Centred variables in non-linear regressions

[Loic DUBOIS]

Hello everyone,

I hope you're all well.

I have a question about the centring of variables when there is interaction.

I know that in traditional linear models, the continuous variables that make up the interaction must be centred to make it easier to interpret the coefficients. My question is: should this also be done for non-linear models such as quantile regressions and fractional logit regressions (fracreg logit)?

For example

FOR =β_0+β_1 DUMGEND+ β_2 DIVERSITY+ β_3 DUMGEND*DIVERSITY+Ycontrol_(i,t)+ α_i+ δ_t+ ε_(i,t)

Here DUMGEND is a dummy that takes 1 or 0 and DIVERSITY is a continuous variable. So I've centred DIVERSITY in the linear regression framework, should I also do this in non-linear regressions?

Thanks in advance

[Daniel Schaefer]

Mean centering continuous variables in general will make the will make the intercept more interpretable. The intercept is the predicted value when all else is zero, so it is desirable to have zero equal the mean - particularly in situations where zero lies outside of the domain of the uncentered continuous variable. Mean centering continuous variables is desirable in interactions specifically because it reduces collinearity between the first and second order terms. The same collinearity issues arise in nonlinear logit regressions.

So yes, it's probably a good idea to mean center your continuous variables in a logit or one if its generalizations.

[Joao Santos Silva]

Dear Loic DUBOIS,

Please note that standard quantile regression (e.g., qreg) estimates a linear model.

Best wishes,

Joao

[Loic DUBOIS]

Thank you both for your answers,

I hope you have a good week,

Loïc