Thank you Prof. Geiser -- I really appreciate what you do with this channel :) Just a comment -- your explanation of non-positive-definite problems as model misspecification is one that I also saw on the Mplus Discussion Forum before, but I never really understand what that means, even after watching your video. I guess a negative residual means that the model explains >100% of the variance of the DV -- this is just something I cannot wrap my head around... What kind of misspecification would lead to this problem? If there is a more in-depth and intuitive (non-technical) exhibition of the internal workings of this problem, I guess that would make it easier to understand.
@QuantFish
2 жыл бұрын
Hi Shi Yu, Thanks for watching! An example of a misspecification that could lead to this problem is when you have too many factors (latent variables) for a given variable (an overly complex and/or overparameterized model). In that scenario, the factors may "overexplain" a variable, leading to > 100% explained variance (which, as you point out correctly, is impossible). Conceptually, you could think of this as the "response" to overparameterizing a model--the negative error variance is a symptom of the unnecessary factors/parameters. It can also happen as a result of other kinds of misspecifications. Christian Geiser
@shiyu5769
2 жыл бұрын
@@QuantFish Thank you Prof. Geiser for the explanation -- it does help understanding. However, it is interesting to compare it with simpler statistical models like regression. In my impression, in a model like regression, although adding too many explanatory variables and parameters may cause undesirable outcomes such as overfitting, I have never seen R2 over 100%. That is, for the more basic models like regression, no matter how "wrong" the model is, nothing like the non-positive-definite problem in SEM occurs. I wonder why is SEM different in this respect?
@QuantFish
2 жыл бұрын
@@shiyu5769 Conventional OLS (linear) regression is saturated (just identified, zero degrees of freedom). There is a 1-to-1 correspondence between available information (variances, covariances, means) and parameters to be estimated. In other words, all the available information in the data is exactly reproduced by the estimated regression parameters. In contrast, many CFA/SEM models are overidentified (non-saturated, > 0 df) such that fewer parameters are estimated than there is available information in the data. Therefore, CFA/SEM models can show misfit (whereas OLS regression models cannot--they always reproduce the observed covariance and mean structure perfectly by definition). Sometimes, the misfit of overidentified models in CFA-SEM is "minimized" by estimating parameter values that are out of bounds (improper estimates) such as negative residual variances which result in R^2 > 1.0. What this typically means is that the model is in some way misspecified--improper parameter estimates are usually a symptom of model misspecification.
@shiyu5769
2 жыл бұрын
@@QuantFish Thanks Prof. Geiser for your detailed explanation!
@sabaqaamozinfo.learningent1480
2 жыл бұрын
.Plz suggest solution for this error. 👇 THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS -0.184D-16. PROBLEM INVOLVING PARAMETER 210. THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE SAMPLE SIZE.
@QuantFish
2 жыл бұрын
It could mean your model is unidentified or too large/complex (has too many parameters) for the given sample size. Christian Geiser
@sabaqaamozinfo.learningent1480
2 жыл бұрын
@@QuantFish thanx for quick response..yeah it has too many parameters becoz i m running a moderated serial mediation, where all variables are continuous, but the results and model fit indices are very good... HOw to Remove this error Sir ?
@QuantFish
2 жыл бұрын
@@sabaqaamozinfo.learningent1480 I wouldn't trust the results for a model that produces this error message in Mplus. The results may be incorrect. Reducing the complexity of the model may help. Best, Christian Geiser
@sabaqaamozinfo.learningent1480
2 жыл бұрын
@@QuantFish thank you so much Sir. i ll have to say ur platform is a great source of learning mplus for me
Пікірлер: 10