|
|
ML and WLSMV estimation with categori... |
|
Message/Author |
|
|
A colleague used ML with a path model with 5 MVs, 2 of which were categorical. WLSMV is preferred for such models. Sample size was large (1198). Under ML: DEMENTIA ON AGED 0.048 0.010 4.643 NP 0.812 0.086 9.488 TDP 0.388 0.084 4.618 HS 1.038 0.280 3.714 Under MLR: DEMENTIA ON AGED 0.048 0.010 4.773 NP 0.812 0.084 9.704 TDP 0.388 0.085 4.575 HS 1.038 0.282 3.685 Under WLSMV: DEMENTIA ON AGED 0.025 0.007 3.299 NP 0.518 0.057 9.166 TDP 0.037 0.075 0.489 HS 0.431 0.088 4.903 So, under ML, MLR, and WLSMV, AGED, NP, and HS had effects with similar z ratios. But, the effect of TDP on DEMENTIA differed tremendously. Under ML and MLR, TDP had a robust direct effect with z around 4.60 and p < .0001. Under WLSMV estimation, TDP had a direct effect with z of 0.49 and p = .625, so not close to significance. Any ideas about why results for the direct effect of TDP varied so much across ML, MLR versus WLSMV? |
|
|
Hard to say without seeing the full output (send to Support). Perhaps TDP was binary and perhaps it was declared as categorical in which case ML(R) uses the observed TDP version as the predictor whereas WLSMV uses the latent continuous response variable behind the observed. |
|
|
Dear Dr Muthen, I have a similar question. I have a binary outcome, which is specified as categorical in the inputfile. However, using MLR or WLSMV produces very different results. I am using multiply imputed data so the difference could not be lying in the fact that MLR and WLSMV handle missing data differently. How can I decide which estimator I should be using? I am running a twolevel regession with only two predictors. Kind regards, Aurelie |
|
|
Can't really say without seeing the two full outputs and thereby see the specifics. But see the FAQ: Estimator choices with categorical outcomes |
|
Back to top |
|
|