missing = 'fiml' and estimator = 'MLM' or 'MLMVS'

[Lim Yonghao]

Dear list,

It seems like 'MLM' or 'MLMVS' cannot be run when missing is set to 'fiml'. I assume that for these estimators, the only option, other than multiple imputation, is listwise deletion. I am just wondering is there a reason why handling missing data with FIML is not implemented for MLM or MLMVS?

Thank you for your attention.

Yonghao

[yrosseel]

On 04/13/2013 09:01 AM, Lim Yonghao wrote:
> Dear list,
>
> It seems like 'MLM' or 'MLMVS' cannot be run when missing is set to
> 'fiml'. I assume that for these estimators, the only option, other than
> multiple imputation, is listwise deletion. I am just wondering is there
> a reason why handling missing data with FIML is not implemented for MLM
> or MLMVS?

MLM(VS) is for complete data only. For missing data, use estimator="MLR".

Yves.

[Lim Yonghao]

Dear Yves,

Thank you for the prompt reply. After reading Yuan & Bentler (2000) briefly, i understand that the Yuan-Bentler correction is similar to the Satorra-Bentler correction (scaled chi-square) for complete data (or the MLM option under estimator). Is there a way to get lavaan to provide the Yuan-Bentler correction equivalent for the adjusted chi-square where the correction factor and df are both corrected (similar to mean and variance adjusted chi-square)?

Once again, thank you very much for the attention.

Yonghao

[yrosseel]

On 04/14/2013 09:00 AM, Lim Yonghao wrote:
> MLM option under estimator). Is there a way to get lavaan to provide the
> Yuan-Bentler correction equivalent for the adjusted chi-square where the

> correction factor and df are both corrected for missing data (similar to

> mean and variance adjusted chi-square)?

So you wish to have a 'Satterthwaite' version of the Yuan-Bentler
correction (MLR)? I'm not sure if this is possible (let alone
necessary), and I have no time to investigate this right now. If you
would have a reference showing how this can be done, I'll implement it
right away.

Yves.

[Lim Yonghao]

Dear Yves,

I apologised for the late reply. I am not aware of any source that document the implementation of the Satterthwaite version of the Yuan-Bentler correction. However, i read about the performance of that correction in a simulation study conducted by Savalei (2010). In that study, Savalei (2010) found that the Satterthwaite version of the Yuan-Bentler correction (called the adjusted chi square in that paper) has the best performance in model evaluation, both in terms of type 1 error and power, when the sample sizes are small and when there are incomplete non-normal data. 

I am guessing that the same sample fourth-order moment weight matrix and the residual weight matrix used in the Yuan-Bentler correction are applied to the formula in the Satterthwaite correction (i.e. MLMV). I cannot be sure of this as i do not understand the technical aspects of the correction fully.

Thank you very much for the attention.

Yonghao

Savalei, V. (2010). Small sample statistics for incomplete nonnormal data. Extensions of complete data formulae and a Monte Carlo comparison. Structural Equation Modeling, 17 (2), 241-264.

[yrosseel]

On 04/20/2013 03:07 PM, Lim Yonghao wrote:
> Dear Yves,
>
> I apologised for the late reply. I am not aware of any source that
> document the implementation of the Satterthwaite version of the
> Yuan-Bentler correction. However, i read about the performance of that
> correction in a simulation study conducted by Savalei (2010). In that
> study, Savalei (2010) found that the Satterthwaite version of the
> Yuan-Bentler correction (called the adjusted chi square in that paper)
> has the best performance in model evaluation,

I'll have a look at the Savalei paper. If I can figure out the formulas,
I will implement a Satterthwaite variant of MLR.

Yves.