the optimizer warns that a solution has NOT been found!

[epd]

Hi Guys,

I need your help for my CFA. I have the following model

OMSHC.bi <- 'GnF =~ OA1 + OA12 + OA13 + OA18 + OA20 + OD4 + OD6 + OD7 + OD10 +
OS3 + OS8 + OS9 + OS17 + OS19
Att =~ OA1 + OA12 + OA13 + OA18 + OA20
DHS =~ OD4 + OD6 + OD7 + OD10
ScD =~ OS3 + OS8 + OS9 + OS17 + OS19 '

When I run it with my N=286, I get the subject warning. However, when I simulated it with a dataset earlier with same N using below code, it was okay.

---- sim2 = simulateData(model, sample.nobs = 286) sim2 = round(sim2, 0) #Run CFA fit <- cfa(OMSHC.bi, data = sim2, estimator = "MLMVS", orthogonal=TRUE) summary(fit, fit.measures = TRUE, ci = TRUE)

----

Hope to get your support. Thank you in advance.

[Jasper Bogaert]

Hi,

No exact answer, just a quick thought. I have tried to run your model with simulated data and it indeed worked. The output seemed to be ok (as far as I know). This makes me think that the problem is situated in your data. If you use simulated data, you will get nice normally distributed data and it will all work fine (if the model is ok). Maybe this is not the case for your data, could you try and check for abnormalities or problems in your data?

Best wishes,

Jasper Bogaert 

PhD student and teaching assistant

Department of Data Analysis (PP01)

Faculty of Psychology and Educational Sciences

Ghent University

[epd]

Hi Jasper,

 

Thank you very much for your response.

 

I believe the warning is due to the data. I have tried to change the optimization method from the default NLMINB to BFGS, and it says that there are negative latent variable variances, especially for the factor 'Att.' When I tried to double-check with a PCA and scree plot, I got the data to load to three factors (and not four). I have tested other models (i.e., three-factor and unidimensional), and the data fit (albeit mixed/poorly) these two different models. I would like to believe that the Heywood case/ negative lv variances are due to structural misspecification, or in other words, the bifactor solution is not feasible.

 

I would like to hear your views if my reasoning could be considered a possibility.

 

Thank you.

 

Best,

Esau

 

 

 

[epd]

Also, I must mention that the Mardia's test I have done demonstrated that the data presents multivariate non-normality.

[Jasper Bogaert]

Hi Esau,

I am hesitant to give concrete advise as I am not very experienced and because I do not have a clear answer to your problem. Nonetheless, here are some thoughts I have:

1) Have you considered performing explorative factor analysis to make sure you find a correct factor structure?

2) Do you use the orthogonal = TRUE option (in the cfa function) intentional?

3) You mentioned that (some of) the items are not normally distributed, maybe you could try using the satorra bentler correction and a scaled test statistic?

I hope these thoughts may help you find a solution.

Best wishes,

Jasper Bogaert 

PhD student and teaching assistant

Department of Data Analysis (PP01)

Faculty of Psychology and Educational Sciences

Ghent University

[epd]

Thanks, Jasper!

My apologies if I have not given much of a context, I have preregistered my methods, hypothesis, cutoffs, models, and inference criteria. As I am replicating studies, the orthogonal =TRUE is intentional.

On the other hand, I have done EFA on the side, for sanity checks, and the scree plot and factor loadings supports the three factor structure.

I will try running your suggestion on the side. 

Best,

Esario

[Jasper Bogaert]

Hi Esario,

You are welcome and no problem at all. I hope one of the suggestions can help. You may also want to wait for responses from someone else (with more experience).

Best wishes,

Jasper Bogaert 

PhD student and teaching assistant

Department of Data Analysis (PP01)

Faculty of Psychology and Educational Sciences

Ghent University

Op zondag 13 maart 2022 om 04:23:52 UTC+1 schreef epd: