A Mediation Effect Analysis Challenge: Categorical IV, Continuous MV, and Ordinal DV

[Jae Park]

Dear all,

Hi! My name is Jae and I have an inquiry about mediation effect analysis in regression (NOT SEM).

I am using these variables: one categorical IV (Generation), one continuous MV (level of trust), and ordinal dependent variable (frequency of political participation).

While some may argue that my MV, level of trust itself, must be considered ordinal variable,
let's consider it continuous for now.

I was able to find commands to use for a mediation effect analysis with a categorical IV with STATA here:
http://www.ats.ucla.edu/stat/stata/f..._mediation.htm

But the only problem is that this page only talks about the case that DV is a continuous variable (therefore regression analysis).

I've looked up some methodology papers to deal with this problem but the closest thing I could find was a paper from Vanderbilt Univ.
Unfortunately, it does not offer detailed mathematics for ordered logit or logistic regression.


I am writing here to ask if anyone's done it before with STATA.
Do any of you happen to know any papers regarding this model?

Any tips will be graciously taken and extremely helpful.


Thank you so much for your time!

[Mike Lacy]

The -khb- package should do the trick: -ssc install khb-. It allows many regression procedures, including -ologit- and -oprobit-.

[Jae Park]

Thank you so much for your feedback, Mike! I will try that for sure. I'm reading few articles about KHB beforehand. I appreciate your help!

[Roman Mostazir]

Noting your reservation about "[NOT SEM]", -sem- will not be able to fit this anyway as -sem- is for continuous outcome. But -gsem- will fit it.

[Jae Park]

Dear Roman,
Thank you for your feedback. I was not familiar with the SEM in the first place. I'm a little ashamed and glad that you corrected me here.

[Jae Park]

I was able to find commands to use for a mediation effect analysis with a categorical IV with "STATA" here:

I would like to acknowledge my mistake in this statement and correct "STATA" to "Stata".

[wbuchanan]

Jae Park the difficulty with these types of models in general has to do with the scaling of the parameters. For example, the threshold parameters will be in logits/odds-ratios but how would you compare those values with a parameter estimating the mean of the mediating variable? This also affects the decomposition of the parameters into direct and indirect effects. However, just based on the description of the variables it isn't clear if you are intending to test a mediating relationship (e.g., the indirect effect of generation on political participation via level of trust) or a moderating relationship (e.g., whether or not the effect of generation on political participation varies across values of level of trust). If what you actually needed is moderation, you can use factor variable notation to get the interaction between the IV and level of trust and can subsequently use the margins and marginsplot commands to illustrate the relationships.

[Jae Park]

wbuchanan, Thank you for your comment. As a matter of fact, I do intend to analyze the indirect effect of trust in governmental organizations on the causal relationship between generation and political participation. While moderation effect by interaction term (of trust and generation) does not hypothesize that generation and level of trust has a significant correlation, I argue that generation and level of trust has a statistically significant relationship. The main idea of my hypothesis is that generation is a sub-category of our society that generates cultures and customs, which, I suppose, would affect one's level of trust. As a result of ANOVA analysis, I was able to reject the null hypothesis. As of now, I plan to stick with my original model of mediation analysis while the analysis itself is quite challenging.

EDIT: However your point about the parameter estimation issue rang a bell. I will have to work on the rationalization of my mediation analysis and improvement of the model.

[wbuchanan]

Jae Park you can find some decent information via www.statmodel.com where Muthén discusses some of the issues with scaling that result from categorical (e.g., nominal/ordinal) mediators. In your particular case, it may actually make it easier to treat your level of trust variable as an ordinal scale measure (which it sounds like it is) since it will take care of the scaling issues for you (e.g., the parameter estimates would all be logits/probits/odds-ratios and would be on the same scale). It isn't clear how you measured level of trust, but if it is some type of latent you could also potentially consider a multiple mediation model using the individual items (which would also have the benefit of allowing you to test whether there is a specific indicator that appears to be driving the results).