Indirect Effects - Calculation and Sa...
Message/Author
 Rob White posted on Monday, April 26, 2010 - 10:44 am
Hi,

1. Is there a reference available that reports the Mplus method for calculating indirect effects from a SEM with intermediate outcomes that include binary, categorical, count and continuous measures?

2. How can I save the estimated indirect effects, total effects and their standard errors in an ascii file? I don't see them in the RESULTS ARE output (or the ESTIMATES ARE) and really do not want to have to calculate these myself.

Thanks very much.
 Linda K. Muthen posted on Tuesday, April 27, 2010 - 9:51 am
1. There is no reference that I know of.
2. You do not need to calculate the indirect effects and their standard errors yourself, you can take them from the output.
 Rob White posted on Wednesday, April 28, 2010 - 9:11 am
1. Where can I read an explanation of how Mplus is calculating indirect effects from a SEM with intermediate outcomes that include binary, categorical, count and continuous measures?

2. I see that the indirect effects and their standard errors are in the .txt output. But I want to retrieve them from the ascii file that is generated by RESULTS ARE. They are not there. How can I save the indirect effects and their standard errors in an ascii file?

Thanks very much.
 Linda K. Muthen posted on Wednesday, April 28, 2010 - 10:26 am
1. I don't think this has been described anywhere. It comes from basic statistical principles.

2. The indirect effects are not saved. There is no way to save them.
 craig neumann posted on Thursday, February 10, 2011 - 1:04 pm
Hi Linda,

A reviewer wants to know how Mplus calculated indirect effects in an SEM we tested.

I assume that calculation of the indirect effect of X predicting Z via Y involves simply multiplying the direct effects.

So say X predicts Y (beta = .25) and Y predicts Z (.55), then the indirect effect of X on Z equals .25 x .55 = .137, correct?

Thanks...
 Linda K. Muthen posted on Thursday, February 10, 2011 - 1:34 pm
Yes, an indirect effect is the product of two regression coefficients.
 Mrinalini Rao posted on Sunday, March 13, 2011 - 8:22 pm
Hi Linda,
I am interested in indirect effects in a mediation model across 5 waves. I am using an effects-coding method of scaling.

However, in the output the estimates are too small to see the values.

a) Is there an option to see more than 3 decimals for indirect effects?

b) I multiplied my data by a 100 hoping to resolve this problem, but it does not converge when I do this. Is there a reason that a simple rescaling to see more significant digits is having convergence problems?

Thanks,
Mrinalini
 Linda K. Muthen posted on Monday, March 14, 2011 - 4:11 pm
There is no option to prints more than three decimals. The saved results contain more than three decimals.

Try multiplying by 5 or 10. The scale of the variables can impact convergence.
 Patchara Popaitoon posted on Tuesday, March 15, 2011 - 3:49 am
Hi Linda,

In relations to your response to Craig's question re: calculations for the indirect effects, which is the product of two regression coefficients. I wonder how the system generates the significant level for the indirect effect. Take A --> B --> C, what if the A-->B path is nonsignificant? and what if these two paths (A-->B and B-->C) are significant at different levels.

Thanks.
 Linda K. Muthen posted on Tuesday, March 15, 2011 - 10:37 am
Mplus uses the Delta method to calculate the standard error of the indirect effect. See the Bollen book to see how that is calculated.
 Ari J Elliot posted on Thursday, January 08, 2015 - 1:50 pm
As part of a mediation analysis I would like to report standardized indirect effects as unstandardized coefficients and SEs are extremely small. I am trying to figure out how best to use bootstrapping, if at all, to evaluate standardized indirect effects.

When using bootstrapping in conjunction with MODEL INDIRECT, is it appropriate to use the p-values for the standardized indirect effects using what I presume are bootstrapped SEs (the same way as with Delta method SEs and p-values) ?

Also, am I correct that examining whether Bias-corrected or standard bootstrapped CIs contain zero is not appropriate for standardized indirect effects?

