Back-transforming standard errors using the Delta method

Often, we want to back-transform estimated means from an analysis to the original scale the response variable was measured on. For example, you may have square-root transformed your response variable prior to analysis by ANOVA in order to approximate your residuals to a Normal distribution, or you may have analysed count data using a Poisson generalized linear model (GLM) with a log link function. In both types of situations, the treatment means (also called predicted or adjusted means) you obtain from your analysis, and their standard errors, are on the transformed scale. However, when reporting your results, you may want to transform them back to the original scale to aid interpretation.

Transforming an estimated mean back to the original scale is fairly straightforward: we just apply the inverse function of the transformation. For example, to “reverse” the square-root transformation, we take the square. Or to “reverse” the natural log transformation, we take the exponential.

Unfortunately this is not the case, and doing so would result in a standard error that is completely wrong! We can, however, calculate an approximate standard error for our back-transformed mean, using what is known as the Delta method.

Formula for the approximate standard error of a back-transformed mean

Let’s denote the estimated mean and its standard error, on the transformed scale, by \(\bar{z}\) and \(se_{\bar{z}}\), respectively. In addition, let’s denote the function used for transformation by \(g()\), its inverse by \(g^{-1}()\), and its first derivative by \(g'()\).


On the back-transformed scale the estimated mean is given by:

\[\bar{y} = g^{-1}(\bar{z})\]

And an approximate standard error of \(\bar{y}\) is given by:

\[se_{\bar{y}} = \frac{se_{\bar{z}}}{|g'(\bar{y})|}\]

Note: The bars surrounding \(g'(\bar{y})\) tell us to take the absolute value of \(g'(\bar{y})\). For example, if \(g'(\bar{y}) = -2\), then \(|g'(\bar{y})| = 2\).

Example 1: Square-root transformation

Suppose our estimated mean on the square-root scale is 3.12, and its standard error is 0.14.


For the square root transformation, \( g^{-1}(x) = x^2 \) and \( g'(x) = \frac{1}{2\sqrt{x}} \)


Thus, on the back-transformed scale our estimated mean is:

\[\bar{y} = 3.12^2 = 9.73,\]

and its approximate standard error is:

\[ se_{\bar{y}} = \frac{0.14}{\left| \frac{1}{2\sqrt{9.73}} \right|} = 0.87. \]

Example 2: Natural log transformation

Suppose our estimated mean on the natural log scale is 1.938, and its standard error is 0.096.

For the natural log transformation, \( g^{-1}(x) = \exp(x) \) and \( g'(x) = \frac{1}{x} \)


Thus, on the back-transformed scale our estimated mean is:

\[\bar{y} = \exp(1.938) = 6.945,\]

and its approximate standard error is:

\[ se_{\bar{y}} = \frac{0.096}{\left| \frac{1}{6.945} \right|} = 0.667. \]

Warning: Proceed with Caution
However, it is extremely important to realise that these back-transformed standard errors are only approximate. You need to consider that this approximation is not always very good. Indeed, for highly non-linear transformations, the resulting value can be severely misleading!

Yes, there is a far better and far easier way!


Confidence intervals are a very effective way to describe the uncertainty in an estimate. And what’s more, it is very easy to back-transform a confidence interval of a mean1. To get back to the original scale, we simply apply the inverse function, \( g'() \), to both the lower and upper limit.


For example, if the 95% confidence interval of our estimate mean on the square-root scale is (2.84, 3.40), then the 95% confidence interval of the back-transformed mean is:

\[ (2.84^2, 3.40^2) = (8.07, 11.56) \]

Note, however, that this back-transformed confidence interval will often no longer be symmetric about the back-transformed mean.

Final Remarks

So… after all that, in conclusion, if you’re back-transforming means from an analysis, provide their back-transformed confidence intervals too. They are easy to calculate and provide an appropriate and interpretable description of uncertainty. In Genstat, the BACKTRANSFORM procedure can be used to calculate back-transformed means along with their confidence intervals. It also calculates approximate standard errors – but use these with caution! The procedure does provide a way of checking whether the approximate standard errors are adequate. It calculates confidence intervals using the approximate standard errors, and plots them alongside the back-transformed confidence intervals (as per above). If the back-transformed

Picture of Dr. Vanessa Cave
Dr. Vanessa Cave
Dr. Vanessa Cave is an applied statistician interested in the application of statistics to the biosciences, in particular agriculture and ecology, and is a developer of the Genstat statistical software package. She has over 15 years of experience collaborating with scientists, using statistics to solve real-world problems. Vanessa provides expertise on experiment and survey design, data collection and management, statistical analysis, and the interpretation of statistical findings. Her interests include statistical consultancy, mixed models, multivariate methods, statistical ecology, statistical graphics and data visualisation, and the statistical challenges related to digital agriculture. Vanessa is a past President of both the Australasian Region of the International Biometric Society and the New Zealand Statistical Association, on the Editorial Board of The New Zealand Veterinary Journal and an honorary academic at the University of Auckland. She has a PhD in statistics from the University of St Andrew.

Comments

2 Responses

  1. Thanks for your question.

    If you’d like to cite this blog post, you can use the following reference: Cave, Vanessa. Back-transforming standard errors using the Delta method. VSNi, Feb. 2023. If possible, please include a link to this article. Hope this helps!

Leave a Reply

Your email address will not be published. Required fields are marked *

Related articles

How much of the variation in your field trial comes from the treatments, and how much comes from...
What should you do when count data are more variable than your model expects? Vanessa Cave and David...
Precision breeding can deliver scientifically impressive traits, but practical value depends on more than technical success. Owen McIntosh...