Interaction methods

[ALKEBSEE RADWAN]

dear all,
I have a question regarding understanding the outcome of interaction coefficients.
I first used a split method to examine the moderating effects of internal control system using a dummy variable (icw= 1 if the internal control system is weak, zero otherwise). I found that the main variable is negatively significant for sample of weak internal control and insignificant for those of weak internal control system.
A reviewer asked to use the interaction method not the split method.
I did it, but I got confused by the outcomes: The interaction variable’s coefficient comes out positive and significant.

Do you think it’s consistent?

I need an explanation for the ideal situation with both style.

looking forward to receiving your responses

[Carlo Lazzaro]

Alkebsee:
as per FAQ, please share what you typed and what Stata gave you back (via CODE delimiters, please). Thanks.

[Clyde Schechter]

I found that the main variable is negatively significant for sample of weak internal control and insignificant for those of weak internal control system.

This statement contradicts itself. Surely one of these refers to the non-weak internal control system group. But which one?

The interaction variable’s coefficient comes out positive and significant.

Do you think it’s consistent?

Quite possibly. The interaction coefficient is the difference between the expected outcome for the weak internal control group and the other group. Since you coded the group indicator as 1 for weak internal control and 0 for other, this interaction coefficient says that the effect of whatever variable you interacted it with is higher (which might be more positive, or less negative) in the weak internal control group. So, in a general sense, this may well be consistent with what you found in the two-regressions approach. But other considerations might also be in play, and those considerations are matters of guesswork because your post has given so little information about what you did and what results you got.

To get a clear and definite response you need to show the actual regression commands and results for the original two models and the interaction model.

Added: Crossed with #2.

[ALKEBSEE RADWAN]

Oh very sorry, it should be as follows:
I found that the main variable is negatively significant for the sample of strong internal control and insignificant for those of weak internal control system.

I will try to give more details. as for the split method result it is shown as follow

the main var (TD) coefficient is -0.066(1.10) when ICW=1, while it is -0.098***(3.06) when ICW=0

When i am doing the interaction method the results of the interaction variable (TD*ICW) is +0.014**(1.94)

I hope it is clear now for you Clyde Schechter Carlo Lazzaro

Thank you very much

[ALKEBSEE RADWAN]

Quite possibly. The interaction coefficient is the difference between the expected outcome for the weak internal control group and the other group. Since you coded the group indicator as 1 for weak internal control and 0 for other, this interaction coefficient says that the effect of whatever variable you interacted it with is higher (which might be more positive, or less negative) in the weak internal control group. So, in a general sense, this may well be consistent with what you found in the two-regressions approach. But other considerations might also be in play, and those considerations are matters of guesswork because your post has given so little information about what you did and what results you got.

Added: Crossed with #2.

I think you have provided a meaningful explantion here. expecially when you said (..., or less negative).

thank you very much

[ALKEBSEE RADWAN]

Clyde Schechter Again, I have a question regarding understanding the outcome of interaction coefficients.
I first used a split method to examine the moderating effects of internal control system using a dummy variable (FC= 1 if the firm has financial issues, zero otherwise). I found that the main variable is negatively significant for the sample of non-financial issues, while negative and insignificant for those of firms with good situation. That is, the main variable coefficient is -0.0xx(-x.xx) when FC=0, while it is -0.00x(-x.xx) when FC=1
A reviewer asked to use the interaction method not the split method.
I did it, but I got confused by the outcomes: The interaction variable’s coefficient comes out positive and insignificant (0.0xx, (x.xx)).
numbers inside the parantheses reperesent T-value.
when I used the sub-sample method and my conclusion was that firms with financial issues can’t be postive interms of the main variable, That is, the main variable coefficient is significant -0.0xx(-x.xx) when FC=0, while it is insignificant -0.0xx (-0.xx) when FC=1.

But when I used the interaction method, the interaction variable’s coefficient comes out positive and insignificant (0.0xx, (x.xx)).
Note:
My query is: Can I argue that both methods are consistent?

[Carlo Lazzaro]

Alkebsee:
as per FAQ, please share what you typed and what Stata gave you back (via CODE delimiters, please). Thanks.

[Clyde Schechter]

There are two ways to use interaction models, and, as Carlo notes, without seeing the actual code and output we can't tell which one you did. Moreover, you are saying you got 0 coefficients for both subsets in the subset analysis and 0 for the interaction coefficient. That's perfectly consistent--but since exactly zero coefficients actually are very unusual in real life, I suspect you have made some typographical errors in reporting your results in #6. Also a truly 0 coefficient can never be "statistically significant."

So, yes, we do need to see the actual code and Stata output here to know for sure what is going on.