As Peter mentioned in this reply


Sometimes we do have a priori information about the causality relationship between features/predictors

Using the fire fighter example: if we want to predict (note, this is for prediction, not for causality inference) $Y$ the damage of a fire incident in dollars term, given number of floors $x_{floor}$ that caught fire, and number of fire fighters involved $x_{fighters}$

$Y= \beta_{1} * x_{floor} + \beta_{2} *x_{fighter}$

Now we know that more floor caught fire would cause more fire fighters to be called. However, the above linear model does not take advantage of this piece of knowlege (at least, not explicitly)

So, here is the my question: is there a common practice to explicitly include causality knowledge among features (note: not causality between feature and target) into a linear predictive model?