lavaan WARNING: some estimated ov variances are negative for "ordered =T"

[Saeed Abbas Shah]

Hello everyone,

I am applying SEM using lavaan package for ordinal data.

(Situation 1)

When I am using ordered = TRUE, I get warning : lavaan WARNING: some estimated ov variances are negative and as per the output, one item (item8) of latent variable (LV3), I get R squired as NA, Variance Negative and for the same item (item8) std.loading on its latent variable (LV3) is 1.039 (more than 1)

But

(Situation 2)

When I apply SEM with estimator = DWLS on same data, I do not get that warning and none of the above issues mentioned appear in output. 

Question. Could you please help me to know possible reason/solution for situation 1 because as per the manual of Lavaan it is recommended to use ordered = TRUE for ordinal data?

data = item1+item2+item3+item4+item5+item6+item7+item8

LV1 =~ item1+item2+item3 ( each item has 3 responses = 3-2-1)

LV2 =~ item4+item5+item6 ( each item has 3 responses = 3-2-1)

LV3 =~ item8+item9 ( each item has binary responses = 1-0)

## mymodel

mymodel<-'LV1 =~ item1+item2+item3

                    LV2 =~ item4+item5+item6

                    LV3 =~ item8+item9

##Regression

                     LV3~LV1+LV2'

(Situation1)

sem_model = sem(mymodel, ordered = TRUE)

(Situation2)

sem_model = sem(mymodel, estimator = "DWLS")

Thanks 

Regards

[Chesnut, Ryan]

Others can probably speak to this more proficiently, but when you just request the DWLS estimator without declaring variables as ordered, lavaan doesn't know to treat your variables as ordered; by default, it will treat them as continuous, I believe. So, you will get different results when you simply request the DWLS estimator vs specifically declaring variables as ordered and using the DWLS estimator.

As for the Heywood case, I would guess that it has to do with LV3 having only 2 indicators. Best practice is to have at least 3 indicators per latent factor. In reality, however, that doesn't always work, but when you only have 2 indicators per factor, it can cause issues because that latent variable has to borrow information from other parts of the model to be identified. This might be further complicated by the fact that the items for this factor are binary.

There are a number of posts on the lavaan forum that talk about negative ov variances and most of them point to a 2012 paper in Sociological Methods & Research that discusses how you can look at the 95% confidence interval around the offending estimates to see if it includes plausible values. If it does, then the issue may be due to sampling error rather than model misspecification. Here is a link to post in the forum that might be helpful to you.

https://groups.google.com/g/lavaan/c/QywYJTTFTm0/m/C-TkX6WgCgAJ

---

Ryan P. Chesnut, PhD

Assistant Research Professor

Clearinghouse for Military Family Readiness

The Pennsylvania State University

402 Marion Place

University Park, PA 16802

Phone: 814-865-9637

Email: rpc...@psu.edu

Website: www.militaryfamilies.psu.edu

---

From: lav...@googlegroups.com <lav...@googlegroups.com> on behalf of Saeed Abbas Shah <s.a...@iba-suk.edu.pk>
Sent: Monday, December 13, 2021 8:35 AM
To: lavaan <lav...@googlegroups.com>
Subject: lavaan WARNING: some estimated ov variances are negative for "ordered =T"

 

--
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/1678cdd1-ff58-4003-87b6-c0177f5abb6cn%40googlegroups.com.