In the Box-Cox transformation parameter $\lambda$ is defined by likelihood function. But I cannot understand what exactly is maximized in this case? What is the purpose of maximum-likelihood in this case?


2 Answers 2


This family of transformations combines power and log transformations, and is parametrised by $\lambda$. Note that this is continuous in $\lambda$. The aim is to use likelihood methods to find the “best” $\lambda$.

Maybe it is best to provide an example, so let's assume that, for some $\lambda$ we have $E(Y ^{(λ)} ) = X\beta$ together with the normality assumption. Then, given data $Y_1, . . . , Y_n$ (ie the untransformed data), the likelihood is

$$ (2\pi \sigma^2)^{-n/2}\exp\left(-\frac1{2\sigma^2}(Y^{(\lambda)}-X\beta)^T(Y^{(\lambda)}-X\beta)\right)\prod_{i=1}^nY_i^{\lambda -1}$$

where the product at the end is the relevant Jacobian which will clearly differ in size for different values of $\lambda$, and so we want the optimal one for it to be consistent with our data. For each $\lambda$, fitting the linear model gives $\hat{\beta}{(\lambda)} = (X^TX)^{-1}X^TY^{(\lambda)} , RSS(λ) = (Y^{(\lambda)})^T(I_X)Y^{(λ)}$ , and $\hat{\sigma}^2 (λ) = RSS(\lambda)/n$ (the maximum likelihood estimate).

The profile log-likelihood for $\lambda$, obtained by maximising the loglikelihood over $\beta$ and $\sigma^2$, is therefore

$$ L_{max}(\lambda)= c - \frac{n}{2}\log(RSS(\lambda)/n)+ (\lambda-1)\sum_{i=1}^n \log(Y_i)$$

And so... we treat this as we usually treat log-likelihood functions: values of $\lambda$ close to the maximising value $\hat{\lambda}$ of $\lambda$ are consistent with the data.

  • $\begingroup$ What is X? What is beta? $\endgroup$
    – railgun
    Commented Aug 18, 2023 at 22:16
  • $\begingroup$ @railgun X is the input/predictor variables, and beta is the regression coefficients. $\endgroup$ Commented Oct 30, 2023 at 2:46

This is a good question. One can argue that the model used to estimate the box-cox transformation, something like $$ y_i^{(\lambda)} = \beta_0 + x_i^T \beta +\epsilon_i, \quad 1=1,\dotsc,n $$ with the error term $\epsilon_i$ independent and identically distributed with a normal distribution , zero mean and some variance. This is problematic as a statistical model Peter McCullagh wrote a paper about that https://projecteuclid.org/euclid.aos/1035844977) and I will come back and try to write about that, but no time now.

For one thing, the $\beta$ parameters and the variance will depend on the transformation parameter $\lambda$, but more important, the meaning of the model will change with changing $\lambda$. But still "estimating" $\lambda$ could be a meaningful thing to do, as help in modeling. It could still be that it is really not estimation in a scientific sense (since the $\lambda$ parameter do not reflect or represent anything in the reality we are modeling, it just indexes a family of models).

But the most obvious thing that happens when varying $\lambda$, is that the size of the $y^{(\lambda)}$ will change. That must be accounted for, and the jacobian is introduced for that reason. A post with details is How do I get the Box-Cox log likelihood using the Jacobian?

(When time (after easter or later) I will come back and (try to) explain my maybe somewhat cryptic comments above)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.