Logistic regression and marginal effects differences

#1
Hello all!

First I would like to say that I have learned a lot from reading this forum, thank you for the great resource on statistics!

I do however have a question that I cannot seem to get my head around. In my research I am running a logistic regression which yields nice results, but I would also like to see the marginal effects of the results.
The problem: When I do run the marginal effects for my data there seems to be quite a big difference between the logistic regression and the corresponding marginal effects. There are quite some coefficients (logistic) that are significant which turn not significant when I look at the marginal effects.

From what I can find this is not uncommon since they are essentially different hypotheses. However, when I look at earlier research that uses basically the same variables as I do they seem to always give the same significance (logistic & marginal effects), which is not the case for me. Are there other differences that I am not aware of that could cause this? Are the marginal effects so small that maybe they are not significant in that output?

Setup: I am using STATA 14.0/SE, running xtlogit regressions and obtaining marginal effects using 'margins, dydx(*)'. Dependent variable is a dummy variable with values of 0 and 1.

Thanks in advance & all the best,

Marcus
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
@Marcus Diller - glad you are enjoying the site. I have my own systemic issue here. In particular what do you mean when you write "marginal effects". This concept in general has been confusing to me per a definitional perspective. Are you meaning the difference between type I and type III effects, so after controlling for another variable, the target exposure has an effect of blank versus when not initially controlling for a variable in the same model? Or are you referencing comparing effects between the saturated and a nested model?

Thanks!
 
#3
Hi hlsmith,

Thank you for your fast reaction :)! I will try to explain it as detailed as possible in a reduced format to make clear what I try to do:

- Dependent variable, dummy: 0 if 'not effective', 1 if 'effective
- Dataset has entries for certain 'events' on a company, which are proposed by a certain stakeholder
- Each event can either be effective or not effective, as noted by the dummy
- Multiple firm controls

I run a logistic regression to see if certain stakeholders are more likely to propose an 'effective' event. So in the regression output for example I find that a certain stakeholder has a negative & significant coefficient, implying that this stakeholder is more likely to propose 'not effective events' (It is my first time using logistic regressions, should my interpretation be wrong I would love to hear :)).

However, I want to be able to interpret this coefficient in more ways than only its direction. In related research I find that besides the coefficient, also the 'marginal effect' is reported. Using this value related research states, for example 'events proposed by stakeholder x are ...% more likely to be effective'.

EDIT: So my issues then is that some of the coefficients are significant, while quite a lot of the corresponding marginal effects are not. Would I still be able to interpret the marginal effects if they are not significant, while their corresponding coefficients in the original regression are? I do not see such a discrepancy in related research.

I hope that makes it clear,

All the best,

Marcus
 
Last edited: