marginal effects vs log(odds) in logit

[Mehdi Shiva]

Hello statalist,

I have a dummy variable (happened=1, not happened=0) for my dependent variables and a number of continuous and dummy explanatory variables. I am using a logit regression. I have standardized all my continuous variables in order to get a better interpretation. However, I do not understand why the marginal effects report different values compared to standardized coefficients? I would appreciate your help.

[Richard Williams]

It would help to see your commands and output. But, by default, margins is going to give you predicted probabilities, whereas the model coefficients talk about the effects of variables on the log odds of an event occurring.

[Mehdi Shiva]

It would help to see your commands and output. But, by default, margins is going to give you predicted probabilities, whereas the model coefficients talk about the effects of variables on the log odds of an event occurring.

Thank you Dr. Williams for the respond. Here is the commands I use:
logit intconf $totalmissingpyears_std ,robust cluster(gwno)
margins, dydx(*) post

I wonder, which one is more informative: If I say one standard deviation of x1 increases the log odds of Y by 50% or to say one standard deviation of x1 increases the probability of Y happening by 4%?
(please let me know if its more convenient to copy my results here as well)

[Richard Williams]

There are any number of ways to make your results easier to understand. All have their pros and cons. There is nothing that says they have to be mutually exclusive and you can only use one. You can look at the sign and significance of coefficients; odds ratios; adjusted predictions; marginal effects; and predicted probabilities for prototypical cases. You can combine these with x-standardized coefficients as you seem to be doing. Personally, I think marginal effects of continuous variables are one of the least useful things to use (the interpretation is not as easy as you imply), but economists seem to like them. The latest version of Long and Freese's book overs many of the options:

http://www.stata.com/bookstore/regre...ent-variables/

For continuous variables I am also a big fan of the mcp command. See

http://www.stata-journal.com/article...article=gr0056

For highlights see

http://www3.nd.edu/~rwilliam/stats3/Margins03.pdf

[Mehdi Shiva]

There are any number of ways to make your results easier to understand. All have their pros and cons. There is nothing that says they have to be mutually exclusive and you can only use one. You can look at the sign and significance of coefficients; odds ratios; adjusted predictions; marginal effects; and predicted probabilities for prototypical cases. You can combine these with x-standardized coefficients as you seem to be doing. Personally, I think marginal effects of continuous variables are one of the least useful things to use (the interpretation is not as easy as you imply), but economists seem to like them. The latest version of Long and Freese's book overs many of the options:

http://www.stata.com/bookstore/regre...ent-variables/

For continuous variables I am also a big fan of the mcp command. See

http://www.stata-journal.com/article...article=gr0056

For highlights see

http://www3.nd.edu/~rwilliam/stats3/Margins03.pdf

Thank you very much. I like mcp as well but sometimes people are looking for just "a simple number"!

[Ryan Daly]

I'd recommend reading upon Kennedy (2008), Enders (2010), or Wooldridge (2013) on marginal effects and standard coefficients in regards to dummy variables. Look for explanation on dummy variables within logit models and how to interpret against marginal effects.

From what I can recall and write in one line, you essentially have a mathematical/statistical transformation of data on the coefficients in regards to the logit/log(odds) results. Correct me if I'm wrong, and I apologize if I'm too vague.