Message/Author 

Anonymous posted on Thursday, October 03, 2002  4:03 pm



How may one evaluate the significance level of correlations among the variables printed in the estimated sample statistics of LGM modeling with missing data? Does Mplus compute significance tests of these correlations? Thanks! 

bmuthen posted on Friday, October 04, 2002  6:57 pm



With missing data, you can obtain the estimated sample covariance matrix and the corresponding s.e.'s through type = basic missing and using the H1SE option (see user's guide). However, these s.e. pertain to covariances, not correlations. To get the correlation s.e.'s that I think you are asking for requires a bit more work using either of two alternatives. First, if you are familiar with the "Delta Method" you can get the s.e.'s via this method via the covariance matrix among the parameter estimates by requesting H1TECH3. Second, you can get the s.e.'s through a model (so not using type = basic) that has Sigma = Lambda * Psi * Lambda', where with p variables Lambda is p x p diagonal and Psi is full p x p with unit, fixed diagonal elements. This can be done by saying that you have one factor for each variable and that you have zero residual variances. Lambda will contain your standard deviations and Psi will contain your correlations. The Psi estimates and s.e.'s are the ones you want. 

S. Oesterle posted on Tuesday, August 30, 2005  3:12 pm



I am using ANALYSIS: TYPE=MISSING BASIC; to estimate correlations among several continuous and categorical variables with missing data. Is it appropriate to use the resulting S.E. FOR CORRELATION MATRIX to calculate the significance of the correlations by calculating the ratio of the estimated correlation divided by its SE and comparing it to the critical value? 


Yes. 


When using TYPE IS TWOLEVEL, is there any way to print standard errors for the correlation matrix in the SAMPSTAT output? Thank you. 


No. these are not given. 


Thank you for your response. As a followup question, how might I compute significance of the correlation values? 


You could specify all of the WITH statements in the MODEL command and ask for STANDARDIZED in the OUTPUT command. You will obtain standard errors for the standardized coefficients which are correlations. 

Tybeert posted on Wednesday, June 15, 2011  5:13 am



I used this method to obtain standardized coefficients and their correlations at the within level by specifying all of the WITH statements at the within level. If it is true that these are correlations, I would expect them to be exactly the same as the within correlations in the SAMPSTAT output. However, they are not. My question is why this difference exists, which of the two outputs are the ‘real’ correlations, and which of these correlations I should report. Thank you very much. 


This method does not work for multilevel models. You should use the correlations from SAMPSTAT or TYPE=BASIC. 

Tybeert posted on Thursday, June 16, 2011  12:16 am



Thanks. Is there any other method that I can use to calculate the standard errors for the SAMPSTAT correlations in the case of a multilevel model? 


Not that I am aware of. 


My output is not showing a column for statistical significance under Model Results (parameter estimations). I am getting Estimate, SE, and EST/SE but no significance. What should I specify in my input? 


You will obtain pvalues automatically if you are using a version of the program that has them. Otherwise you can look up the est/se in a ztable. 

Back to top 