# What exactly is the Horwitz equation?

During a laboratory exercise in a course on analytical chemistry, we were asked to calculate the $z$-score.

$$z=\frac{X_{\text{lab}}-X_{\text{cons}}}{S_{\text{target}}}$$ where

• $X_{\text{lab}}$ is some measured quantity (e.g., concentration),
• $X_{\text{cons}}$ is the consensus value for that $X_{\text{lab}}$,
• $S_{\text{target}}$ is known as the benchmark standard deviation

One way of optaining the $S_{\text{target}}$ was to use the so-called Horwitz equation:

$$S_{\text{target}}=2^{(1-0.5\log C)}.$$ $C$ denotes the concentration as a dimensionless mass fraction.$^{[1]}$

On the mathematical relationship, M. Thompson comments$^{[2]}$

As well as being useful, the Horwitz Trumpet is a feature of considerable theoretical interest. It is hard to avoid the assumption that a simple mathematical law that describes the behaviour of large numbers of methods over at least six orders of magnitude of analyte concentration must have some inherent meaning and deserves serious consideration. So far, though, nobody has managed to explain the strange empirical exponent from basic principles, although several people have made conjectures. Are we seeing the manifestation of a physical law here, or is there a psychological basis, perhaps to do with our perception of fitness for purpose?

There is a sure-fire paper in Nature waiting for somebody!

• In statistical terms, why have we chosen such a form for $z$?

This might include an explanation of the relationship with the Horwitz equation.

• What are the several conjectures that have been made?

[1] Carlos Rivera, Rosario Rodríguez. Horwitz Equation as Quality Benchmark in ISO/LEC 17025 Testing Laboratory

[2] M. Thompson. The amazing Horwitz function. Royal Society of Chemistry. (2004)

• Look i am from mexico, so my English is not perfect but horwitz was the editor of the AOAC that is a book that contains test methods for various matrixes, during this time Horwitz recouncade data of validations of these methods and obtained an estimate of the behavior of the variability of the results and described it in the AOAC with the equation you have, so that the result of that equation says the estimated dispersion of the values with respect to your concentration of the analyte, for more information checks the appendix of validation of the AOAC. I will review the site conditions and if – Emmanuel Jul 30 at 22:20
• I think I agree with M. Thompson that the Horwitz equation is useful but disagree rather intensely with the idea that is is "of considerable theoretical interest." – Curt F. Jul 31 at 16:16

The Horwitz function 1, also known as the Horwitz “Trumpet”, is an empirical relationship that has, thus far, not been derivable from fundamental principles. As noted by Thompson 2, and as shown below in Fig. 1 (from 2), the relative standard deviation of reproducibility (RSDR) increases as analyte concentration decreases:

Horwitz formulated his expression after noticing a pattern in the relative standard deviations reported in many (thousands at this point) collaborative trial results. The following figure (a screenshot from Hanley [3, p. 12037]) provides the essential equations and background information:

As Hanley notes, the Horwitz function appears to be in the Tweedie family of exponential dispersion fluctuation scaling relationships (https://en.wikipedia.org/wiki/Tweedie_distribution ). Specifically, the Horwitz function may be a compound Poisson-Gamma process, since the exponent, $$\alpha$$ in equation 2, is approximately 0.85. (More accurately, Thompson 1 notes that it is $$1 - (log2)/2$$).

Assume measurements are Gaussian distributed, with $$\mu$$ as the population mean and $$\sigma$$ as the population standard deviation, and assume both parameters are unknown. Let $$X_{lab}$$ be a single measurement, i.e., a random sample from the Gaussian distribution. Then $$z = (X_{lab} - \mu )/\sigma$$. So z, which is the so-called ‘z score’, is simply the signed difference parsed in units of $$\sigma$$. A small z score means the measured result is close to the true value, i.e., $$\mu$$, assuming no systematic error. A large z score means the measurement has relatively poor accuracy.

But both $$\mu$$ and $$\sigma$$ are unknown, so $$X_{cons}$$ is used as an estimate of $$\mu$$ and $$S_{target}$$ is used as an estimate of $$\sigma$$. The result is a rough, and possibly biased, estimate of z. Frankly, this is too crude for me.