lavaan WARNING: some estimated ov variances are negative

[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. 

Q1. 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?

Q2. Is it technically right to use just estimator = DWLS instead of 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

[Shu Fai Cheung]

I read the other post before reading this one. I believe the responses there, and my example, may have already answered your question.

I don't know the answer to Q1. It usually requires the examination of the data and model at hand. No general solution.

For Q2, the simple answer is no. Setting the estimator to DWLS is not enough because analyzing ordinal variables in SEM involves more than using DWLS as the estimator.

Hope this helps.

-- Shu Fai