AMOS can handle missing data without imputation as well.

All you have to do is check the "Estimate means and intercepts" box on the "Estimation" tab in the "Analysis properties" menu.

This setting (maximum likelihood estimation) generally performs about equally well as multiple imputation.

See: Article Missing Data

If you still want to use multiple imputation, I would check the AMOS manual.

Hello! It is advised that you utilize AMOS' own procedures for multiple imputation of missing cases (see Arbuckle, 2015). If you wish to proceed with your analyses without considering multiple imputation, you should proceed with Full Information Maximum Likelihood Estimation (FIML), which is the "royal" way to handling missing data (see The Wiley Handbook of Psychometric Testing- 2018). FIML is activated by clicking the "Estimate Means and Intercepts" option in the "Analysis Properties".

Regards

I will suggest mean value replacement.

Heiko Breitsohl , Ioannis G. Katsantonis , Noor Ul Hadi Thank you for your inputs. I used the single imputation with regression method. But now I am trying to perform a multiple imputation using the Stochastic regression method.

i am planning to use the imputation number to group them and get the result.

Any input on that will be appreciated.

Hello, again! As far I am able to tell, the IBM products, including AMOS , have limited capacities with pooling multiple imputed datasets. AMOS is not able to pool all the imputed datasets. The only procedure that could be applicable, I think, is to compute by hand the various estimates that you want by following several formulas. See Arbuckle (2015); he has examples. Otherwise you may transfer your analyses to R. However, Byrne (2016) suggests model-based procedures for handling missing values with AMOS, such as FIML. As you may be more familiar with the graphics user interface of AMOS and you may not want to do calculations by hand, then I highly recommend FIML.

I hope that I may have offered some help.

With regards.

Ioannis G. Katsantonis I performed the analysis using FIML, but the problem is- even if my variables are nominal, the imputed data converts the data into continuous, while maintaining the measure status.

For ex.-

In original- Gender- 0- Male ; 1- Female

In imputed- Gender- 0 (Male); 0.44; 0.45; 0.46; 0.47; 1(Female)

Any input, how we can fix that?

Hanish Kodali , FIML is used with your raw data including the missing values. There is no reason for multiple imputation, if you use FIML.

Ioannis G. Katsantonis Heiko Breitsohl

Thank you for your inputs in past.

I am writing my paper, so I redid my whole analysis, just to be sure but I have some issues and I was wondering you have a better insight on it.

General Info-

1) exposure, outcome and one of the mediators are Latent variables.

2) 2 mediators are continuous observed variables

3) 4 covariate as confounders- age, gender, income, and race

Steps

1) Using complete cases (n = 757), performed a regular SEM- mediation analysis in AMOS (without confounders)

2) After eliminating all the variables with Factor loading >0.30 and adjusting for co-variance, got a good model fit.

3) Used the same skeleton, used the incomplete data (n= 1330) to conduct FIML imputation

4) I got a complete dataset (n= 1330) with complete columns for all the variables and newly constructed latent variables.

Issues-

1) After imputation, convert the nominal data to continuous data

For ex.-

In original- Gender- 0- Male ; 1- Female

In imputed- Gender- 0 (Male); 0.44; 0.45; 0.46; 0.47; 1(Female)

2) In the imputed dataset, there are 3 new columns for all the 3 latent variables, which was calculated on its own during the imputation.

For my analysis, which one shall I use

a) the same old skeleton and adding all the unobserved variables and perform the SEM

b) just stick to the newly SEM-FIML created the latent variables from imputed data.

I will really appreciate a response on this issue.

Hello!

In your place, I would have preferred option no. A. Also, I would call fiml as estimation method because it is not imputation. It is further advised to read a couple of papers on fiml to get a better idea of missing value handling:

Enders,& Bandalos (2001);

Enders (2001). You can mail me to send them to you. Best wishes.

Let me be more clear. I suggested to you to specify your model as you described with all latent and observed variables on the basis of your initial dataset with missing and complete values and then to simply activate FIML.