I have a fitted linear regression of a Box-Cox transformed dependent variable, using an indicator variable as one of the two predictors :

$$ g(Q, \lambda) = \hat{\beta_0} + \hat{\beta_1}P+ \hat{\beta_2}I +\epsilon = f(P,I) +\epsilon $$

This generates two parallel lines of slope $\beta_1$ and intercepts $\beta_0$, when $I=0$, and $\beta_0+\beta_2$, when $I=1$. $g(Q, \lambda)$ would be the Box-Cox transformation of $Q$ by parameter $\lambda$

I am interested in estimating how $Q$ (untransformed) changes, given a value of $P = \bar{P_0}$ when $I$ changes from $0$ to $1$. Note that $\bar{P_0}$ is actually the average of the values of $P$ where $I=0$, eventually I need to do the same for $\bar{P_1}$, i.e. the average of $P$ when $I=1$, but I guess that's just repeating the same procedure.

Since, I need $Q$ in the untransformed space, It is my understanding that I need to correct for bias in the back transformation process, for this I am fiddling with Duan's smearing, using the following formula to find $Q$ as a function of $I$:

$$ Q(I) = \frac{1}{n}\sum_{i=1}^nh(f(\bar{P_0},I) + \hat{\epsilon}_i, \lambda) $$

where $f$ is the linear regression model above, $h = g^{-1}$ is the inverse of the Box-Cox transformation and $\hat{\epsilon}_i$ are the residuals of the fitted linear regression.

Hoping to be using the right approach, there are a few things that I am finding unclear:

  1. The domain of the inverse transformation of Box-Cox is not the entire number line. in my formula above, it requires $f(P_0,I) + \hat{\epsilon}_i > \frac{-1}{\lambda}$. I have a few points where this condition doesn't hold, and my reverse transformation is impossible. This doesn't seem to be an issue that is considered in the (admittedly sparse) literature I have found on the topic (e.g. this article), so I'm not sure how to deal with it. At the moment I am substituting the entire $h(\cdot, \lambda)$ term with $0$ in these cases because logically that makes the most sense to me given the data ($Q$ and $P$ are volumes, so cannot be negative) and the shape of the Box-Cox transformation. Is there a better approach?

  2. How should I deal with the indicator variable when taking the $\hat{\epsilon}_i$ to add to the sum? Should I use all the residuals regardless of the value of $I$ that I am testing for, or only the residuals of the data points with $I$ matching the one I am testing at the specific moment? I guess that mostly means asking if I should consider the two lines that come out of the regression as two separate models or as a single one. I can see benefits and pitfalls in both approaches and admittedly the results I get are not that different. Still would be good to understand what the best practice is.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.