Thank you!
 Bengt O. Muthen posted on Thursday, January 08, 2015 - 5:05 pm
Small unstandardized values may be due to x variables on a large scale. A simple solution is to use Define to scale down such variables by dividing by say 10.

You cannot use p-values for standardized to represent significance of unstandardized; they have different sampling distributions.
 Ari J Elliot posted on Thursday, January 08, 2015 - 5:38 pm
Thank you very much for the reply and suggestion.

So would it be ok when reporting standardized indirect effects to use p-values in the output corresponding to the standardized coefficients, and is this appropriate when SEs are generated using bootstrapping?
 Bengt O. Muthen posted on Thursday, January 08, 2015 - 7:04 pm
Yes. But we don't provide bootstrap SEs for standardized.
 Cheng posted on Wednesday, April 25, 2018 - 2:24 pm
Dear Linda,
May I know whether I should concern about a small estimate value for standardised specific indirect effect, with significant p-value? For Example, an estimate value for A->B->C->D is 0.001, and the p-value is 0.003. I expect such small estimate value, the p-value should be not be significant. How this p-value was estimated, based on estimate/SE or the estimate value? Thank you.
 Bengt O. Muthen posted on Wednesday, April 25, 2018 - 5:20 pm
The p-value is based on Estimate/SE. Instead, try the bootstrapping option to get better confidence intervals.
 Tor Neilands posted on Tuesday, November 13, 2018 - 10:50 pm
Hi,

I'm helping a postdoc to fit a complex SEM with multiple correlated mediators and a binary outcome using the BAYES estimator. Exposures (X's) are either binary or continuous with meaningful units, so we are using STDY standardization to interpret standard deviation changes in the underlying latent variable y* per unit changes in Xs. We are using MODEL INDIRECT to compute the traditional (i.e., non-counterfactual) indirect effects.

Under Bayesian estimation, Mplus helpfully supplies both 95% credible intervals for indirect effects (both raw and standardized) and asterisks those effects which are significant at p < .05.

For some of our results, the lower confidence limits print on the output as .000, yet the asterisk appears, presumably signifying that the CI boundary value for the effect is > 0, yet small enough to round to .000 on the printed output.

Can you recommend a way I can get the result at more significant digits? (if we sought to report and interpret the raw coefficients, I'd calculate them manually using MODEL CONSTRAINT with a multiplier of, say, 1000 [something I've done before when working with raw coefficients when this situation occurs] but I don't think that would be easily implemented for a standardized result where the variance expression for Y* is very complicated).

Thanks so much,

Tor Neilands
 Bengt O. Muthen posted on Wednesday, November 14, 2018 - 4:50 pm
Try using a longer run (more FBiterations or Biterations) to get more precise results.
 Tor Neilands posted on Thursday, November 29, 2018 - 9:00 am
Thanks, Bengt.

The original run was based on 2 chains and 500,000 FBiterations (default convergence was around 25,000 iterations) and had two significant indirect effects with zero values for the LCL. I then tried 2 million FBiterations and one of the two significant indirect effects with a lower CI boundary of zero became .001. I then tried 5 million iterations, this time with a single chain to speed run time, but the solution reverted back to having both effects showing an LCL of zero.

I'm not sure whether there's anything else to try at this point. We can always footnote those effects in our write-up with an explanation or report both raw and standardized results instead. For the future, a useful feature enhancement in a subsequent version of Mplus could be to add raw and standardized direct, indirect, total indirect, and total effects - and their CIs - to the RESULTS and BPARAMETERS SAVEDATA files or something similar that would allow the user to obtain those results at more than 3 decimals of precision if the need arises.

Thanks again and best regards,

Tor Neilands
 Bengt O. Muthen posted on Thursday, November 29, 2018 - 10:26 am
Noted on the wish list. But one can argue that 0.001 and 0.001 should not make a difference in the substantive conclusions - in neither case is the CI clearly bounded away from zero. Practically, in both case it is uncertain if the effect is larger than zero. With a wider CI than 95% it would reach into the negative.