How can I get “cfi.robust” instead of “cfi.scaled” in ordered CFA ?

799 views
Skip to first unread message

岡林秀樹

unread,
Mar 1, 2021, 9:49:39 PM3/1/21
to lavaan

As Terrence mentioned on another log, “The correct calculation of robust fit measures proposed by Brosseau-Liard & Savalei (2014) has only been proposed for mean-adjusted chi-squared statistics (estimators MLM, MLR, WLSM, ULSM), not mean- and variance-adjusted (or "scaled & shifted") statistics (estimators MLMV, WLSMV).  WLSMV is the default for ordinal data, so that is why cfi.robust, rmsea.robust, etc., will be NA.  The *.scaled fit measures are simply using the original formulas for non-robust chi-squared, but plugging in the robust chi-squared (which, as the articles explain, will not be consistent estimators of the true population values).”

 So, I think that “cfi.robust” is the better fit index than “cfi.scaled” because cfi.robust will be consistent with the true population value. Is my understanding right?

 Then, I wrote the following R code and conduct it. But I can get the error message.

 ******************************************************************************************

configural.fit <- cfa(model, data=combined.dat, estimator="MLM", group="ng",

                      ordered=c("O1", "O2","N.O3", "O4", "O5", "N.O6",

                                "N.SP1", "N.SP2", "N.SP3", "N.SP4", "N.SP5","N.SP6",

                                "N.CP1", "CP2", "N.CP4",

                                "CP3", "CP5", "CP6",

                                "N.SS1", "N.SS2", "N.SS3", "N.SS4", "N.SS5","N.SS6",

                                "CS1", "CS2", "CS3"))

 Error in lav_options_set(opt) :

  lavaan ERROR: estimator ML for ordered data is not supported yet. Use WLSMV instead.

 ************************************** 

I set the estimator as “MLM”, but lavaan did not recognize it.

Would you tell me what is wrong on my R code?

Then, would you tell me how to get “cfi.robust”?

 

Sincerely,

Hideki 

Chesnut, Ryan

unread,
Mar 1, 2021, 10:00:25 PM3/1/21
to lav...@googlegroups.com
You’ve declared your variables as ordered but you are using an estimator that requires (or assumes) they are continuous. If you want to use MLM, then remove the ordered argument. Of course, that may not be a good idea if the variables really do need to be treated as ordered.

Sent from my iPhone

On Mar 1, 2021, at 9:49 PM, 岡林秀樹 <hideki.okab...@gmail.com> wrote:


--
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 on the web visit https://groups.google.com/d/msgid/lavaan/881a7018-4de9-4bb1-b4ea-ed808b343be8n%40googlegroups.com.

岡林秀樹

unread,
Mar 2, 2021, 2:06:41 AM3/2/21
to lavaan

Thank you very much for your suggestion. 

I understand that we should use WLSM, ULSM, and WLSMV estimator for ordered variables.

I confirmed which estimator gets which goodness of fit (“*.robust and *.scaled” or “scaled” only).

プレゼンテーション1.jpg

     WLSMV is a default estimator for ordered variables. But it can get only scaled fit measures.

     If the scaled fit measures will not be consistent estimators of the true population values as Terrence suggested 

    and robust fit measures are better than scaled fit measures, we should use WLSM or ULSM to get robust fit measures instead of using WLSMV. 

     How do you think about this?

     My question is which fit measures is the best for ordered CFA (scaled or robust). 

     Then, which estimator is the best for ordered CFA?

 

Sincerely,

Hideki


2021年3月2日火曜日 12:00:25 UTC+9 rpc...@psu.edu:

Patrick (Malone Quantitative)

unread,
Mar 2, 2021, 11:39:11 AM3/2/21
to lav...@googlegroups.com
Hideki,

The "robustness" in MLM/MLR is robust with respect to deviations from normality in the outcomes. Using WLSMV estimator means such deviations are not applicable to ordered outcomes--with the probit link they are modeled as reflections of an underlying continuous normal variable. As the error told you, "estimator ML for ordered data is not supported yet." So WLSMV is the best bet even if it's not perfect.

If this is critical to you, I believe Mplus can estimate ML with ordinal data by numeric integration.

Pat



--
Patrick S. Malone, Ph.D., Malone Quantitative
NEW Service Models: http://malonequantitative.com

He/Him/His

岡林秀樹

unread,
Mar 2, 2021, 9:16:14 PM3/2/21
to lavaan

Dear Pat,

 

Thank you very much for your comment. I understand that ULS and WLS estimators should be used for ordinal data.

Although WLSMV is set as a default estimator, there are other variants such as WLSM, ULSM, WLSMVS, ULSMV, ULSMVS as described in the following table.  I would like to know which is the best among them. If WLSMV is the best, I would like to know the reason why it is considered the best among the several variant estimators. 

プレゼンテーション1.jpg

Then, there are two adjusted fit measures for ordinal data such as *.scaled and *.robust. Which one should be referred as fit measures for ordinal CFA?  Only WLSM and ULSM give us both *.scaled and *.robust measures as described in the table.  If *.robust measures is better than *.scaled ones, should we use WLSM or ULSM instead of WLSMV?

 Would you please give me any suggestions about them?

Sincerely,

Hideki

2021年3月3日水曜日 1:39:11 UTC+9 mal...@malonequantitative.com:
Reply all
Reply to author
Forward
0 new messages