:
Test statistic 406.880
Degrees of freedom 21
P-value (Chi-square) 0.000
Model Test Baseline Model:
Test statistic 406.880
Degrees of freedom 21
P-value 0.000
summary(fitted_ml, fit.measures = TRUE)
And part of our output is:
Parameter Estimates:
Standard errors Standard
Information Expected
Information saturated (h1) model Structured
If we fit with a robust estimator (e.g. MLR):
fitted_mlr <- lavaan::sem(model, d, estimator = "MLR")
summary(fitted_mlr, fit.measures = TRUE)
We get:
Standard errors Sandwich
Information bread Observed
Observed information based on Hessian
MLR is a robust method, and is a type of sandwich estimator. You might have come across sandwich estimators in regression when they are used to account for clustering and heteroscedasticity of variance (i.e. when residuals are not i.i.d, using the sandwich package in R, they’re also called robust estimates or Huber-White estimates). There are some equations in this presentation http://www.gvptsites.umd.edu/uslaner/robustregression.pdf.
If ML is used (and assumptions are satisfied, and the sample is large) the expected and observed information converges to the same value. Expected information is faster to calculate and more stable. Sandwich estimators use the inverse of the model information (the ‘bread’) and an estimate of their variance (the meat / tofu) [see the presentation for why it's a sandwich].
At least that’s my understanding.
You can ask lavaan to use expected information with robust estimates:
fitted_mlr_expected <-
lavaan::sem(model, d, estimator = "MLR", information = "expected")
summary(fitted_mlr_expected, fit.measures = TRUE)
Standard errors Sandwich
Information bread Expected
Information saturated (h1) model Structured
--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/lavaan/CACXLW8_KSj5ZwC8dz4WXspe4vY9QUoHnWPYoSgRWEyeevsn5NA%40mail.gmail.com.
To view this discussion visit https://groups.google.com/d/msgid/lavaan/CAMtGSxmx0zzHvkiHM%2BMsCBe%3Dc9A-Tu63rTqBWV4ojcGcQass6g%40mail.gmail.com.