4
$\begingroup$

I am using OLS on log(outcome) to predict future outcomes.  My question is which method of correction to backtransformed predictions will least bias the modeled effect of change in predictor values.

  I know that transforming outcomes is fraught.  An answer of “don’t transform” or “don’t use OLS” is often or even usually better.  But they don’t answer my question or solve the problem.  For my application, both the predicted log(outcomes) and the predicted linear, or backtransformed, outcomes are used downstream in the larger model.  Also, part of what I am doing is assessing at both scales whether predictions of withheld data are more accurate if log predictions are backtransformed, or linear predictions are log transformed.  I hope you will agree that the question is valid even if transformations in general, and OLS on log(outcomes) in particular, are very imperfect tools.  

  Here are possible correction approaches.  I am interested to consider another still if that is better:  

 

\begin{equation} k = \frac{{\sum}y}{{\sum}e^{\widehat{ln(y)}}} \end{equation}  

The outcome I am modeling ranges over 7 orders of magnitude and is almost exponentially distributed.  The methods listed above yield constants that differ widely.    

For “best” method I have two criteria.  The mean of the linear predictions should be unbiased, which is what all the corrections aim to improve.  Second, both linear and log forms of predicted outcomes should respond to variations in predictor values as realistically as the sample enables.  A constant correction factor multiplicatively affects linear-scale response sensitivity.  My math isn't strong enough to state this generically, but here is an example. If a coefficient for a predictor x of log(outcome) is 3, then a one unit increase in predictor x causes a 3-unit additive increase in predicted ln(outcome), and with no correction a tripling of the backtransformed predicted outcome.  If the correction factor (k, above) is 2, then the same change in x causes the backtransformed and corrected predicted outcome to increase instead by a factor of 6.  I have no independent way to know if 3 or 6 is more accurate.

So I hope the best choice can be determined theoretically, or at least to understand why it can’t.  And if you know of a reference I have not found that describes the effect of bias correction method on accuracy of predictor sensitivity, please point.

$\endgroup$
3
  • $\begingroup$ I often apply "match first moment", but maybe there are indeed better approaches - good question! $\endgroup$
    – Michael M
    Commented Feb 18, 2023 at 17:17
  • 2
    $\begingroup$ "I am using OLS on log(outcome) to predict future outcomes" tells you the answer, because it says you are interested in getting good agreement between the log predictions and the log values. No back transformation is needed. If you were really interested in predicting the actual values (as you suggest at "the mean of the linear predictions should be unbiased," if by that you intend the arithmetic mean), you wouldn't be using OLS on the logs for this analysis. $\endgroup$
    – whuber
    Commented Feb 18, 2023 at 18:02
  • 1
    $\begingroup$ @whuber, thank you. This goes into a climate model and it simulates fire emissions. Fire emissions affect cloud microphysics (cloud condensation nuclei) in proportion to a log scale. The same emissions affect plant tissue & photosynthetic capacity, and greenhouse gas concentrations, in proportion to linear emissions. So both matter. I realize there are trade-offs - that's a main point of the project. But I want to make the best assessment I can of costs to estimates of one consequence of good fits to the scale of the other. $\endgroup$
    – InColorado
    Commented Feb 19, 2023 at 17:26

2 Answers 2

3
$\begingroup$

There are two common ways to backtransform a model $\ln y=X\beta+e$ into a point forecast:

  • $\hat y=e^{X\hat \beta}$
  • $\hat y=e^{X\hat \beta+\sigma^2_e/2}$

The latter one is theoretically inspired by lognormal distribution when $e\sim \mathscr N(0,\sigma^2_e)$. The problem is that we don't know the variance and have to estimate it. Therefore, it is not clear whether this form is actually better in practice. Empirical studies show that often the first form is more optimal (see The role of the log transformation in forecasting economic variables).

That is assuming certain kinds of loss functions, such as squared error loss $\mathscr L(\hat y,y)=(\hat y-y)^2$. In practice, apart from mathematical convenience there's no reason to use such loss function. Other loss functions lead to different optimal forecasts.

In my opinion, the best solution is to do away with point forecast, and use instead distributional forecast. That way you don't worry about this stuff.

$\endgroup$
5
  • 1
    $\begingroup$ added the reference $\endgroup$
    – Aksakal
    Commented Feb 18, 2023 at 17:35
  • $\begingroup$ Given the constraints of the larger model within which I am working, I don't have the option of a distributional forecast, though I agree in principal. Why, though, is option 4, moment matching, not best still? And isn't the first option you list biased - the premise of Duan, Miller, Newman etc. $\endgroup$
    – InColorado
    Commented Feb 18, 2023 at 17:35
  • $\begingroup$ moment matching doesn't address the question I raised: why should your point forecast to be a mean? why not median? or some other value? $\endgroup$
    – Aksakal
    Commented Feb 18, 2023 at 17:44
  • $\begingroup$ The backtransformed version should be a mean on a linear scale just as OLS leads to the untransformed simulation being a mean on a log scale, or at least I believe that is the best unbiased estimator on a linear scale. The outcomes that the linear simulations subsequently predict downstream in the larger model respond at a near-linear scale. $\endgroup$
    – InColorado
    Commented Feb 18, 2023 at 17:49
  • $\begingroup$ why should it be linear? e.g. in your case is the cost of error symmetrical? is it quadratic or linear or something else? $\endgroup$
    – Aksakal
    Commented Feb 18, 2023 at 17:50
1
$\begingroup$

Your original model is a conditional expectation $E[\log y|X] = XB$, and you are now interested in $ \hat y $. Knowing that $\log y|_X \sim \mathcal N (XB,\sigma^2)$, then $$y|_X \sim \text{lognormal}\left(\exp\left(XB+\frac{\sigma^2}{2}\right), (\exp(\sigma^2)-1)\exp(2XB+\sigma^2)\right),$$

which is the second result in Aksakal's answer.

If the assumptions of the model are correct, then this is distributionally correct, although efficient estimation of $B$ and $\sigma^2$ also comes into play.

One important point is that this estimate is not expected to be optimal with regards to $\text{MSE}$, since it was not derived to be so, because this optimality is guaranteed in the $\log$ scale instead.

$\endgroup$
1
  • $\begingroup$ True, and I believe this summarizes (well) the Miller84 approach. Given that the criterion of true log-normality of residuals is not decisive, my question is whether and why it is better to use Miller, Duan, or, especially, moment-matching. $\endgroup$
    – InColorado
    Commented Feb 24, 2023 at 19:22

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.