Dominance analysis with imputed values and fixed effects

[Hannah Glinka]

Dear Statalists,
I am doing dominance analysis (using domin-ado from SSC by Luchman¹ in Stata 16.1) for the first time and regarding this 3 questions arised which I hope some of you can help me with.

1) My dataset contains imputed values. I was wondering, if using these imputed values in dominance analysis can cause any problems? Are there any arguments against using imputed values in dominance analysis?
I am using following command:

mi estimate, post: domin dependvar independvar1 independvar2, reg(reg, abs(independvar3)) fitstat(e(r2_p))

2) Moreover I was wondering if I can consider in dominance analysis that my model contains fixed effects? Explicit: Can I use the regular regress-command with fixed effects as an option for dominance analysis as well? See command above.

3) Also: Is R² still a valid indicator for goodnes of fit after doing dominance analysis with fixed effects (if possible)?

Can any of you help me with these issues?
Thank you in advance!

Best wishes,
Hannah

¹Luchman, Joseph, 2021: Determining relative importance in Stata using dominance analysis: domin and domme. The Stata Journal 21,2: 510-538. DOI: 10.1177/1536867X211025837. (https://journals.sagepub.com/doi/full/10.1177/1536867X211025837)

[Joseph Luchman]

Hi Hannah,

Some responses.

1) Recommend reading over -domin-'s helpfile. The -mi- option allows an analyst for data that are -mi set-. -domin- does not accept the -mi estimate- prefix directly.

The only downside to using multiply imputed data in the dominance analysis/DA is the time it takes to complete (which could be substantial depending on the number of imputations and the number of independent variables). There is no conceptual downside - if the multiply imputed data is valid for the underlying model, then it will be valid for the dominance analysis based on it.

2) If by fixed effects, you mean indicator or dummy codes for some set of categories (assume this is what is meant; i.e., in the style of -xtreg, fe-), then yes.

The example provided will not work, however. You will need to call a command that estimates a fixed effects model or generate them yourself if using -regress-

3) This depends on the R² (or more broadly fit statistic) that is being used. Note that -e(r2_p)- is not a value eretuned by -regress-.


Dominance analysis is just a method for subdividing a fit statistic based on independent variables (or parameters) in a model. If the fit statistic applies to the model, then any independent variables (or parameters) in it could be dominance analyzed. This method, as some in the data science community put it, is "model agnostic" - any model that produces a fit statistic and has predictors or parameters could be used.

[Hannah Glinka]

Hi Joseph,
thank you for your answers, this helps a lot!

Best wishes,
Hannah