New variants of marginal effects available for logistical regressions

ByTrond Pedersen

It is now possible to calculate marginal effects for logistic regression analyzes in other ways than the default value “overall”, using the new option “mfx_at ()”.

Logistic regression analyzes run by the logit, probit and mlogit commands provide logistic coefficient estimates by default. However, many people prefer to look at marginal effects, as these are easier to interpret. For this, the option mfx() is used, where you can select the variants dydx, dyex, eydx, and eyex. More info about these variants can be found in the user manual chapter 5.6.2 or by using the command help logit (or one of the other logistical commands).

By using the new option mfx_at(), you can override the default measure. The following variants are available:

  • mfx_at(overall) (mean of the marginal effects measured over all x values) (default measure if this option is not used)
  • mfx_at(mean) (marginal effect measured at mean value of x)
  • mfx_at(median) (marginal effect measured at median value of x)
  • mfx_at(zero) (marginal effect measured at 0-value of x)

The mfx_at() option is usually used in combination with mfx(), for example:

logit high_income male married age high_wealth, mfx(dydx) mfx_at(mean)

However, you may also just use mfx_at(). Thus, the standard variant mfx(dydx) is used.

The following alternative regression expressions will present the same marginal effect values:

logit high_income male married age high_wealth, mfx(dydx) mfx_at(overall)
logit high_income male married age high_wealth, mfx_at(overall)
logit high_income male married age high_wealth, mfx(dydx)