Anonymous posted on Monday, March 12, 2007 - 9:23 pm
I have a construct that has 3 items that are dichotomous and other 4 items that are measured on a 3-point scale. My measurement model for that factor has all the items for that particular construct as is. To bring this items on the same scale, I standardized these variables. When I run the Mplus model with Standardize variables, I am not able to get the model to converge. However, when I run the unstandardized variables, I can get it to converge. What should I be doing? I am not very comfortable putting items with different scales on the same construct with out standardization.
You should not standardize these variables. Just use them as is. Put them on the CATEGORICAL list. The numbers for dichotomous and 3-category variables have no numerical meaning. They just represent categories so standardizing them has no meaning.
I apologize if this questions has previously been addressed, but I had a question regarding a CFA analysis I am performing. I have a continuous latent variable measured by 8 indicator that are both continuous and categorical (4 ordered categories). When I look at the factor loading and the loading standardized using the the variance of the continuous latent variable (Std), there seems to be a distinct quantitative difference between the categorical and continuous variables. Would you recommend I use the StdYX standardization instead? If so, can you please explain why. Thank you!
With categorical indicators you are considering probit or logit slopes which are on different metrics than slopes for continuous indicators. If you use StdYX you will get results that are more similar because then you consider an underlying continuous y* variable. In that scale, categorical items may nevertheless have smaller coefficients due to being less reliable indicators.
In Mplus user's guide, it is said that STDY cannot be obtained for categorical outcomes and weighted least squares estimation. I run CFA on (only) categorical outcomes with WLSMV estimation (delta) on Mplus 5, and the output shows STDY estimates. Is that normal?
I am running a factor analysis on categorical variables and would like to have standardized parameters. I have two questions. 1. In the Mplus documentation, I have not been able to find the definition of the standardized parameter arrays for tau, the threshold parameters. Are these defined similarly to the standardized parameter arrays for nu (intercepts)?
2. STDY is not available for categorical outcomes and WLS estimation. However, STDYX is available. In my model there are no background (X) variables. Is it correct that STDYX is equal to STDY when there are no X variables? And how can STDYX be available when STDY is not?
Is it possible to standardize variables in Mplus prior to starting your modeling? I know that there is the option to center variables prior to modeling. Can I also standardize variables prior to modeling (put them in z-score format)?
Yes, this was added to the DEFINE command in Version 6.1.
The STANDARDIZE option standardizes continuous variables by subtracting the mean from each value and dividing by the standard deviation. Following is a example of how the STANDARDIZE option is used:
STANDARDIZE y1 y5-y10 y14;
where the variables y1, y5 to y10, and y14 will be standardized. The order of variables for the list function is taken from the order in the USEVARIABLES statement. If there is no USEVARIABLES statement, the order is taken from the NAMES statement. When a variable on the STANDARDIZE list is used in other transformations, the original values of the variable are used. Standardization takes places after all other transformations have been completed.
Joe King posted on Tuesday, December 20, 2011 - 11:44 am
how do you interpret standardized parameter estimates (STDY) when the variable associated with that parameter estimate is standardized.
I ran a 2-factor CFA. Factor 1 (F1) had 4 items, and factor 2 (F2) had 5 items. Both factors were latent continuous. The model fit indices were all good. All of the factor loadings were greater than .3. I see the coeffecient for F1 with F2 was 1.057 using both the STDYX and STD. I am unable to find a discussion that addresses what to do with this coeffecient greater than 1. Can you provide some counsel? Thanks
JPower posted on Tuesday, April 09, 2013 - 9:25 am
Hello, I have been examining a CFA with 5 latent factors, each with multiple ordinal indicators. The model is estimated using WLSMV. I have now extended the CFA to a MIMIC model and have included a dichotomous variable. I have regressed each of the latent factors ON this new dichotomous variable.
I would like to be able to report STDY for these new relationships (because the variable is dichotomous I only want to standardize with respect to the latent factors). However these estimates do not appear in my output.
