Plausible Values in SemTools

[Ale Dea]

Dear lavaan users,

I am using fitting a factor model in lavaan, then I am using the function plausibleValues to get 5 plausible values for each individuals. I would like to run regressions with these plausible values with them on the RHS. Something like

Y = a +b*PlausibleValue + epsilon

Is there a command that does this? Or shall I run this regression for 5 times, one for each plausible value and then average over the 5 coefficient b that I obtain from each regression to get the b coefficient? In stata you can do this with repest (https://ideas.repec.org/c/boc/bocode/s457918.html)

Thanks

[James Ades]

Are the plausible values from imputation? There is a function in SJmisc that will create a kind of proxy score for those five values that enter into the regression. I also think there is something in MICE that will handle separate values (in something like lme4); I don't know if this extends to Lavaan.

[Mauricio Garnier-Villarreal]

After estimating th plausibel avlues, you run the models like multiple imputations. You can use the sem.mi()/cfa.mi() functions from semTools to run the model for how many plausible values you exported. This way you get average estimates, and SE that account for variability

Plausible values are "kind" of multiple imputations for factors, accounting for variability due to factor indeterminancy. But not exacly imputations as the factor scores are completelt missing

[Ale Dea]

This is very useful, so it is possible to analyze them using the command sem.mi() which is meant for imputed values (but plausible values are always imputed given that the factor score is unobserved - i.e. missing)

Ho does sem.mi() correct the standard errors? I thought I had to bootstrap the entire procedure, namely factor model, plausible values, and regression to get standard errors

[Mauricio Garnier-Villarreal]

By treating the plausible values as missing, we use the multiple imputation Rubin's rules for combining estimates and SE. For estimates its the average across data sets, for SE it adjust/penalize them for between data sets variability.

You can lok at Enders (2010) for details about the MI combining rules

Enders, C. K. (2010). Applied missing data analysis. Guilford Press.

[Terrence Jorgensen]

so it is possible to analyze them using the command sem.mi() 

Yes, see the multiple-group example on the ?plausibleValues help page.

Terrence D. Jorgensen

Assistant Professor, Methods and Statistics

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

http://www.uva.nl/profile/t.d.jorgensen

[Ale Dea]

I have noticed that plausible values are not available for categorical indicators yet. May I please ask if this is something that it is planned to be implemented in next version of semtools? And do plausible values require different assumptions when working with categorical indicators - instead of continuous ones?

[Terrence Jorgensen]

is this something that it is planned to be implemented in next version of semtools?

It can be in semTools as soon as it is available in lavaan.  As of now, the sampling (co)variances of factor scores are unavailable for factors whose indicators are latent response variables (underlying observed ordinal response variables).  I don't think this is high on Yves' priority list, so it might be available sooner from blavaan, where it is a high priority for the next version (although that could still take quite some time).

And do plausible values require different assumptions when working with categorical indicators - instead of continuous ones?

The latent response variables are assumed normal, but the fact that they are unobserved adds a layer of complexity to deriving / approximating their sampling variability.

[Ale Dea]

Are plausible values available for categorical indicators in the new version of lavaan?

[Terrence Jorgensen]

Are plausible values available for categorical indicators in the new version of lavaan?

Do you mean semTools::plausibleValues()?  No, lavaan only provides point estimates of factor scores when indicators are ordinal, and blavaan does not yet handle ordinal indicators.  When either package changes to provide distributions of factor scores with ordinal indicators, I can update semTools to capitalize on it.