| Message/Author |
|
|
| John C posted on Wednesday, June 20, 2018 - 8:48 am
|
|
|
I have two questions on Sensitivity Analysis, based on the example in the Mplus textbook chap 3.5.
1 - It is always necessary to include a moderator (Z) in addition to the X,M, and Y variables?
2 - Can one get sensitivity plots when the mediator is binary? I tried this but got no sensitivity plots in that case, although perhaps there is a specific estimator that has to be used (I specified ML with bootstrap)? |
|
|
|
|
1. No, but it seems reasonable.
2. This is not available yet in Mplus. |
|
|
|
|
Hello there,
Could someone please tell me how to view the sensitivity plots?
I am using Mplus on Windows, but I do not see the option anywhere.
Is there something else I have to do?
Thanks,
Andre. |
|
|
|
|
Apologies for the double post; I figured out that I was not using the loop command appropriately.
However, how do I get the sensitivity analysis plot without a moderator?
Thanks again,
Andre. |
|
|
|
|
You use the Plot command option
Type = Plot3 Sensitivity;
and then look for the View plot menu option "Sensitivity plots". |
|
|
|
|
Thanks for the reply, Bengt.
I am able to get the plots now.
My model has 2 mediators and 3 outcomes. If I wanted to get sensitivity analyses for all, would I just have to ask for them individually in the "Model Indirect" command? (From, my reading, multiple sensitivity plots cannot currently be generated).
Would there be anything wrong with doing this?
Thanks again. |
|
|
|
|
| The current implementation of the Sensitivity analysis is for 1 mediator and 1 outcome. Doing 1 outcome at a time should be fine. But if your model truly has more than 1 mediator, then this approach cannot be used. |
|
|
|
|
Thanks again for the response.
Why is that? (Could you point me in the direction of the literature?)
I have read the Imai 2011 and 2013 papers.
Also, is there an alternative approach that you/anyone might be able to suggest that I could use to deal with multiple mediators?
And in the worst case scenario, would it be useful to drop 1 mediator at a time and run the sensitivity analysis like that? (Understanding that it would fundamentally change some things about the model, but maybe it would be useful from a discussion point of view about the robustness of mediation through one variable on the outcomes?) |
|
|
|
|
I have not seen/studied literature on sensitivity analysis with more than 1 mediator. It may well be possible. I don't know if there are alternatives. I agree that doing 1 mediator at a time changes the model.
You can ask on SEMNET. |
|
|
|
|
| Thanks again. Will do |
|
|
|
|
Hello,
Can you recommend any approach for a sensitivity analysis of the assumption of no mediator outcome confounding when the outcome is binary?
Is any kind of sensitivity analysis appropriate with a binary outcome?
Also, what is the reason why PLOT: TYPE = plot3 sensitivity is not available for binary outcomes? Is it because the residual covariance isnt available for the binary outcome?
thank you |
|
|
|
|
| Yes, I think there are articles by Tyler Vanderweele dealing with the binary sensitivity case. Our RMA book refers to a 2010 article by him "Odd ratios..." in Amer J of Epi. |
|
|
|
|
| PS I haven't explored the binary case. |
|
|
|
|
Thank you, very useful reference. I also found Vanderwheele references another paper of theirs 'Bias formulas for sensitivity analysis for direct and indirect effects' - also useful. But i cant understand the formula well enough to attempt to apply this practically in mplus - is there anything i can do to address the question of unmeasured confounding? Or is this a lost cause?
I think i should look both at exposure-mediator confounding and mediator-outcome confounding.
Thank you for any advice. |
|
|
|
|
| I think maybe David McKinnon at ASU has worked in this area. |
|
|
|
|
Thanks for McKinnon suggestion its useful.
some related questions:
when a unmeasured m-y confounder exists it causes the residuals for m & y to correlate. Where in the residuals output can i see these correlations (sorry im confused by the output)?
How can i judge the size of these correlations? And if they are small, does this go someway to evaluating risk of unmeasured confounding as small?
Is it possible to run sensitivity plots for x-m confounding?
if i treat my binary outcomes as continuous then i can run the sensitivity plots - is this at all acceptable?
The 'sensitivity plot is used to determine at which rho values the ci does not include zero'- how do you judge whether the value is too strict and therefore leads you to question the result? do you look at the overall residual correlation values for those 2 variables and see see how it compares?
sorry for all the never ending questions about this!
thank you |
|
|
|
|
Q1-Q2: You don't find this in the output because the model is just-identified and will have zero residuals. The residual correlations gets absorbed in the model parameters so that they are distorted. Hence the sensitivity analysis.
Q3: We don't have that.
Q4: Gives a rough approximation.
Q5-Q6: No firm rules. It's up to your judgement. |
|
|
| Back to top |