Prior to including the dichotomous variable, my CFA outputs did include STD, STDY and STDYX estimates. In this new MIMIC model, STD and STDYX estimates are output, but there are no associated SE, Est/SE or p-values. Additionally, STDY estimates no longer appear at all. Do you know why this might be? I have included STAND in the OUTPUT section of my code.
With WLSMV and covariates, limited standardized output is given. You will need to create StdY by dividing StdYX by the standard deviation of x.
JPower posted on Friday, April 12, 2013 - 10:19 am
Hello, I have a couple of additional questions about standardization in CFAs and MIMIC models estimated using WLSMV.
1) In a CFA model with all ordinal indicators, I add a "with" statement between 2 indicators. For this estimate, the raw and STD estimates are the same. The STDYX and STDY estimates are the same. Am I correct in understanding that the raw and STD estimates for this relationship represent the correlation between the y* indicators (the underlying continuous indicators that are standardized by default in WLSMV)? Does it follow then that the STDYX and STDY represent the correlation between the observed ordinal indicators? Am I correct in assuming that it is the corr between the y*'s that I want to report?
2) In a MIMIC model, I have regressed a latent variable and one of its indicators on a continuous covariate. For the path from the covariate to the latent variable, I am reporting the STDYX estimate (both latent variable and covariate standardized). For the effect of my covariate on the indicator, I am unclear what the estimates represent in terms of y* and what is being standardized. Again, the raw and STD estimates are the same, but the STDYX is different. Is it correct that the raw and STD estimates represent the effect of the covariate on y*? And that the STDYX refers to the effect of the covariate on the observed indicators? Which should be reported?
1) WITH between 2 factor indicators (or any DVs) refers to their residual covariance. So, standardizing wrt Y gives the correlation between the residuals for the 2 indicators. That's worth reporting.
2) Yes and yes.
It is easy to understand what goes on if you recall that STD is standardization wrt the latent variables only. The latent variables are not involved in the residual covariance, nor in the direct effect of a covariate on a factor indicator.
RuoShui posted on Friday, September 27, 2013 - 9:53 pm
I extended my SEM model into a MIMIC by including a continuous observed score (on the same scale as the other factors)and a z score of socioeconomic status. However, there is no Est/SE or P-value any more for the stdyx output. Would you please let me know why this happens? Thank you very much.
When reporting the standardized effects of a set of predictors/covariates, some continuous and some dichotomuous, on a latent variable with continuous indicators, should I report the SDTYX for the continuous and the SDTY for the dichotomuous, or should I report the SDTY for all the predictors (the whole covariate model) in order to be consistent, since some of them are dichotomuous?
In your post dated September 28th 2013 you say that the next update will include standard errors for standardized estimates in conditional models. I wonder if you have an approximate time scale for when the next update will be released?
I was wondering why, or why not, to standardize continuous variables prior to entry into a CFA or SEM model. I've seen people create z-scores because they say it improves model fit. Why would that be? And is that acceptable practice? Thanks.
Jessica T. posted on Thursday, April 20, 2017 - 12:08 pm
I am running LVMM with 7 subscales from 2 different measures - one of which uses a Likert scale from 0-4 and the other has a Likert scale from 1-6. Each subscale has a different number of questions in it, so the ranges for each subscale vary even within the particular measure. I would like to know whether the standardized means take into account that each of these subscales was different to begin with? In other words, are the scores standardized within each subscale and also compared to other subscales (e.g. will subscale 1's standardized mean that has a range of 0-8 be comparable to subscale 2's mean that has a range of 4-24). Sorry if the wording of the question is confusing. Thanks.
This general analysis question is better suited for SEMNET.
Min posted on Thursday, January 11, 2018 - 8:20 am
Dear Dr. Muthen
I ran a MIMIC model using the following commands in estimation. I also included the standardized command in the output. However, I still only have unstandardized coefficients. Would you let me know how I can have standardized coefficients in Mplus output and/or compute them using the unstandardized output?
I attain different estimates and p-values in regression analysis depending upon whether the data is standardised or not. More specifically, I have tried using the 'Define: standardize' option towards the beginning of the syntax or the STDYX option at the end of the syntax, and the estimates and p-values differ. Would you know why this is the case?