I have the following dilemma:

I understand-ish what marginal effects are, also the calculation of it, derivation of the sigmoid function and how to interpret it (as a the change in probability by increasing your variable of interest by "a little bit", this little bit being 1 for discrete vars or by a std(x)/1000 for continuous ). Now, the part I find tricky is to corroborate the results of the marginal effects by hand and recalculating the probabilities for x=0 and then x=1 (for example) and then get a difference in probability equal to the marginal effect I got earlier, I am particularly stuck with dummy variables since If I increase one, I have to decrease the other one, so I am not so sure how to work around it and interpret it. (this question also applies for highly correlated variables)

To make it more clear, let's say I have the following dataset:


[1. , 0. , 0. , 4.6, 3.1, 1.5, 0.2],
[1. , 0. , 1. , 5. , 3.6, 1.4, 0.2],
[1. , 1. , 0. , 5.4, 3.9, 1.7, 0.4],
[1. , 0. , 1. , 4.6, 3.4, 1.4, 0.3],
[1. , 1. , 0. , 5. , 3.4, 1.5, 0.2],
[1. , 0. , 0. , 4.4, 2.9, 1.4, 0.2],
[1. , 0. , 1. , 4.9, 3.1, 1.5, 0.1],
[1. , 1. , 0. , 5.4, 3.7, 1.5, 0.2],

Var_0 = What will be the intercept.

Var_1, var_2 = (2/3 binary dummies), one dropped to avoid co linearity.

Var 3+ = Normal continuous variables


[ 7.56986405,  0.75703164,  0.27158741, -0.37447474, -2.79926022, 1.43890492, -2.95286947]




0.0769344 , 

Marginal effect = p*(1-p) * B_j

Now let's say that I am interested in the marginal effect of var_1 (one of the dummies), I will simply do: p*(1-p) * 0.7570

Which will result in an array of length n (# of obs) with different marginal effects (which is fine because I understand that the effects are non constant and non-linear). Let's say this array goes from [0.0008 to 0.0495]

Now the problem is, how can you verify this results? How can I measure the marginal effect when the dummy goes from values 0 to 1?

You could argue that I could do two things MEM and AME methods:

  1. MEM: Leave all the values at its mean and then calculate all over again for var_1 = 0 and then for var_1 = 1 (MEM method)

    (you can't really do this because that you will be assuming that you can have some observations where var_1 and var_2 will be equal to 1 at the same time, which incorrect since the mean for a dummy is like a proportion of how many "1s" there are for that column)

  2. AME: Leave as observed, but changing all the values of var_1 to 0 (making all the values of var_2 = 1) and then do the opposite (var_1 = 1, var_2 =0, you have to do this since it can't belong to two categories at the same time), and then take the average of the results (AME method) (Side comment:One thing I am not sure if it is the average between the difference in marginal effects when var_1 = 0 and then 1, or if it is an average between the probabilities when var_1 =0 and then 1, I used both, but probability I think it makes more sense to me)

Now, if I try the 2nd approach I get very different results to what I originally got ( which were values between [0.0008 to 0.0495]), it gives me values between [0.0022 to 0.1207], which is a massive difference.

To summarise:

  1. How can do a mathematical corroboration to get the same values I got initially from the theoretical formula of marginal effect = p* (1-p)* B_j, which was ([0.0008 to 0.0495]). There should be a method to arrive to the same number (I am using AME at the moment)

  2. How can I interpret these original values in the first place? Because if I take 0.0495, I am basically saying, if I increase var_1 by 1-unit (from 0 to 1), I will have a 4.95% increase in probability of my event happening, the problems is that it doesn't consider that to make the 1-unit increase I need to, by default, decrease the other dummy variable (var_2), so I will be doing something of a double-change in the variables or like a double marginal effect at the same time.


For connected dummies $d$ and $x$, you might want to calculate this average of finite differences:

$$AME_x =\frac{1}{N} \cdot \sum_{i=1}^N \left[ \hat p(d=1,x=0,z=z_i)-\hat p(d=0,x=1,z=z_i) \right],$$

where $\hat p(.)$ is the predicted probability from the logit model. I don't know what encoding you are using, so replace that for the ones and zeros above.

Note that std(x)/1000 for continuous variables is not quite right. If you recall the definition of derivatives, the limit of the change goes to zero. You are considering a tiny perturbation, not one of particular size that depends on the SD of x.

  • $\begingroup$ Hello Dimitry, thanks for you comment, that is currently the approach I am taking with binary encoding. (although I think in the formula you forgot to divide by "n"). The problem I have with this approach is that you can calculate the marginal using the theoretical formula `p*(1-p)*B_j using the unaltered version of you dataset (without evaluating for x = 0 and then x=1) that should give you in theory the "correct" marginal effect, and by doing some method (I am choosing AME) you should be able to arrive at the same result, the problem is that the result of AME is not the same, not even close $\endgroup$ – Felipe Araya Feb 10 '20 at 17:42
  • $\begingroup$ With regards to the std(x), yes you are right. Some people say 0.0001/10000 as the increase, some others say the std(x)/1000, I am not so sure which one to take since I haven't seen any conclusive answer for it.What would you recommend? $\endgroup$ – Felipe Araya Feb 10 '20 at 17:49
  • $\begingroup$ Thanks for the $1/N$ correction. I think you need to predict $p$ and $1-p$ with the corresponding dummy/dummies to zero, and not own data. Even with that change, the derivative and the finite difference formula will not agree exactly if the function is very curvy. Derivatives are for epsilon changes, finite differences are for 1 unit changes. $\endgroup$ – Dimitriy V. Masterov Feb 10 '20 at 18:20
  • $\begingroup$ To put in another way, take the continuous variable case. Suppose we have $E \left[ y \vert x \right] =F(a + b \cdot x + c \cdot z ).$ Then $\frac{\partial E \left[ y \vert x \right]}{\partial x} =F'(a + b \cdot x + c \cdot z ) \cdot b.$ This tells you the change when $x$ increases by 1. But when you evaluate it, you don't calculate $F'(a + b \cdot (x+1) + c \cdot z ) \cdot b,$ but use the original value of $x$. For dummies, you have to do something similar and use a sensible baseline for the change ($0 \rightarrow 1$). $\endgroup$ – Dimitriy V. Masterov Feb 10 '20 at 18:40
  • $\begingroup$ I would just say "for small change in x" for the continuous derivatives. You can actually calculate finite differences for continuous variables to where you can add the same particular value to everyone's value of $x$ that you can pick to be whatever you want. $\endgroup$ – Dimitriy V. Masterov Feb 10 '20 at 18:42

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.