I'm running a fracreg regression like this:
fracreg probit meanperunitFLEVLOW100 ZFL ZNonfinancial_MCs ZFinancial_MCs c.ZFL#c.ZNonfinancial_MCs c.ZFL#c.ZFinancial_MCs c.ZFL#c.ZNonfinancial_MCs#c.ZFinancial_MCs stdmeanROE_ROA_ stdmeansales_growperunit stdmeanVA_SALES_ stdmeanROA stdmeanEBIT stdmeanANC stdtotalHrsForm ZBusiness_age i.CEO_Dummy i.intStage i.Sector_Dummy , difficult technique(dfp) het(ZFL ZNonfinancial_MCs ZFinancial_MCs )
You can see that I have created the interactions using the # operator. The variables in the interaction terms are continuous (there are three two-way interactions and one three-way interaction). I use standardized variables because each variable's metrics are different, and I want to avoid the potential impact it might have on estimates.
The question is that I created manually the variables with the interaction terms, without considering the signs. And the results are obviously the same. However, I question on the validity and perhaps should warn against using the # operator in two-way interactions when both variables may have negative values. In short, let's say the interaction is X1*X2. So when we have values (-2) in both the interaction is (+4). And when we have (+2) in both, the interaction is also (+4).
Should we accept this?
I think we should do this manually by recalculating the interaction with negative value when both variables are negative:
gen interaction=X1*X2
replace interaction=-(1)*X1*X2 if X1<0 & X2<0
I think the interpretation this latter one interaction is different to that obtained by using the operator #. With the operator, what is relevant is the absolute value when both have negative values (having too much of both or having very few of both). On the other hand, when using the interaction built manually, when both variables have negative values, the interaction is negatively magnified, so having very few of both will be in a pole, and having too much of both will be in the opposite pole. Also, X1=(-2) and X2=(+3) will be equal to X1=(+2) and X2=(-3) in both cases, and I agree with this idea.
The question is: how should interactions be built when speaking about continuous variables that both may have negative values? Should the interaction # operator be used only when both may have values equal or above zero?
Thanks
Miguel
fracreg probit meanperunitFLEVLOW100 ZFL ZNonfinancial_MCs ZFinancial_MCs c.ZFL#c.ZNonfinancial_MCs c.ZFL#c.ZFinancial_MCs c.ZFL#c.ZNonfinancial_MCs#c.ZFinancial_MCs stdmeanROE_ROA_ stdmeansales_growperunit stdmeanVA_SALES_ stdmeanROA stdmeanEBIT stdmeanANC stdtotalHrsForm ZBusiness_age i.CEO_Dummy i.intStage i.Sector_Dummy , difficult technique(dfp) het(ZFL ZNonfinancial_MCs ZFinancial_MCs )
You can see that I have created the interactions using the # operator. The variables in the interaction terms are continuous (there are three two-way interactions and one three-way interaction). I use standardized variables because each variable's metrics are different, and I want to avoid the potential impact it might have on estimates.
The question is that I created manually the variables with the interaction terms, without considering the signs. And the results are obviously the same. However, I question on the validity and perhaps should warn against using the # operator in two-way interactions when both variables may have negative values. In short, let's say the interaction is X1*X2. So when we have values (-2) in both the interaction is (+4). And when we have (+2) in both, the interaction is also (+4).
Should we accept this?
I think we should do this manually by recalculating the interaction with negative value when both variables are negative:
gen interaction=X1*X2
replace interaction=-(1)*X1*X2 if X1<0 & X2<0
I think the interpretation this latter one interaction is different to that obtained by using the operator #. With the operator, what is relevant is the absolute value when both have negative values (having too much of both or having very few of both). On the other hand, when using the interaction built manually, when both variables have negative values, the interaction is negatively magnified, so having very few of both will be in a pole, and having too much of both will be in the opposite pole. Also, X1=(-2) and X2=(+3) will be equal to X1=(+2) and X2=(-3) in both cases, and I agree with this idea.
The question is: how should interactions be built when speaking about continuous variables that both may have negative values? Should the interaction # operator be used only when both may have values equal or above zero?
Thanks
Miguel
Comment