Can a mediating variable be mediated by another mediating variable?

[Maudrene Tan]

Hi All,

I hope that someone can help me with this because I can't seem to find anything online.

SO the question is as follows: Is it possible to have a mediating variable being mediated by another mediating variable?

I have 1 DV (depression), 1 IV (poor health) and 3 MV (control, stress, fams). However, I noticed that control, also mediated the relationship between stress and depression. Hence, I wrote my SEM equation as follows:

sem (poorh-> fams) (poorh -> stress) (poorh -> control1) (poorh -> k10_tot) (fams -> k10_tot) (stress -> control1) (stress -> k10_tot) (control1 -> k10_tot), nocapslatent

The output of my results are attached.

Because the indirect effect between poor health and depression is no longer significant, does it mean that this model is null?

I will really appreciate some assistance on this. Thanks.

[Phil Bromiley]

You might look at the FAQ. Statalist participants are somewhat cautious about opening files that could contain malware including Word documents. They also like the statements and actual results.

You say your DV is depression, but I don't see it in the sem statemen - is it k10?. You certainly can have a mediating variable mediated by another variable, if the two are properly identified. With your nocapslatent, I can't tell what is latent and what is measured - you may be building a lot of estimates on very few measured variables. Just because the direct influence of health on K10 becomes insignificant doesn't mean the model is null. Actually, "model is null" is not a term I see much, but it would normally apply to being unable to reject the hypothesis that all the parameters are zero or something - the aggregate test SEM folks like. The specific path becoming insignificant might mean that the influence of Poorh on k10 is largely mediated by the other variables. You need to estimate a total effect (i.e., include the mediating paths) to understand poorh's influence on k10.

[Maudrene Tan]

Thank you so much for this. I'm new to SEM and the help rendered is greatly appreciated. I wanted to show the diagram that I used to create the model. That was why I attached the Word document.

Apologies for not being clear. Yes, the DV is K10. Let me paste the results that I got here to give you an idea of what my output was.



. sem (poorh -> fams) (poorh -> stress) (poorh -> control1) (poorh -> Dep) (fams -> Dep) (stress -> control1) (stress -> Dep) (control1 -> Dep), nocapslatent
(58 observations with missing values excluded;
specify option 'method(mlmv)' to use all observations)

Endogenous variables

Observed: fams stress control1 k10_tot

Exogenous variables

Observed: poorh

Fitting target model:

Iteration 0: log likelihood = -7050.4519
Iteration 1: log likelihood = -7050.4519

Structural equation model Number of obs = 1013
Estimation method = ml
Log likelihood = -7050.4519

-------------------------------------------------------------------------------
| OIM
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
Structural |
fams <- |
poorh | -.0480998 .0135619 -3.55 0.000 -.0746807 -.021519
_cons | .9039747 .0282806 31.96 0.000 .8485457 .9594038
------------+----------------------------------------------------------------
stress <- |
poorh | -.1101989 .0238074 -4.63 0.000 -.1568607 -.0635372
_cons | 1.822457 .0496456 36.71 0.000 1.725154 1.919761
------------+----------------------------------------------------------------
control1 <- |
stress | -.2000265 .0511408 -3.91 0.000 -.3002606 -.0997924
poorh | -.1410787 .0391588 -3.60 0.000 -.2178286 -.0643288
_cons | 3.017059 .1233552 24.46 0.000 2.775287 3.25883
------------+----------------------------------------------------------------
Dep <- |
fams | -.8524807 .2755156 -3.09 0.002 -1.392481 -.31248
stress | 1.18689 .155889 7.61 0.000 .8813527 1.492427
control1 | -.9128712 .096379 -9.47 0.000 -1.101771 -.7239718
poorh | .3009289 .1197807 2.51 0.012 .0661631 .5356948
_cons | 13.09623 .5152412 25.42 0.000 12.08638 14.10609
--------------+----------------------------------------------------------------
Variance |
e.fams | .149903 .0066607 .1374006 .1635431
e. stress | .4619491 .020526 .4234209 .5039831
e.control1 | 1.223879 .0543812 1.121803 1.335243
e. Dep | 11.20162 .497727 10.26737 12.22089
-------------------------------------------------------------------------------
LR test of model vs. saturated: chi2(2) = 29.00, Prob > chi2 = 0.0000

