2
$\begingroup$

How does Stata create their margins dy/dx output? I would like to recreate some Stata output in SAS. The results of a logistic regression give a coefficient of 1.4921 for my dummy variable, and an odds ratio of 4.447 (or the exponentiation of 1.4921). Stata has a margins dy/dx tool that gives a value of 0.1309 for this variable. I interpret this to mean there is a 13.9% difference between the probability of each outcome for this 0 1 indicator.

How can I generate this 13.9% difference outside of Stata?

$\endgroup$
  • $\begingroup$ I might be able to have a go at answering if you put your whole output here. $\endgroup$ – Jeremy Miles Jan 28 '15 at 18:12
  • $\begingroup$ I'd recommend checking Stata manuals. They are freely available online as pdf files. $\endgroup$ – Roberto Ferrer Jan 28 '15 at 18:32
6
$\begingroup$

For a continuous variable, the marginal effect of $x_k$ in a logit is

$$\Lambda(X'\beta)\cdot \left[1-\Lambda(X'\beta)\right]\cdot\beta_k,$$ where $$\Lambda(z)=\frac{\exp{z}}{1+\exp{z}}.$$

By default, Stata actually calculates the average of this over the estimation sample, but I will use the mean value of x in what follows (marginal effect at the mean rather than average marginal effect) to get the point across:

. sysuse auto, clear
(1978 Automobile Data)

. logit foreign mpg

Iteration 0:   log likelihood =  -45.03321  
Iteration 1:   log likelihood = -39.380959  
Iteration 2:   log likelihood = -39.288802  
Iteration 3:   log likelihood =  -39.28864  
Iteration 4:   log likelihood =  -39.28864  

Logistic regression                               Number of obs   =         74
                                                  LR chi2(1)      =      11.49
                                                  Prob > chi2     =     0.0007
Log likelihood =  -39.28864                       Pseudo R2       =     0.1276

------------------------------------------------------------------------------
     foreign |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         mpg |   .1597621   .0525876     3.04   0.002     .0566922     .262832
       _cons |  -4.378866   1.211295    -3.62   0.000    -6.752961   -2.004771
------------------------------------------------------------------------------

. di _b[mpg] * exp(_b[mpg]*21.2973+_b[_cons])/(1+exp(_b[mpg]*21.2973+_b[_cons]))* (1-exp(_b[mpg]*21.2973+_b[_cons])/(1+exp(_b[mpg]*21.2973+_b[_cons])))
.03175262

. margins, dydx(*) atmean

Conditional marginal effects                      Number of obs   =         74
Model VCE    : OIM

Expression   : Pr(foreign), predict()
dy/dx w.r.t. : mpg
at           : mpg             =     21.2973 (mean)

------------------------------------------------------------------------------
             |            Delta-method
             |      dy/dx   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         mpg |   .0317526   .0103945     3.05   0.002     .0113798    .0521254
------------------------------------------------------------------------------

For dummy variables, Stata uses a finite difference method (average of predicted probabilities with $x_i$ set to 1 minus predicted probability with $x_i$ set to zero).

Standard errors are a bit more complicated.

$\endgroup$
  • $\begingroup$ Useful answer, although I'm not sure about the clarity of the link between the opening equations and the formula actually used in the di part of the Stata code. It took me 10 good minutes to decompose the code to reconstitute the equation :) $\endgroup$ – Fr. Mar 19 at 9:22
  • $\begingroup$ @Fr. What was confusing about it? Is it the means? I suppose the index function coefficient in the code comes first rather than last, but that is the only difference. Please feel free to propose an edit that you think works better. $\endgroup$ – Dimitriy V. Masterov Mar 19 at 14:15
  • $\begingroup$ I think one way to make the code easier to parse is (1) to define $z$: sca de z = _b[mpg] * mu_x + _b[_cons], (2) then $\lambda_z$: sca de lambda_z = exp(z) / (1 + exp(z)), and finally (3) di lambda_z * (1 - lambda_z) * _b[mpg] to display the estimated effect. $\endgroup$ – Fr. Mar 20 at 16:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.