[Vasilis Papas]

Maybe i dont have things quit clear in my head. For example in the first specification of

xtreg Loan_ratio Loan_ratio_lagged c.RD_ratio_lagged##i.NIRP , fe robust

, by doing

margins, dydx(RD_ratio_lagged NIRP)

the value that i will get for RD_ratio_lagged will show me the average change in Loan_ratio for a unit change in RD_ratio_lagged, regardless of whether negative rates are active or no. It does not show me the difference in the average marginal effect between NIRP=0 and NIRP=1, correct? And if i want to know what is the difference in the average effect for a unit change in RD_ratio_lagged between the periods of positive and negative ratesi believe i should use either

margins NIRP, dydx(RD_ratio_lagged)

or

margins, over(NIRP) dydx(RD_ratio_lagged)

? I hope i make any sense

In general i would like to compare how a change in RD_ratio_lagged affects Loan_ratio before and after NIRP and how NIRP affects loan ratio too

And later to compare how a change in RD_ratio_lagged affects Loan_ratio before and after NIRP for different bank groups

[Clyde Schechter]

Maybe i dont have things quit clear in my head. For example in the first specification of Code:

xtreg Loan_ratio Loan_ratio_lagged c.RD_ratio_lagged##i.NIRP , fe robust
, by doing Code:

margins, dydx(RD_ratio_lagged NIRP)
the value that i will get for RD_ratio_lagged will show me the average change in Loan_ratio for a unit change in RD_ratio_lagged, regardless of whether negative rates are active or no. It does not show me the difference in the average marginal effect between NIRP=0 and NIRP=1, correct?

Not correct. It will show you both, in separate rows of the output table, because you asked for both by including both in the -dydx()- argument.

And if i want to know what is the difference in the average effect for a unit change in RD_ratio_lagged between the periods of positive and negative ratesi believe i should use either
Code:

margins NIRP, dydx(RD_ratio_lagged)
or Code:

margins, over(NIRP) dydx(RD_ratio_lagged)
?

Well, "use either" in the sense that either of them will give you a particular "species" of the average effect for a unit change in RD_ratio_lagged conditional on periods of positive or negative interest rates. But they are not the same thing, and depending on exactly what conclusions you are going to try to draw, one of them will be correct and the other will be wrong. (OK, with extremely low probability, by chance they could turn out to be the same number, but don't count on that.) So you still need to figure out which of these is the answer to your question--probably you need to focus your question more clearly than you have so far in order to resolve this. Remember that the difference between them is that with -over()- you do not fully adjust for other variables in the model, where as with -margins NIRP,...-, you do.

[Vasilis Papas]

Thank you once more for your time and patience! I believe things are much clearer for me now.

[Clyde Schechter]

Let me try to also clarify when you would use -margins, over(NIRP) dydx(RD_ratio_lagged)- and when you would use -margins NIRP, dydx(RD_ratio_lagged)-.

Again, let's consider just your simplest interaction model, with NIRP, RD_ratio_lagged, and their interaction, but no other variables (except the fixed effects). The marginal effect of NIRP conditional on RD_ratio_lagged depends on the value of RD_ratio_lagged. Consequently, if the distributions of RD_ratio_lagged are different in the NIRP = 0 and NIRP = 1 conditions, the average marginal effect of NIRP will generally differ between NIRP = 0 and NIRP = 1. So the question we now have to ask is how did it come to pass that the distributions of RD_ratio_lagged differ in these two conditions. There are two different situations.

In situation 1, there is no particular systematic reason why the distributions should differ. It is just the luck of the draw in our particular sample that they differed in the way, and to the extent that they did. In this situation, then, the distribution of RD_ratio_lagged in the entire data set, being a larger sample, is the best representation of the distribution of RD_ratio_lagged, and we should do our calculations using that full data set. In other words, we should use -margins NIRP, dydx(RD_ratio_lagged)-.

In situation 2, the variables NIRP and RD_ratio_lagged are truly related to each other: if a new change of state between NIRP 0 and NIRP 1 were to occur next year, we would expect a concomitant change in the distribution of RD_ratio_lagged. In this case, it is not appropriate to consider the distribution of RD_ratio_lagged conditional on NIRP = 0 to calculate a marginal effect conditional on NIRP = 1 or vice versa. So in this case, -margins, over(NIRP) dydx(RD_ratio_lagged)- gives the appropriate statistic, because it respects the systematic difference in the RD_ratio_lagged distributions between the two NIRP states.

So what you need to ask yourself is which situation better describes the relationships active in the world you are trying to model.