Multinomial Logistic Regression: significant odds ratios and nonsignificant average marginal effects

[Rashad Jafarov]

Dear All,

I am performing a multinomial logistic regression. Below I present an extract from the regression output that is relevant for my question. As you can see from the table below, almost all the coefficients (odds ratios) are significant.

VARIABLES	Only tuition fee credit (full or partial)	Only loan (a fixed amount per month or sporadic)	Tuition fee credit and a loan
Subjective financial situation
Reference: Reasonable
Very bad	2.406***	1.617***	2.893***
	(0.807)	(0.283)	(0.582)
Bad	1.387*	1.262***	1.760***
	(0.233)	(0.097)	(0.186)
Good	1.228	0.584***	0.685***
	(0.167)	(0.041)	(0.072)
Very Good	0.724**	0.313***	0.362***
	(0.098)	(0.022)	(0.039)

Yet, when I compute the average marginal effects of the same variable, across all the equations (the outcomes of the dependent variable), some of the p-values for the average marginal effects indicate that the values which exhibit significant effect on the odds borrowing in the regression output above are not statistically significant.

	dy/dx	Std. Err.	z	P>z	[95% Conf. Interval]
Subjective financial situation
Reference category: Reasonable
Outcome: Only tuition fee credit
Very bad	0.029	0.019	1.530	0.125	-0.008	0.066
Bad	0.008	0.006	1.240	0.214	-0.005	0.021
Good	0.014	0.005	2.590	0.009	0.003	0.024
Very good	-0.002	0.005	-0.410	0.684	-0.011	0.007
Outcome: Only a loan
Very bad	0.035	0.029	1.190	0.234	-0.023	0.092
Bad	0.019	0.013	1.510	0.130	-0.006	0.044
Good	-0.075	0.010	-7.340	0.000	-0.095	-0.055
Very good	-0.136	0.009	-14.600	0.000	-0.154	-0.118
Outcome: Tuition fee credit and a loan
Very bad	0.078	0.021	3.670	0.000	0.036	0.120
Bad	0.039	0.009	4.450	0.000	0.022	0.056
Good	-0.014	0.007	-2.120	0.034	-0.027	-0.001
Very good	-0.038	0.006	-6.290	0.000	-0.049	-0.026

What would you do in such situation. Should I still go ahead and interpret the average marginal effects, despite their statistical nonsignificance?

[Clyde Schechter]

FIrst, in general there is nothing strange or unusual about this. The coefficients are very different entities from the marginal effects.

But before anyone can give you specific advice about your situation, we need to see more information. Show us the exact Stata commands you used and exactly the output you got from Stata. The coefficient output you show appears to be from -estout- or -outreg- or something like that: the original regression table would be easier to work with here. And the marginal effects output is completely uninterpretable without knowing what the command was.

So go back to the Results window or your log file, copy the commands and Stata response and paste them directly into a code block here. (See FAQ #12 7th paragraph for instructions about creating a code block.) Using code blocks will assure that your output is readable and nicely aligned.

[Richard Williams]

We had a discussion along these lines last month:

http://www.statalist.org/forums/foru...s-significance