. estat teffects


Direct effects
-------------------------------------------------------------------------------
| OIM
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
Structural |
fams <- |
poorh | -.0480998 .0135619 -3.55 0.000 -.0746807 -.021519
------------+----------------------------------------------------------------
stress <- |
poorh | -.1101989 .0238074 -4.63 0.000 -.1568607 -.0635372
------------+----------------------------------------------------------------
control1 <- |
stress | -.2000265 .0511408 -3.91 0.000 -.3002606 -.0997924
poorh | -.1410787 .0391588 -3.60 0.000 -.2178286 -.0643288
------------+----------------------------------------------------------------
Dep <- |
fams | -.8524807 .2755156 -3.09 0.002 -1.392481 -.31248
stress | 1.18689 .155889 7.61 0.000 .8813527 1.492427
control1 | -.9128712 .096379 -9.47 0.000 -1.101771 -.7239718
poorh | .3009289 .1197807 2.51 0.012 .0661631 .5356948
-------------------------------------------------------------------------------


Indirect effects
-------------------------------------------------------------------------------
| OIM
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
Structural |
fams <- |
poorh | 0 (no path)
------------+----------------------------------------------------------------
stress <- |
poorh | 0 (no path)
------------+----------------------------------------------------------------
control1 <- |
stress | 0 (no path)
poorh | .0220427 .0073782 2.99 0.003 .0075816 .0365038
------------+----------------------------------------------------------------
Dep <- |
fams | 0 (no path)
stress | .1825984 .046685 3.91 0.000 .0910976 .2740993
control1 | 0 (no path)
poorh | .0188747 .0552404 0.34 0.733 -.0893945 .127144
-------------------------------------------------------------------------------


Total effects
-------------------------------------------------------------------------------
| OIM
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
Structural |
fams <- |
poorh | -.0480998 .0135619 -3.55 0.000 -.0746807 -.021519
------------+----------------------------------------------------------------
stress <- |
poorh | -.1101989 .0238074 -4.63 0.000 -.1568607 -.0635372
------------+----------------------------------------------------------------
control1 <- |
stress | -.2000265 .0511408 -3.91 0.000 -.3002606 -.0997924
poorh | -.119036 .0390427 -3.05 0.002 -.1955582 -.0425137
------------+----------------------------------------------------------------
Dep <- |
fams | -.8524807 .2755156 -3.09 0.002 -1.392481 -.31248
stress | 1.369488 .1627295 8.42 0.000 1.050544 1.688432
control1 | -.9128712 .096379 -9.47 0.000 -1.101771 -.7239718
poorh | .3198036 .1272479 2.51 0.012 .0704023 .569205
-------------------------------------------------------------------------------

Separately, I noticed that the mediating effect of stress on poorh and Dep was inconsistent. Does this suggest that I should not include this path in the SEM?

Once again, I'm really appreciative for the advice and time that you'd spent guiding me on this. Many thanks!

[Dick Campbell]

I know it must be frustrating to be asked to re-post your question several times, but this output would be a lot easier to read if you could enclose it in CODE tags so that it aligns properly See the FAQ entry 12.3 for instructions on how to do this. I did pull up your original attachment and looked at the the path diagram and at your equations.. I don't understand your statement "stress on poorh and Dep was inconsistent." Do you mean the effects of poor health and stress on depression? And what do you mean by "inconsistent?"

Aside from all that, your model has three intervening variables between poor health and stress. There is nothing inherently wrong with that, and although I don't quite understand your original question ("Can a mediating variable be mediated by another mediating variable?") the model itself is not all that usual as these things go.

Note that FAMS and STRESS have no effect on each other and that both have paths from the sole exogenous variable. This means that you could and probably should allow their errors in equations (e1 and e2) to be correlated. As it stands,the model does not fit the data all that well and you will find that the goodness of fit is likely to increase if you estimate that covariance parameter. Use estat gof to get the goodness of fit statistics.

Finally, you should realize that most models like this are essentially descriptive. While the model makes various causal assertions, you are not really testing them, even if you have longitudinal data. I often tell people that the way to think about these models is to say "If this model represents the way the world works, then here are the quantitative implications."

[Maudrene Tan]

Dear Prof Campbell,

Thanks so much for the explanation. When I analyzed the mediating effect of Stress on Poorh and Dep separately, I noticed that there is a negative indirect effect. I googled it and found that this negative indirect effect is termed inconsistent mediation

Indirect effect = -.151042 .037414 -4.037 .000054

Proportion of total effect that is mediated:  -.47488436
Ratio of indirect to direct effect:           -.32198074
Ratio of total to direct effect:              .67801926

As a result of this large negative indirect effect, the total indirect effect of my model became insignificant, which suggests that I'm unable to reject my hypothesis that together, Stress, FamS and Control mediate the relationship between Poorh and Dep. Given this large negative indirect effect, will it be right for me to not create this path in the model? The estat gof dramatically improved when I removed the path between Poorh and Stress.

Thank you again for the time spent on this.