| Message/Author |
|
|
|
|
| I am currently examining a mediational model that requires the integration algorithm. Since the "model indirect" statement with bootstrapped standard errors is not available with this algorithm, would you have any recommendations for methods of calculating estimates of the indirect effect and possibly even standard errors (e.g. one of the product or difference methods)? |
|
|
|
|
| You can use the Delta method to calculate standard errors. One place you can find this is in Bollen's SEM book. |
|
|
|
|
| To clarify: 1) I can use the product coefficient approach with the deta method for standard errors? 2) Is there any way to test the assumption of a normally distributed sampling distribution for the indirect effect? |
|
|
|
|
You get the delta method standard error if you use assign labels to the 2 slopes that are multiplied (say a and b) and use those labels in Model constraint and the NEW option:
Model constraint:
New(ab);
ab = a*b;
This gives you a delta method SE for the indirect effect a*b.
Regarding your last question - I don't know. Check with Dave McKinnon at ASU who works with this. I wouldn't worry too much about non-normality unless the sample is rather small. |
|
|
|
|
Hi Charles,
One thing you could do is formulate an empirical distribution of the product based on your raw parms and se's - Dave MacKinnon has a new program that does this for the Asymmetric Confidence Interval test for mediation (in SPSS, SAS and R - and I imagine Mplus in the not-too-distant future  ). Here's the link:
http://www.public.asu.edu/%7Edavidpm/ripl/Prodclin/
If you use this, one rough (but pretty good) clue that I would look see is if the distribution approaches normality in the output are in the critical values of the 2.5th and 97.5th percentiles of the empirical distribution of your product - if they deviate from |1.96|, it's not normal. From what I remember, as the sizes of the ratio of the parms/se's (i.e., the Z-statistics) for the individual paths get larger, the product tends toward normality. If you have big effects and/or big sample (as Bengt alludes to), you'll be more likely to see it approach normality. You could also edit Dave's code to a) output the distribution and b) do formal tests for normality on your empirical distribution (i.e., the normality test that comes with SAS Proc Univariate). The ref for the program is here:
MacKinnon, D. P., Fritz, M. S., Williams, J., & Lockwood, C. M. (in press)
Distribution of the product confidence limits for the indirect effect program PRODCLIN. Behavior Research Methods.
His 2002 Psych Methods paper and/or his 2004 MBR paper talk about conditions when the distribution should approach normality. Hopefully, other mediation folks can weigh in if I have erred anywhere in this post...... |
|
|
| Mark Prince posted on Thursday, August 09, 2012 - 6:26 pm
|
|
|
I see here that the last post was in 2006. Is there currently an Mplus equivalent of PRODCLIN?
thank you. |
|
|
|
|
| What is PRODCLIN? |
|
|
| Mark Prince posted on Friday, August 10, 2012 - 1:22 pm
|
|
|
| PRODCLIN is Dave MacKinnon's program that computes the Asymmetric Confidence Interval test for mediation. |
|
|
|
|
| Mplus does have those. See the CINTERVAL option of the OUTPUT command. |
|
|
| Mark Prince posted on Friday, August 10, 2012 - 7:55 pm
|
|
|
| Thank you! That is what I was looking for. |
|
|
|
|
Hi,
We are running a moderation and mediation analysis with latent factors.
For the moderation, we are using general random for type and estimator MLR and algorithm integration. However, the software does not provide the "traditional" indices of fit (e.g., we don't see the CFI, TLI, RMSEA and SRMR). Should we rely simply on AIC and BIC (which are comparison indices I think?)
Similarly, we are running a mediation with the same latent factors, using type general and estimator ML and bootstrapping; We also do not see in the output the "traditional" indices of fit.
Note: several of the indicators on the exogenous factor are categorical. I wonder if it might have something to do with the traditional indices of fit not being provided.
Thank you
Simon |
|
|
|
|
| Yes, with categorical outcomes there is no overall fit index because you no longer fit the model to data summarized into a covariance matrix - only raw data will do. You can use BIC and also TECH10. |
|
|
|
|
Hi Dr. Muthen,
Thanks for the answer above. I was wondering, how do we interpret the information from TECH10 to assess model fit? Is this about looking at the number of cases in which standardized residuals are above 1.96 (absolute values) ?
And then running a potentially improved/different models, inspect the standardized residuals and compare the number of cases above 1.96 (absolute values) between the two models?
Thank you
Simon |
|
|
|
|
| You got it. |
|
|
| Back to top |
