Dr. Muthen, What is the difference between STDY Standardization and STDYX Standardization in the standardized model results? Which of the two should I interpret when I have direct and indirect effects unto one Dependent latent variable? Thank you, Linda
Unless you have a preponderance of zeroes, estimators like MLR that are robust to non-normality should handle skewness for continuous indicators. Floor or ceiling effects for categorical indicators are handled by categorical data methodology.
I am estimating a logistic regression (MLR) with some covariates that are continuous and others that are binary. I am using exclusively observed variables and no indirect effects. It is a simple logistic regression as you use to estimate in Stata or SPSS.
Should I use StdYX for continous and StdY for the binary covariates even when the two types of covariates are in one and the same model? I have a strange feeling by using for some covariates the coefficients from one part of the output and for the other covariates the coefficients from an other part of the output. Is this really advisable?
If I understand right, I should use the coefficients under the output block titled StdYX for the continuous covariates while for the binary covariates I should use the non-standardized coefficients at the top of the output?
Is it really advisable to use for some covariates one specific part of the output while for other covariates a different part of the output is used?
I am trying to make sense of differences in the p-values and meanings of the unstandardized vs. standardized (STDYX) covariances in the following output. All the variables in the model are continuous and both have missing data, so I am using FIML. Estimator = MLR. Typically, I like to report the STDYX results because they’re in an easily understood metric, but I am unnerved by such differences. I have several similar examples in regression models.
If the only difference is that STDYX is calculating the covariance having rescaled each variable to have M=0 and SD=1, I would expect the p-values to agree perfectly. Since that’s not the case, I’m unsure what Mplus is doing…how these inconsistent p-values occurred. Also, which results are “more correct.” And is the answer any different if I’m doing regression models with continuous DVs and predictors?
Thanks for your input!
Y WITH X 0.031 0.019 1.645 0.100
STDYX Standardization Y WITH X 0.322 0.156 2.062 0.039
Leaving Bayes est aside for the moment, which results would you put more stock in, the unstandardized (n.s.) or the standardized (sig)? Is it arbitrary because they're both "correct"? Sigh, I'm sure this is where the people who advocate getting rid of null hypothesis testing would pipe up.
So you're suggesting that Bayes is "better" than either the regular unstandardized or standardized solutions? I should trust the result that is more consistent with the Bayes result? If so, does that imply that I should be using Bayes all the time?
By the way, I've been using MLR because the IV and DV distributions are nonnormal.
Bayes does not assume that the sampling distributions of the parameters are normal as frequentist methods do. It is not clear how robust Bayes is to non-normality of the data. If your data are non-normal and the sampling distributions of the parameters are non-normal, you may be best off with ML and bootstrapping.
I am running a multiple mediation model with two dummy independent variables, 5 continuous mediators and 4 latent outcome variables. When reporting standardized output am I right in thinking I should report STDY for the path estimates between the independent variables and the latent factors and STDYX for the path estimates between each of my mediators and the latent factors? Also when reporting standardized indirect effects should I use STDY as the independents are binary variables?
Dear Dr. Muthen, I have a practical question regarding the numerical values I am getting in a two-level model. Although the observed scores of the variables in my model are between 1 and 5, and although I have already centered the scores at the higher level, the values that I am getting at the between level are excessively high under the STDYX standardization section (the same applies under the STDY standardization section). Specifically, although the paths from the between level to the outcomes seem reasonable (i.e., .71 and -.74) the intercepts are as high as 11.46 and 10.56. In the non-standardized solution the respective intercepts fall within logical boundaries (i.e., 3.56 and 3.04). I am getting exactly these values under the STD standardization section. In those sections however the paths from the between level to the outcomes are excessively high (.995 and -.965). So, my question is what am I missing (or doing wrong)? I thank you for your consideration
yvette xie posted on Monday, January 16, 2017 - 12:50 am
In my model, I have continuous latent variables as DV, and both continuous and bivariate variables as IV. Should I report the results in STDYX for continuous covariates and STDY for bivariate covariates?
yvette posted on Monday, January 16, 2017 - 5:09 pm
Thank you for your prompt reply.
