SEM categorical item issue

[Taman Powell]

Hello, I'm a bit of a novice at lavaan and SEM and was hoping for some assistance with this problem. Basically I have a SEM that was working fine, now I’ve added 4 binary items to the model (each one coded as 1 or 0) and the model dies with the following error:

Error in lavsamplestats@cov[[g]][cbind(row.idx, col.idx)] :
  subscript out of bounds
In addition: Warning messages:
1: In lav_data_full(data = data, group = group, group.label = group.label,  :
  lavaan WARNING: exogenous variable(s) declared as ordered in data: locSZ locGD locMA locOT
2: In lav_partable_check(lavpartable, categorical = categorical, warn = TRUE) :
  lavaan WARNING: parameter table does not contain thresholds

The actual SEM model is:

modelDef <- '
Control =~ contr4 + contr6 + contr3 + contr9      # latent variable definitions
TranSuc =~ prac10 + prac7 + prac11
HCAvail =~ resoua12 + resoua11
HCDevl =~ perfor8 + perfor9
HCStr =~ confid4 + confid1 + confid3
Perform =~ perfor20 + perfor18 + perfor19


TranSuc ~ Control                                 # Model
HCStr ~ HCAvail + HCDevl
Perform ~ TranSuc + HCStr + Age + Sector + staffchil + staffhkl + locSZ + locGD + locMA + locOT
‘

With the locSZ,  locGD,  locMA, locOT fields being the categorical items. I’ve identified these in the SEM as follows as I think I need to do, but not having any joy.

modelFit <- sem(modelDef, data = dataset, ordered=c("locSZ","locGD", "locMA","locOT")
                , sample.cov = true, estimator = "WLS")

Prior to adding these 4 items it was all working fine!

Any tips would be greatly appreciated.

[yrosseel]

On 12/01/2015 10:59 PM, Taman Powell wrote:

> Perform ~ TranSuc + HCStr + Age + Sector + staffchil + staffhkl + locSZ
> + locGD + locMA + locOT
>
> With the locSZ, locGD, locMA, locOT fields being the categorical
> items. I’ve identified these in the SEM as follows as I think I need to
> do, but not having any joy.
>
> |
> modelFit <-sem(modelDef,data

> =dataset,ordered=c("locSZ","locGD","locMA","locOT")

> ,sample.cov =true,estimator ="WLS")

Only endogenous (dependent) variables need to be declared as 'ordered'.
Not exogenous predictors. If they are binary (coded 0/1), you can just
leave them as they are.

See: http://lavaan.ugent.be/tutorial/cat.html

Yves.

[Taman Powell]

Hi Yves,

Thanks for your response. I’ve tried just leaving them binary coded. The model works when I just use 1 of the 4 items, but when I add a second I get an error:

Error in lav_samplestats_from_data(lavdata = lavdata, missing = lavoptions$missing,  : 

  lavaan ERROR: Gamma (NACOV) matrix is not positive-definite

I’ve tried this will all different combinations but still seem to get the same error.

Sorry if this is a really basic question. But it’s killing me!

Best, Taman 

--
You received this message because you are subscribed to a topic in the Google Groups "lavaan" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lavaan/QxnaM5eyNv8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to lavaan+un...@googlegroups.com.
To post to this group, send email to lav...@googlegroups.com.
Visit this group at http://groups.google.com/group/lavaan.
For more options, visit https://groups.google.com/d/optout.

[yrosseel]

On 12/02/2015 05:54 PM, Taman Powell wrote:
> Thanks for your response. I’ve tried just leaving them binary coded. The
> model works when I just use 1 of the 4 items, but when I add a second I
> get an error:
>
> Error in lav_samplestats_from_data(lavdata = lavdata, missing =
> lavoptions$missing, :
> lavaan ERROR: Gamma (NACOV) matrix is not positive-definite

Are the binary variables perfectly correlated?

Could you post a snippet of the data (just enough to show this error)?
Only then can we investigate.

Yves.

[Taman Powell]

Hi Yves,

I’m actually just trying to control for geographic area in the data. So I’ve setup four items where only one will be true relating to the area that the company is in. I’m attaching a csv with the first 40 lines of data. The fields are E, F, G, H. Only one in each row will be positive relating to the geographic area that the company is in. Maybe there is a better way to do this, but this seemed to make sense.

Thanks again for your help with this. In final revision (well should be) of quite a big paper for me and this is the key outstanding question.

Best, Taman

[yrosseel]

On 12/02/2015 07:09 PM, Taman Powell wrote:
> I’m actually just trying to control for geographic area in the data. So
> I’ve setup four items where only one will be true relating to the area
> that the company is in.

It would seem three columns are enough (and use 0 0 0 for the 4th area).

Yves.

[Taman Powell]

Three would be enough, but even when I just use 2 of the items it returns a:

Error in lav_samplestats_from_data(lavdata = lavdata, missing = lavoptions$missing,  : 

  lavaan ERROR: Gamma (NACOV) matrix is not positive-definite

I’m just trying to see whether geographic area has an impact on the model - like the attached figure (geographic areas being the four items to the right).

Best, Taman

[yrosseel]

On 12/02/2015 07:51 PM, Taman Powell wrote:
> Three would be enough, but even when I just use 2 of the items it returns a:
>
> Error in lav_samplestats_from_data(lavdata = lavdata, missing =
> lavoptions$missing, :
> lavaan ERROR: Gamma (NACOV) matrix is not positive-definite
>
> I’m just trying to see whether geographic area has an impact on the
> model - like the attached figure (geographic areas being the four items
> to the right).

Is it possible that the information contained in the two location
variables somehow overlaps with information contained in other parts of
the model?

If you wish, you can email me the full data + script directly, and I can
have a look at it.

Yves.

[yrosseel]

On 12/02/2015 07:56 PM, Taman Powell wrote:
> Hi Yves,
>
> I don’t think there is any overlap with other parts of the model.
> Basically model was looking at completely different things and then had
> a query regarding whether geographic area had an impact, so setup the 4
> items to look at this. Anyway, I’m attaching the complete dataset and
> code. I really appreciate your help with this.

Ok. The estimator is the problem: "WLS". Your sample is not large enough
for this estimator. You really should use ML, or even better MLR (robust
ML). It works, I tried.

WLS (in the continuous case!) will treat all variables (including the
exogenous 0/1 variables) as stochastic and Gaussian. That will almost
certainly give problems if you include 0/1 variables...

Yves.

[Taman Powell]

Thank you! OMG, it works. I didn’t think about trying to change this. Changed almost everything else!! Thanks again!!!! If you’re in London sometime let me buy you several beers :-)

[Mandy]

Dear Yves, 

My question is out of topic but I was wondering whether you could tell me the equations of how to manually calculate 

(i) the Satorra-Bentler scaling correction factor by which to divide the original ML chi-square test statistic

and 

(i) robust standard errors (the ones that show in lavaan package under MLM)

Thanking you in advance