I have a follow-up question. Some studies argued that we can no longer use standardised coefficient to compare the effect sizes of different IVs within a study. Then, why should we report the standardised solution?
Is it unusual or problematic to have all standardized values using LVMM to be positive? I understand that it is typical for at least some values to be negative and want to be sure that having all positive values for each indicator wasn't indicative of an error or problem with the data? This is using STDyx. Thank you.
I'm running several models without an interaction and one model with a latent variable interaction. Before entering the interaction I was reporting the stdyx results of my models (in my opinion easier to interpret). However, when I add the interaction effect to my model (XWITH) MPlus does not show me the stdyx results, only the model results. Is there a way to get the stdyx-results for a model with XWITH. If not, would you advise me to report the model results solely for the model with the interaction effect or do I need to be consistent and report the model results for all my models instead of the stdyx-results?
Thank you for your quick reply. I’ll try out running my model in the MPlus 7.4 version soon. I have one follow-up question: how does it come that the p-value for the model results and the stdyx results can be different (I use z-scores)? Sometimes there are only slight differences, however, two parameters in my model have a non-significant standardized coefficient (e.g., p = .069) whereas the non-standardized coefficient is significant (e.g., p = .025). Which one should I interpreted and report?
Standardized coefficient can have different significance than unstandardized
This is also discussed in our new book.
Hyunsik Kim posted on Thursday, February 01, 2018 - 7:44 pm
I'm running two types of survival analyses (semi-parametric model): (1) A model where continuous survival time is regressed on a latent factor (continuous indicators; factor variance is fixed at 1 and mean at 0), controlling categorical (gender) and continuous (age) CVs. (2) A model where continuous survival time is regressed on factor SCORES saved (continuous indicators; factor variance is fixed at 1 and mean at 0 when saving the factor scores), controlling categorical and continuous CVs.
In each model, which standardized output should I look at?
I would recommend both stdyx and stdy. For the continuous predictor use stdyx (so the predictor is standardized) while for the binary predictor use stdy so the predictor is not standardized. If that is too confusing you can standardize the continuous predictor with the define command and then use stdy.
ywang posted on Monday, February 19, 2018 - 1:06 pm
Dear Dr. Muthen,
If there is a dummy mediator variable and the Y and X are continuous, should I report the stdyx or std? Thank you very much!
If you have a binary mediator, the indirect and direct effects are best handled via counterfactually-defined effect formulas as discussed in our mediation book. But if you are not ready to go there, you can use WLSMV where instead the mediator is considered as the continuous latent response variable M* underlying the binary mediator M. In this case, you should use stdyx (which is the same as stdy for Y ON M*).
I don't know what it means to determine which growth curves to include in a parallel process model - substantive motivation would seem to be prime, not whether you use raw or standardized coefficients. This question may be more suitable for SEMNET.
Thank you for your response. Sorry for my lack of clarity. In my parallel process mediation model, I have three growth curves. I am trying to determine if change in one variable mediates the relationship between change in two other variables. I have one growth model with count data, one growth model with categorical data, and one growth model with continuous data.
Before running a parallel process model, it must be determined if all growth curve models have significant slopes. If the slope is significant, it is included in the model. If the slope is not significant, it is not included in the model. I am trying to determine if I should use the standardized or unstandardized p-values to decide if slopes are/ are not significant.
I can reach out to the SEMNET if this is not clearer. Thanks again for your help!
Usually, one uses the p-values or CIs of the unstandardized coefficients. Usually, they are the same as for the stand'd. If they are not, I would use bootstrapping so I could look at non-symmetric CIs because the 2 versions may differ in how close they approximate a symmetric distribution. This is discussed in our RMA book.